Every earlier module of this course has asked a system to remember something. The carbon file of A03 needs a home and a data quality score. The sustainability-linked facility of B03 needs its KPIs, targets and margin grid stored so that a test result changes the price. The agency team of B17 needs a compliance certificate logged, a margin reset keyed once and propagated everywhere, and a declassification flag that reaches the data warehouse the same week. The reporting team of E06 needs one record per counterparty with its CSRD status and Taxonomy KPIs, and loans tagged so that a Green Asset Ratio can be traced down to contracts. This module is about the machine that does all of that: the operating model, the data and the IT of the credit chain.
Where it sits. There is no single “ESG operations desk” in Meridian Bank. The work is spread across functions that a student will meet under different names in every bank:
For an asset manager, the same picture is familiar. A fund house has a front office (research, portfolio management, trading), a middle office (trade capture, performance, risk), a data management function built around a securities master that holds one golden record per instrument, and support functions. A bank’s credit chain has the same skeleton. The front office is coverage, structuring and syndication; the middle office is credit administration, loan servicing and agency; the equivalent of the securities master is the counterparty master (one golden record per legal entity and group, with its identifiers, sector, size and now its ESG attributes) completed by a facility master (one record per loan or commitment, with its terms and its labels). The support functions are Risk, Finance, Compliance and IT. Throughout this module we use that mapping, because it shows the student who has worked on the buy side where each ESG data point lands in a bank.
Clients. The functions above have no external client. Their customers are the coverage banker who wants a pre-filled ESG section in the credit memo, the credit analyst who needs a transition score, the agency officer who needs a margin reset to flow into the interest calculation, Finance who needs a reconciled Green Asset Ratio, Risk who needs financed emissions by sector for the climate stress test (see E08), and the supervisor who wants to see lineage from a published figure to the contracts.
How it earns money. It does not; it spends it. The relevant economics are those of a cost centre: people, licences, IT run costs and change budgets. Orders of magnitude in the case of this module (case assumptions): the ESG data work in Meridian’s corporate credit chain consumes about 37 full-time equivalents a year spread across six teams, and the bank pays EUR 2.6m a year for three partly overlapping ESG data licences. A transformation programme to industrialise this costs EUR 14m over three years. The question this module answers is the one a Chief Operating Officer asks: is that money well spent, in what order, and how do we know the result is controlled?
WHO DOES WHAT — The data owner of a data domain is a senior business person (for counterparty ESG attributes at Meridian, the head of ESG data and analytics) who is accountable for its definition and quality. The data steward is the operational person who maintains rules and resolves quality issues. The data architect designs the model and the flows. The business analyst writes requirements. The product owner decides the order of delivery. The control owner in Risk designs the second-line test, and the internal auditor samples it. In a job interview, being able to say who owns a field, who changes it and who checks it is worth more than knowing the vendor names.
A realistic week in November 2026. Monday: the programme’s steering committee, chaired by the Chief Operating Officer with the Chief Risk Officer and the CFO, reviews the business case of section 5 and must choose between a “big-bang” release in 2028 and a wave plan that starts with the carbon file. Tuesday: the data stewards run the monthly quality dashboard; 6% of active labelled facilities still lack a verification status in the facility master. Wednesday: an architecture review rejects a request by one business line to store sustainability-linked KPI targets in a spreadsheet attached to the loan record, because the spreadsheet cannot be reconciled or audited. Thursday: model risk management reviews the pilot of a language model that extracts emissions figures from client sustainability reports. Friday: Internal Audit presents its 2027 plan to the Audit Committee; ESG data lineage for the first year-end under the revised Pillar 3 templates is one of its themes.
WHY IT MATTERS ON THE DESK — A banker who knows where the data lives knows why a margin reset was late, why a green loan disappeared from the bank’s sustainable finance volume, and why the supervisor’s on-site team asked for the lineage of a single number. In the next decade most of the sustainable-finance jobs inside banks will be data jobs wearing a business title.
Before any ESG field is added, a bank already runs a large, regulated information chain for credit. This section describes it plainly: what a target operating model is, what the credit chain does at each step, which systems carry which data, and what regulators expect of risk data in general.
IN ONE SENTENCE — A target operating model (TOM) describes how a function will work in the future along a few fixed dimensions (processes, organisation, data, technology, governance and sourcing), so that a transformation can be planned as the gap between today and that picture.
INTUITION — Think of a restaurant that wants to add allergen labelling. The menu (the products) does not change. What changes is the kitchen: who checks each ingredient on delivery, where the information is written, how the waiter learns it, who signs the label. A TOM is the redesigned kitchen, drawn before the builders arrive.
The six dimensions are used by every consulting firm and every bank’s change function, with minor variations in names:
| Dimension | Question it answers | Typical deliverable |
|---|---|---|
| Process | Which steps, in which order, with which controls? | Process map, RACI matrix |
| Organisation | Who does each step, with which skills? | Role descriptions, sizing in FTE |
| Data | Which data, defined how, owned by whom? | Data model, data dictionary |
| Technology | Which applications and interfaces? | Target architecture |
| Governance | Who decides, who checks? | Committees, policies, controls |
| Sourcing and location | In-house, outsourced, offshore? | Sourcing plan |
A TOM is not a slide; it becomes a roadmap (what is delivered when), a business case (what it costs and returns) and a benefits register (who commits to which saving). The case of section 5 builds all three for ESG data in the credit chain.
A corporate loan passes through roughly nine steps between the first conversation and the last report. Each step creates or uses data.
Each step has its own application, often bought from a different vendor and built at a different time. Large banks typically use a market loan servicing platform for syndicated and bilateral corporate loans, a separate rating tool, a limit and collateral system, and a data warehouse that receives nightly extracts from all of them. The chain works because some data are mastered in one place and referenced elsewhere.
Three kinds of data move along the chain:
PITFALL — “We have the data” often means “someone has typed it somewhere”. If the sector code of a counterparty is maintained in the CRM, overwritten in the rating tool and corrected again in the data warehouse, the bank has three answers and no golden source. The supervisor’s question is always: which one did you report, and why that one?
A critical data element (CDE) is a field whose error would materially distort a risk decision or a report: the counterparty’s PD grade, the exposure amount, the sector code, and, as we will see, the counterparty’s emissions and the facility’s sustainable-finance label. Each CDE has a definition, an owner, quality rules and documented lineage (the path from the source to every report that uses it).
Data quality is measured on a small number of dimensions. The ones used by most banks and by the ECB are completeness (is the field filled?), accuracy (is it right?), validity (does it respect the format and code list?), consistency (is it the same in every system?), timeliness (is it available when needed?) and uniqueness (is each object recorded once?). A data quality rule tests one dimension of one CDE, for example “the sector code of every active corporate counterparty is a valid NACE code” or “the emissions figure of a counterparty has a source document and a reporting year”.
The most useful single number for a chain is the first-time-right rate: the share of records that pass every step without needing correction. Because errors at each step compound, the rate falls quickly with the number of manual steps.
FORMULA
is the first-time-right rate of a record through the chain; the probability that step handles the record correctly (its accuracy); the number of steps that touch the record; the expected number of records with at least one error; the number of records processed.
INTERACTIVE LAB — First-time-right along a chain of manual steps
See how per-step accuracy and the number of manual hand-offs determine the share of records that reach the report without any error.
| Parameter you can move | Range | Default |
|---|---|---|
| Number of steps touching the record | 1 to 12 steps | 5 steps |
| Accuracy of each step | 90 to 100 % | 98 % |
| Records processed per year | 100 to 50000 records | 10000 records |
| Result at default settings | Value |
|---|---|
| First-time-right rate | 90.39 % |
| Records with at least one error | 961 records |
Try it. Keep accuracy at 98% and double the steps from 5 to 10: one record in five now needs fixing. Removing a manual hand-off often does more than training the people at each step.
With five steps at 98% each, only 90.4% of records arrive clean; one record in ten needs fixing somewhere. Raising accuracy to 99.5% per step, the level of a system-enforced validation, brings the chain to 97.5%. This is the arithmetic behind the most common recommendation in any operating-model review: capture data once, as early as possible, with validation at the point of entry, and copy it electronically afterwards.
INTUITION — A data chain is like a relay race in which each runner drops the baton two times in a hundred. Five runners finish cleanly nine times in ten. You win more by shortening the race than by coaching each runner.
The cost of an error also depends on where it is caught. A heuristic often quoted in data quality practice, the “1-10-100 rule”, says that preventing an error at entry costs one unit, correcting it later in the process costs ten, and dealing with its consequences after it has left the bank costs a hundred. The ratios are not a law, but the direction is reliable: an error in a facility’s margin caught at booking costs a few minutes; the same error found after interest has been paid to twenty lenders costs a reconciliation project and a client letter (B17, section 5.5); an error in a published ratio costs a restatement.
IN ONE SENTENCE — BCBS 239 is the Basel Committee’s set of fourteen principles, published in January 2013, requiring systemically important banks to be able to aggregate risk data accurately, completely and quickly, and to report it to management, including in a crisis.
The principles fall into four groups: governance and infrastructure (principles 1 and 2: board ownership, a data architecture that supports aggregation in normal times and stress); risk data aggregation capabilities (3 to 6: accuracy and integrity, completeness, timeliness, adaptability); risk reporting practices (7 to 11: accuracy, comprehensiveness, clarity and usefulness, frequency, distribution); and supervisory review (12 to 14). They were written after the 2008 crisis showed that large banks could not tell their supervisors, in days, their total exposure to a failing counterparty.
Implementation has been slow everywhere. The Basel Committee’s progress report of 28 November 2023 on 31 global systemically important banks found that significant work remained at most of them. In the euro area the ECB launched a targeted supervisory strategy on risk data aggregation and risk reporting (RDARR) in 2022, ran an on-site inspection campaign covering about a third of directly supervised banks, and published in May 2024 its Guide on effective risk data aggregation and risk reporting. The Guide imposes no new law but sets the benchmark: a management body that owns the topic, a scope that covers all material legal entities and risk types, complete data lineage from reports back to source systems, clear data ownership, quality controls with thresholds, and limits on uncontrolled end-user computing (spreadsheets and desktop tools). The ECB’s findings were consistent: insufficient board attention, legacy systems, weak lineage, unclear ownership, recurring manual corrections.
WHY IT MATTERS ON THE DESK — ESG data is not exempt from any of this. When the supervisor asks for financed emissions by sector or for the share of the corporate book with reported emissions, it is asking for an aggregation of risk data. A bank that keeps its ESG data outside the BCBS 239 framework is building a second, weaker chain beside the first, and will be told so.
The documents produced by an operating-model team are as concrete as a term sheet: the process map with control points; the RACI matrix (who is Responsible, Accountable, Consulted, Informed for each activity); the data dictionary with the definition, owner, format and quality rules of each CDE; the lineage map; the target architecture diagram; the business case; the roadmap; and the control library, which lists each control with its owner, frequency, evidence and test. Internal Audit reads them in that order.
The sustainable layer adds four families of data objects to the chain, each with a different owner and a different rhythm.
| Data object | Examples | Rhythm |
|---|---|---|
| Counterparty ESG attributes | Emissions scope 1, 2, 3; PCAF data quality score; CSRD status; Taxonomy KPIs; transition plan; transition score | Annual, at client reporting |
| Facility labels and terms | Green, social, sustainability-linked; standard followed; KPIs and targets; margin adjustment | At signing, then at each test |
| Asset attributes | Location (physical risk), energy performance certificate, use of proceeds | At signing and on change |
| Evidence | Sustainability report, compliance certificate, verification report, allocation report | Annual or event-driven |
The earlier modules give the list of data points that the chain must capture. From A03: the carbon file (scope 1, 2 and 3 by source, the consolidation approach, the reporting year), the enterprise value including cash (EVIC) or equity plus debt used for attribution, and the PCAF data quality score from 1 (reported and verified) to 5 (estimated from the asset class). From B03 and B17: for each sustainability-linked facility, the KPIs, the sustainability performance targets (SPTs) by year, the test date and certificate deadline, the current margin adjustment and its effective date, the verification status, and the declassification status; for green loans (B02), the eligible category, the allocation register and the reporting obligation. From E06: the counterparty’s CSRD scope status, its published Taxonomy-aligned turnover and CapEx, the use-of-proceeds flag, and for real estate the energy performance certificate, all needed to compute a Green Asset Ratio that reconciles to the balance sheet.
Two design principles carry over from B17. First, every price that depends on an ESG test is derived from one deal-level field, never keyed separately in several places. Second, every change leaves an audit trail: who changed what, when, on which document, validated by whom.
EBA Guidelines on the management of ESG risks (EBA/GL/2025/01). Published in January 2025 under the CRD as amended by CRD6, they apply from 11 January 2026 to most institutions and from 11 January 2027 at the latest to small and non-complex institutions. They are the most direct operating-model text in this field. Paragraphs 23 to 29 ask for information management systems that identify, collect, structure and analyse ESG data within the bank’s overall data governance and IT infrastructure, with arrangements to assess and improve quality; client-level and asset-level data at an appropriate granularity; use of counterparties’ own disclosures, client engagement and third-party data whose methods the bank understands; and, where data are missing, documented estimates or proxies whose use should decline as data improve. For large corporate counterparties paragraph 28 lists the data points to consider: location of key assets and exposure to hazards, current and targeted greenhouse gas emissions (scopes 1, 2 and 3), dependency on fossil fuels, energy and water consumption, energy efficiency of real estate, transition plans, and alignment with international social and governance standards. Paragraphs 56 to 59 place ESG risks inside the internal control framework across the three lines of defence, with Internal Audit giving independent assurance on ESG controls and systems.
CRD6 and prudential transition plans. Article 87a of the CRD, as amended by Directive (EU) 2024/1619, requires plans to monitor and address ESG risks, including those arising from the transition to climate neutrality by 2050 (see E12). Plans need metrics, and metrics need data from the credit chain.
ECB expectations. The ECB Guide on climate-related and environmental risks (November 2020) set thirteen expectations, of which expectation 6 on reporting requires data on climate-related and environmental risks to be aggregated and reported. The ECB’s supervisory priorities for 2026 to 2028 combine two themes this module joins: banks must remediate long-standing RDARR shortcomings, and they must be fully aligned with the climate guide, with attention to ESG data quality and Pillar 3 disclosures. In November 2025 it published periodic penalty payments imposed on ABANCA for missing a climate-risk deadline, a sign that the expectations are enforced (see E01).
Pillar 3 and CSRD reporting. The final draft ITS of 22 June 2026 (see E06) keeps financed emissions by sector with scope 1, 2 and 3 shown separately and moves sector breakdowns to NACE Rev. 2.1. For the operating model this means a sector-code migration in the counterparty master and a scope-3 field that most chains did not carry at counterparty level.
DORA. The Digital Operational Resilience Act, Regulation (EU) 2022/2554, has applied since 17 January 2025. An ESG data vendor, a taxonomy-engine provider or a cloud AI service is an ICT third-party service provider: the contract must meet DORA’s content requirements, the provider must appear in the bank’s register of information, and an exit strategy is required where the service supports a critical or important function.
The AI Act. Regulation (EU) 2024/1689 classifies as high-risk the AI systems used to evaluate the creditworthiness of natural persons. The Digital Omnibus on AI, Regulation (EU) 2026/1744 published on 24 July 2026, postponed the obligations for such stand-alone high-risk systems to 2 December 2027. A model that extracts emissions figures from a corporate sustainability report, or that drafts the ESG section of a corporate credit memo, is not in that category; it remains subject to the bank’s model risk management, to DORA for the provider, and to the AI Act’s general provisions such as AI literacy of staff. A model that scores retail borrowers using ESG variables would be high-risk.
ESAP. The European Single Access Point, Regulation (EU) 2023/2859, is scheduled to open on 10 July 2027 and will progressively give machine-readable access to companies’ published information, including sustainability statements. For banks it is a future golden source for counterparty disclosures, which reduces the value of building scraping tools today.
Practice is uneven and can be described qualitatively with confidence. Most large European banks built their first ESG data capabilities between 2021 and 2024 under deadline pressure (the first Pillar 3 ESG disclosures, the first GAR), often as tactical solutions: spreadsheets, a separate ESG database fed by vendors, manual tagging of labelled loans. Many are now in a second phase, integrating ESG attributes into the counterparty and facility masters and into the BCBS 239 framework. Three tensions dominate the discussions. Regulatory simplification in 2025 and 2026 (Omnibus I, the 2026/73 delegated act, the Pillar 3 final draft) reduced some reporting demands, which weakened the internal case for investment, while supervisory risk-management expectations (EBA guidelines, ECB priorities) did not fall. Vendor data are expensive and frequently overlap. And language models now make document extraction cheap, which changes the build-or-buy calculation for the carbon file but raises model risk and evidence questions.
PITFALL — “Regulation was simplified, so we can postpone the data programme” confuses disclosure with risk management. The GAR may not be published by Meridian before 2028, but the EBA guidelines have applied since January 2026, the climate stress test needs counterparty-level data, and every sustainability-linked margin depends on a KPI record. The data are needed even if fewer of them are published.
The “deal” of this module is the life of the ESG data attached to one sustainability-linked term loan, from the first meeting to the bank’s published figures, in Meridian’s target operating model. The flow shows who captures each data point, in which system, and which control applies.
INTERACTIVE LAB — The ESG data of one sustainability-linked loan, from origination to the published report
Follow each ESG data point of a facility through the credit chain: who captures it, who controls it, where it is stored and who uses it.
| # | Who | Does what | Deliverable |
|---|---|---|---|
| 1 | Coverage banker | Records the opportunity in the CRM and sends the ESG questionnaire through the client portal, limited to the voluntary-standard content for mid-sized clients | Opportunity record, questionnaire request |
| 2 | Client onboarding (KYC) | Confirms LEI, group structure and NACE Rev. 2.1 code in the counterparty master | Golden counterparty record |
| 3 | ESG data analyst | Builds or refreshes the carbon file from the sustainability report with AI-assisted extraction and page references; assigns the PCAF data quality score | Carbon file, data quality score |
| 4 | Second ESG data analyst | Four-eyes review of extracted figures flagged as low confidence and a sample of the rest | Validated counterparty ESG attributes |
| 5 | Credit analyst | Uses the pre-filled ESG section and transition score in the credit memo; records the sustainability-linked structure proposed | Credit application with ESG section |
| 6 | Sustainable finance office | Checks KPIs and targets against the SLLP and the internal framework; assigns the internal classification | Classification opinion |
| 7 | Credit committee | Approves limit, pricing grid and conditions, including the KPI reporting covenant | Credit decision in the limit system |
| 8 | Credit operations | Books the facility; creates the KPI and SPT records and the margin grid in the facility master; second operator validates | Facility record with ESG label and KPI schedule |
| 9 | Loan agency / servicing | Receives the annual compliance certificate and verification report, logs them, computes the adjustment once and lets pricing derive from it | Updated margin adjustment with audit trail |
| 10 | Data steward (CDO) | Runs monthly quality rules on ESG CDEs and opens issues for breaks | Data quality dashboard, issue log |
| 11 | Risk control (second line) | Tests key controls on a sample; challenges proxies and classifications | Control test results |
| 12 | Finance sustainability reporting | Extracts exposures, reconciles to the ledger and computes financed emissions, GAR inputs and Pillar 3 tables | Reconciled sustainability metrics |
| 13 | Internal audit and statutory auditor | Trace a sample of published figures back to contracts and evidence | Audit report, limited assurance opinion |
Origination (steps 1 and 2). The coverage banker’s ESG questionnaire is the first place where data enter. In Meridian’s target model the questionnaire is a structured form in the client portal rather than a spreadsheet sent by email, its content is mapped field by field to the counterparty master, and its scope for mid-sized clients is capped at the voluntary standard’s content so that the request is lawful and proportionate. Onboarding fixes the identifiers. The single most damaging data error in an ESG chain is a wrong legal entity or a wrong group link, because emissions, Taxonomy KPIs and CSRD status are then attached to the wrong company and propagate to every loan in its group.
Carbon file and validation (steps 3 and 4). These are the A03 steps, now partly automated. The analyst works from the source document; the extraction tool proposes each figure with its page reference and a confidence level; the analyst confirms or corrects. A second analyst reviews all low-confidence fields and a sample of the high-confidence ones. The output is a set of counterparty attributes, each with a source, a reporting year, a data quality score and a validation status. Nothing typed outside the tool is accepted.
Credit decision (steps 5 to 7). The credit memo template reads the counterparty attributes and the transition score directly. The analyst comments on them but does not retype them. The sustainable finance office records its classification opinion in a field, not a free-text email, because that field will drive the bank’s internal sustainable finance volume. The credit committee’s decision record includes the reporting covenant (certificate deadline, verification standard), which becomes a monitored obligation in the facility master.
Booking and servicing (steps 8 and 9). Credit operations create the facility with its label, its KPI and SPT schedule and the grid. The control is a four-eyes validation against the signed agreement. During the life of the loan the agency or servicing team logs each certificate, computes the adjustment once, and lets the interest and fee calculations derive from it (see B17 section 5.4). A late certificate automatically triggers the contractual default adjustment and a task, not a reminder in someone’s mailbox.
Data quality, control and reporting (steps 10 to 13). The data steward runs the quality rules monthly on every ESG CDE: completeness (does every labelled facility have a verification status?), validity (is the data quality score between 1 and 5?), consistency (is the label in the facility master the same as in the finance warehouse?), timeliness (was the reset applied within ten business days of the certificate?). Breaks go to an issue log with an owner and a due date. The second line tests key controls on a sample and challenges proxies. Finance extracts, reconciles to the general ledger and computes the metrics described in E06 and E07. Internal Audit and the statutory auditor trace a sample of published figures back through lineage to contracts and source documents.
WHO DOES WHAT — The documents that circulate along this flow are: the questionnaire, the carbon file with its evidence pack, the credit memo, the classification opinion, the credit decision, the facility agreement and its KPI schedule, the annual compliance certificate and verification report, the data quality dashboard, the issue log, the control test report and the audit report. A good operations analyst can name, for each one, the system where it is stored and the field it feeds.
Question. Meridian’s Chief Operating Officer has received a request for EUR 14m over three years to “industrialise ESG data in the credit chain”. She asks the change team for three things before the steering committee: a measured picture of today, a target operating model, and a business case with an NPV at the bank’s 12% hurdle and a payback period.
Data (case assumptions). The scope is the corporate credit chain of the CIB and Corporate Banking: about 3,200 corporate obligor groups with EUR 92bn of drawn exposure. Meridian maintains a carbon file (see A03) for 1,100 large corporates each year. It has 460 labelled facilities (green loans and sustainability-linked loans and RCFs) that require an annual KPI test or allocation check. Each of the 3,200 annual credit reviews contains an ESG section. The bank publishes its financed emissions and Pillar 3 ESG tables twice a year (at 30 June and 31 December). A full-time equivalent works 1,600 productive hours a year at a fully loaded cost of EUR 110,000, so one hour costs EUR 68.75.
The change team spent four weeks with the six teams involved and measured the time per unit of work. Today’s chain is typical of a first-generation build: the carbon file is keyed by hand from PDF reports into a spreadsheet and then uploaded; KPI certificates arrive by email and are re-keyed in the loan servicing system and in a separate register; credit analysts copy ESG figures into the memo; and each reporting cycle ends with weeks of manual reconciliation between the ESG database, the loan system and the finance warehouse.
| Activity | Volume per year | Hours per unit today | Target hours per unit |
|---|---|---|---|
| Carbon file | 1,100 files | 22 | 8.25 (see 5.4) |
| KPI test or allocation check | 460 facilities | 30 | 12 |
| ESG section of credit review | 3,200 reviews | 4 | 2 |
| Reporting cycle reconciliation | 2 cycles | 4,000 | 1,200 |
The target hours are the change team’s estimates for the redesigned process, each justified later: the carbon file by the extraction arithmetic of 5.4, the KPI test by a structured certificate intake that writes once to the facility master, the credit review by a pre-filled template, and the reporting cycle by lineage and automated reconciliations.
INTERACTIVE LAB — Where the ESG data effort goes in Meridian’s credit chain, today and in the target model
Convert volumes and hours per unit into hours, FTE and cost for each activity, today and in the target operating model.
| Parameter you can move | Range | Default |
|---|---|---|
| Carbon files per year | 200 to 3000 files | 1100 files |
| Hours per carbon file today | 5 to 40 h | 22 h |
| Hours per carbon file in the target model | 2 to 40 h | 8.25 h |
| Labelled facilities tested per year | 50 to 2000 facilities | 460 facilities |
| Hours per KPI test today | 5 to 60 h | 30 h |
| Hours per KPI test in the target model | 2 to 60 h | 12 h |
| Credit reviews with an ESG section | 500 to 8000 reviews | 3200 reviews |
| Hours per ESG section today | 0.5 to 10 h | 4 h |
| Hours per ESG section in the target model | 0.5 to 10 h | 2 h |
| Hours per reporting cycle today | 500 to 8000 h | 4000 h |
| Hours per reporting cycle in the target model | 100 to 8000 h | 1200 h |
| Reporting cycles per year | 1 to 4 cycles | 2 cycles |
| Productive hours per FTE | 1400 to 1800 h | 1600 h |
| Fully loaded cost per FTE | 60 to 200 EUR k | 110 EUR k |
| Result at default settings | Value |
|---|---|
| Hours per year today | 58,800 h |
| FTE today | 36.75 FTE |
| Cost today | 4.043 EUR m |
| Hours per year in the target model | 23,395 h |
| FTE in the target model | 14.62 FTE |
| Cost in the target model | 1.608 EUR m |
| Annual labour saving at full run rate | 2.434 EUR m |
| FTE saved at full run rate | 22.13 FTE |
| Cost of one hour | 68.75 EUR |
| Carbon file share of today’s hours | 41.2 % |
Try it. Set the target hours of the carbon file back to 22: the saving falls by about EUR 1m. The carbon file is the largest single lever, which is why section 5.4 looks at it in detail.
Calculation. Today the chain consumes 58,800 hours a year: 24,200 on carbon files, 13,800 on KPI tests, 12,800 on credit-review sections and 8,000 on reporting reconciliations. At 1,600 hours per FTE this is 36.75 FTE, or EUR 4.04m a year. The target model needs 23,395 hours, 14.62 FTE and EUR 1.61m. The labour saving at full run rate is EUR 2.43m a year, or 22.13 FTE.
Reading. Two observations frame the rest of the case. First, the carbon file alone is 41% of today’s effort; any plan that does not change how it is produced will not change the economics. Second, 22 FTE of “saving” is not cash until people are redeployed or not replaced. In a bank of 21,400 employees with natural attrition of several percent a year, this is achievable without redundancies, but the benefits register must name the managers who commit to lower headcount budgets, and the finance business partner must remove the budget. A saving that nobody signs is an estimate.
PITFALL — Do not count the same hour twice. If the KPI test today includes re-keying the margin reset in the loan system, and the error-cost model of 5.3 also counts the cost of fixing wrong resets, the business case must take the time spent on routine keying in 5.1 and only the cost of fixing errors in 5.3. The change team built the two estimates from separate time logs to avoid double counting.
Question. Which objects and fields must the target model hold, where is each one mastered, and what data quality should the bank aim for?
The data model. The change team kept the asset manager’s logic: a few master objects, each with one golden source, and everything else referenced. Four objects carry the ESG attributes.
| Object | Golden source | Owner | Key ESG CDEs |
|---|---|---|---|
| Counterparty | Counterparty master | Head of ESG data | Emissions by scope, reporting year, PCAF score, CSRD status, Taxonomy KPIs, transition score |
| Facility | Loan servicing system | Head of credit operations | Label, standard, classification, KPI schedule, adjustment, verification and declassification status |
| Asset | Collateral system | Head of collateral management | Location, energy performance certificate, use-of-proceeds category |
| KPI observation | KPI register in facility master | Head of loan agency | KPI value, SPT, test date, verifier, result |
Three design choices deserve explanation. Emissions sit on the counterparty, not on the loan. A facility inherits the borrower’s emissions through the attribution formula; storing emissions on each loan multiplies the copies and the errors. The KPI observation is a separate object, because a sustainability-linked facility has one observation per KPI per year, each with evidence; storing only “current adjustment” would lose the history that auditors and the borrower may dispute. The transition score is a risk output, owned by Risk, but it lives in the counterparty master so that the credit memo, pricing and the stress test read the same value.
Every CDE gets a definition in the data dictionary, quality rules with thresholds, and lineage to the reports that use it. For example, the CDE “scope 1 emissions” is defined as tonnes of CO2 equivalent under the GHG Protocol for the borrower’s consolidated reporting perimeter and reporting year, with rules “not negative”, “reporting year within the last two years for an active counterparty”, “source document attached”, “PCAF score present”, and lineage to the financed emissions calculation, the Pillar 3 sector template and the transition score.
The data quality objective. For financed emissions the PCAF score is itself the bank’s quality metric. The current distribution of Meridian’s EUR 92bn drawn corporate exposure by score, and the programme’s target at the end of year three, are case assumptions.
| PCAF score | Meaning (business loans) | Today | Target |
|---|---|---|---|
| 1 | Reported, verified emissions | 18% | 25% |
| 2 | Reported, unverified emissions | 20% | 30% |
| 3 | Physical activity data and emission factors | 12% | 20% |
| 4 | Revenue and sector emission factors | 30% | 20% |
| 5 | Asset-class or sector averages per euro of exposure | 20% | 5% |
INTERACTIVE LAB — Exposure-weighted PCAF data quality score of the corporate book
Compute the weighted data quality score, the share of exposure with reported emissions and the exposure still estimated from averages, today and at the target.
| Parameter you can move | Range | Default |
|---|---|---|
| Drawn corporate exposure | 10 to 200 EUR bn | 92 EUR bn |
| Today: share at score 1 | 0 to 100 % | 18 % |
| Today: share at score 2 | 0 to 100 % | 20 % |
| Today: share at score 3 | 0 to 100 % | 12 % |
| Today: share at score 4 (score 5 is the remainder) | 0 to 100 % | 30 % |
| Target: share at score 1 | 0 to 100 % | 25 % |
| Target: share at score 2 | 0 to 100 % | 30 % |
| Target: share at score 3 | 0 to 100 % | 20 % |
| Target: share at score 4 (score 5 is the remainder) | 0 to 100 % | 20 % |
| Result at default settings | Value |
|---|---|
| Weighted data quality score today | 3.14 1-5 |
| Weighted data quality score at target | 2.50 1-5 |
| Share with reported emissions (scores 1-2) today | 38 % |
| Share with reported emissions at target | 55 % |
| Exposure at scores 4-5 today | 46.0 EUR bn |
| Exposure at scores 4-5 at target | 23.0 EUR bn |
Try it. Move 10 points of today’s exposure from score 5 to score 3: the weighted score falls by 0.2. Collecting energy and production data from mid-sized borrowers is often cheaper than chasing verified reports.
Calculation. Today’s weighted score is 0.18 × 1 + 0.20 × 2 + 0.12 × 3 + 0.30 × 4 + 0.20 × 5 = 3.14; 38% of exposure has reported emissions, and EUR 46.0bn is still estimated from revenue or exposure averages. At the target the score is 2.50, 55% is reported, and EUR 23.0bn remains at scores 4 and 5.
Reading. The improvement comes from three sources that the operating model must support separately: more counterparties publishing (the CSRD, though fewer after the Omnibus, and the ESAP from July 2027), better capture of what is already published (the extraction tool of 5.4), and the questionnaire for mid-sized borrowers asking for energy and production data that move them from score 4 or 5 to score 3. Be ready for an awkward consequence: as scores improve, financed emissions can rise, because sector averages often understate the emissions of a heavy industrial borrower (A03, section 5.8, shows a proxy error in both directions). The steering committee must hear this before the first report, not after.
Lineage, walked once. To make lineage concrete, the change team traced one published figure, the financed emissions of the cement sector at 31 December 2025, back to source. It passed through the Pillar 3 template, a finance-warehouse table, a calculation spreadsheet, an extract from the ESG database, the carbon files of fourteen borrowers and, for three of them, a vendor feed whose methodology had changed in the year. Two manual adjustments had no documented owner. In the target model the same walk has four stops (report, calculation engine, counterparty master, source document), each with an owner and a control.
CASE — The target is deliberately not “everything at score 1 or 2”. Around 45% of the book by exposure is to borrowers that will not publish verified emissions in the next three years: mid-sized companies outside the CSRD after the Omnibus, and subsidiaries reporting only at group level. For them the efficient target is score 3, built from activity data the client already has (energy bills, production volumes), within the content of the voluntary standard.
Question. How many facility records leave the chain with an ESG error today, what does that cost, and how much does the target process save?
Data (case assumptions). The chain holds 9,600 active facility records for the 3,200 obligor groups (three facilities per group on average). Each record’s ESG data pass through five manual steps today, measured by the change team on a sample of 400 records: questionnaire capture (98% accuracy), onboarding of identifiers and sector (97%), carbon file keying (99% at record level, since most facilities inherit an already validated file), booking of label and KPI schedule (96%), and servicing resets and status updates (98%). Today 80% of errors are found before publication, by reconciliations at period end; correcting one costs EUR 400 on average (analyst and operations time). The 20% found after publication or by a client cost EUR 4,000 on average (investigation, restatement of internal reports, client communication, recalculation of fees). In the target model each step is system-validated at 99.5% and detection before publication rises to 95% thanks to monthly quality rules.
INTERACTIVE LAB — Errors in facility ESG records and the cost of finding them late
Compute first-time-right rates, the number of records with errors and the expected annual correction cost, today and in the target process.
| Parameter you can move | Range | Default |
|---|---|---|
| Active facility records | 1000 to 30000 records | 9600 records |
| Today: questionnaire capture accuracy | 90 to 100 % | 98 % |
| Today: onboarding identifiers and sector | 90 to 100 % | 97 % |
| Today: carbon file keying (record level) | 90 to 100 % | 99 % |
| Today: booking of label and KPI schedule | 90 to 100 % | 96 % |
| Today: servicing resets and status | 90 to 100 % | 98 % |
| Target: accuracy of each validated step | 95 to 100 % | 99.5 % |
| Today: share of errors found before publication | 0 to 100 % | 80 % |
| Target: share of errors found before publication | 0 to 100 % | 95 % |
| Cost to fix an error found before publication | 50 to 2000 EUR | 400 EUR |
| Cost to fix an error found after publication | 500 to 20000 EUR | 4000 EUR |
| Result at default settings | Value |
|---|---|
| First-time-right today | 88.54 % |
| Records with an error today | 1,100 records |
| Annual correction cost today | 1.232 EUR m |
| First-time-right in the target process | 97.52 % |
| Records with an error in the target process | 238 records |
| Annual correction cost in the target process | 0.138 EUR m |
| Annual saving on error correction | 1.095 EUR m |
| Errors escaping to publication today | 220 records |
| Errors escaping to publication in the target process | 12 records |
Try it. Set today’s detection to 95% without changing accuracy: the correction cost halves. Better controls at the end are cheaper to build than a new process, but they leave the error volume, and the operational risk, where it was.
Calculation. Today the first-time-right rate is 0.98 × 0.97 × 0.99 × 0.96 × 0.98 = 88.54%, so 1,100 records a year carry at least one ESG error. With 80% caught before publication at EUR 400 and 20% after at EUR 4,000, the expected cost is 1,100.3 × (0.8 × 400 + 0.2 × 4,000) = EUR 1.232m. About 220 errors reach a published or client-facing figure. In the target process the rate is 0.995 to the fifth power, 97.52%; 238 records carry an error, 12 escape, and the cost is EUR 0.138m. The annual saving is EUR 1.095m.
Reading. The weakest step today is booking (96%): an operations officer keys a KPI schedule from a 200-page facility agreement into a system that was not designed for it. The redesign removes this keying by making the sustainable finance office’s term-sheet record the source of the schedule, validated against the signed agreement, not retyped from it. The simulator also shows a cheaper alternative: improving detection alone to 95% halves the cost (EUR 0.638m) without touching the process. The change team presents it honestly as option B. It is rejected not on cost but because it leaves 1,100 erroneous records a year in the systems that price loans and feed risk models between two reporting dates, which the ECB’s RDARR guide would describe as reliance on recurring manual corrections.
PITFALL — Error counts are estimated from a sample, and the cost per error is an average with a long tail: one wrong declassification on a large SLL can cost more than all routine corrections of a year (B17, section 5.5). Present the error saving as a range and never as the anchor of a business case.
Question. Which tools does the target model need, which should be bought and which built, and where does AI genuinely help?
The target architecture has five components around the two masters:
| Component | Function | Choice for Meridian |
|---|---|---|
| Client portal questionnaire | Structured capture from clients, mapped to CDEs | Extend the existing portal (build) |
| Emissions calculator | Attribution, PCAF scores, proxies by sector | Configure the vendor tool used by Finance (buy) |
| Taxonomy engine | Eligibility and alignment tests, GAR inputs | Buy; rules change too often to maintain in-house |
| Document extraction | Read reports and certificates, propose values with page references | Build on the bank’s approved language-model platform |
| Data quality and lineage | Quality rules, dashboards, lineage maps | Extend the CDO’s existing tooling |
Buy or build. The rule the architects apply is simple. Buy where the logic is defined by an external standard that changes often and identically for all banks (Taxonomy screening criteria, PCAF sector factors): a vendor spreads the cost of updates across clients. Build where the logic is specific to the bank’s processes and data (mapping to its masters, its questionnaire, its controls). Under DORA, each bought service becomes an ICT third-party arrangement with a contract review, an entry in the register of information and an exit plan; a taxonomy engine that feeds published figures supports an important function.
AI-assisted extraction. The carbon file is the largest cost (5.1). An analyst today reads a sustainability report of 150 to 300 pages and keys about 60 data points (emissions by scope and category, consolidation basis, base year, targets, energy, Taxonomy KPIs). The measured time is 22 hours per file, of which 15 hours are finding and keying values and 7 hours are analysis (plausibility checks, peer comparison, scoring, commentary).
In the target process a language model reads the report and proposes each value with its page reference and a confidence level. The analyst reviews every low-confidence value in full, about 30% of them, at three minutes each, and spot-checks the high-confidence 70% at half a minute each, by clicking on the page reference. The analysis time is unchanged: the tool removes typing, not judgement.
INTERACTIVE LAB — AI-assisted extraction of the carbon file: time and residual errors
Compare manual keying with model extraction plus human review on hours per file, annual hours saved and expected residual errors per file.
| Parameter you can move | Range | Default |
|---|---|---|
| Data points per carbon file | 20 to 200 points | 60 points |
| Manual: minutes to find and key one point | 2 to 30 min | 15 min |
| Manual: error rate per point | 0 to 10 % | 2 % |
| AI: share of points with low confidence | 0 to 100 % | 30 % |
| AI: minutes to review a low-confidence point | 0.5 to 10 min | 3 min |
| AI: minutes to spot-check a high-confidence point | 0 to 5 min | 0.5 min |
| AI: residual error on high-confidence points | 0 to 10 % | 1 % |
| AI: residual error on reviewed points | 0 to 10 % | 0.5 % |
| Analysis time per file (unchanged) | 0 to 20 h | 7 h |
| Carbon files per year | 100 to 3000 files | 1100 files |
| Result at default settings | Value |
|---|---|
| Hours per file, manual | 22.00 h |
| Hours per file, AI-assisted | 8.25 h |
| Hours saved per year | 15,125 h |
| FTE saved (1,600 hours each) | 9.45 FTE |
| Expected errors per file, manual | 1.20 errors |
| Expected errors per file, AI-assisted | 0.51 errors |
Try it. Set the spot-check time to zero: the file takes 7.9 hours, but nobody has looked at 70% of the values. The half-minute per point is the price of being able to say that a human validated every figure.
Calculation. Manual: 60 × 15 / 60 = 15 hours of keying plus 7 of analysis, 22 hours. AI-assisted: 60 × (0.30 × 3 + 0.70 × 0.5) / 60 = 1.25 hours of review plus 7 of analysis, 8.25 hours, the target used in 5.1. Over 1,100 files the saving is 15,125 hours, about 9.45 FTE. Expected errors per file fall from 1.20 (2% of 60 points) to 0.51 (60 × (0.70 × 1% + 0.30 × 0.5%)).
Reading. The economics rest on two numbers that must be measured in a pilot, not assumed: the share of low-confidence points and the residual error on high-confidence points. If the model is poorly calibrated (it is confident and wrong), the spot-check catches little and errors rise. Meridian’s model risk management therefore classifies the tool as a model, requires a validation on a labelled sample of reports before go-live, monitors the error rate on the spot-checked points monthly, and keeps the page reference as evidence for each value so that an auditor can re-perform the check. The tool does not take credit decisions on natural persons and is not a high-risk system under the AI Act; it is still a model, an ICT service under DORA if hosted by a provider, and a source of figures that end in published reports. Its use is disclosed in the methodology note of the bank’s financed emissions.
INTUITION — The extraction model is a fast junior who has read the whole report and marks every number with a sticky note and a page number. You still sign the file; you just stop retyping it.
The other tools in one line each. The questionnaire moves from a spreadsheet to a portal form whose fields are the CDEs, with validation at entry (units, years, ranges). The taxonomy engine stores the technical screening criteria and the bank’s evidence for each use-of-proceeds asset; its output is a field on the facility and the asset, never a manual flag. The emissions calculator reads counterparty emissions and exposures from the masters and returns attributed emissions with their score; it must be fed, not typed into.
Question. Does the programme earn the bank’s 12% hurdle, and how long does it take to pay back?
Costs (case assumptions). The one-off budget is EUR 14.0m over three years: EUR 5.0m in year 0, EUR 6.0m in year 1 and EUR 3.0m in year 2. It breaks down by workstream as follows: counterparty and facility masters and the KPI register EUR 4.5m; client portal questionnaire EUR 1.5m; taxonomy engine and emissions calculator integration EUR 2.5m; extraction tool with model validation EUR 1.8m; data quality and lineage tooling EUR 2.0m; change management, training and programme office EUR 1.7m. Run costs (licences of the taxonomy engine, model hosting, application support) are EUR 0.5m in year 1 and EUR 1.1m a year from year 2.
Benefits at full run rate (pre-tax). Labour EUR 2.434m (5.1); vendor rationalisation EUR 0.70m (one of the three overlapping ESG data licences is cancelled once the extraction tool covers the data it supplied); error correction EUR 1.095m (5.3). Total EUR 4.229m a year. In the programme’s phased plan, benefits reach 25% of run rate in year 1, 60% in year 2 and 100% from year 3.
Conventions. Cash flows are taken after tax at 25%, treating all costs as deductible when incurred (a simplification; capitalised software would be deducted through amortisation, see 6.4). They are discounted at the 12% after-tax hurdle over years 0 to 7. Payback is the time until cumulative undiscounted cash flow turns positive, interpolated within the year.
FORMULA
is the tax rate (25%); the share of run-rate benefits achieved in year ; the annual benefits at run rate; the one-off cost in year ; the run cost in year ; the hurdle rate (12%). Payback is the first at which .
INTERACTIVE LAB — Business case of the ESG data programme: NPV and payback
Compute the after-tax cash flows, NPV at the hurdle rate and payback of the programme under three delivery plans, with and without an expected avoided supervisory capital add-on.
| Parameter you can move | Range | Default |
|---|---|---|
| Labour benefit at run rate (5.1) | 0 to 6 EUR m | 2.434 EUR m |
| Vendor rationalisation | 0 to 3 EUR m | 0.7 EUR m |
| Error-correction benefit (5.3) | 0 to 3 EUR m | 1.095 EUR m |
| One-off cost year 0 | 0 to 15 EUR m | 5 EUR m |
| One-off cost year 1 | 0 to 15 EUR m | 6 EUR m |
| One-off cost year 2 | 0 to 15 EUR m | 3 EUR m |
| Run cost year 1 | 0 to 3 EUR m | 0.5 EUR m |
| Run cost from year 2 | 0 to 3 EUR m | 1.1 EUR m |
| Delivery plan | phased / bigbang / wave | phased |
| Extra year-1 cost of interim interfaces (wave plan only) | 0 to 3 EUR m | 0.5 EUR m |
| Probability of a supervisory capital add-on avoided | 0 to 50 % | 0 % |
| Size of that add-on | 0 to 50 bp of RWA | 5 bp of RWA |
| Meridian total RWA | 50 to 300 EUR bn | 118 EUR bn |
| Hurdle rate after tax | 5 to 20 % | 12 % |
| Tax rate | 0 to 40 % | 25 % |
| Result at default settings | Value |
|---|---|
| Expected avoided capital cost, pre-tax equivalent | 0.000 EUR m |
| Total benefits at run rate, pre-tax | 4.229 EUR m |
| Annual net cash flow at run rate, after tax | 2.347 EUR m |
| NPV at the hurdle rate, years 0-7 | -1.585 EUR m |
| Payback (undiscounted) | 5.84 years |
| Cumulative cash flow at year 7 | 2.730 EUR m |
| Add-on probability at which NPV is zero | 6.26 % |
| After-tax cash flow year 0 | -3.750 EUR m |
| After-tax cash flow year 1 | -4.082 EUR m |
| After-tax cash flow year 2 | -1.172 EUR m |
| Capital of the add-on if imposed | 59.0 EUR m |
| Annual after-tax cost of that capital at the hurdle | 7.08 EUR m |
Try it. Switch to the wave plan and set the add-on probability to 10%: the NPV turns clearly positive. Then set the labour benefit to 1.7 (30% less): the programme no longer pays back within seven years whatever the plan.
Calculation (phased plan, efficiency benefits only). After-tax cash flows are: year 0, −5.0 × 0.75 = −EUR 3.75m; year 1, (0.25 × 4.229 − 6.0 − 0.5) × 0.75 = −EUR 4.08m; year 2, (0.60 × 4.229 − 3.0 − 1.1) × 0.75 = −EUR 1.17m; from year 3, (4.229 − 1.1) × 0.75 = EUR 2.347m a year. Discounted at 12%, the NPV over years 0 to 7 is −EUR 1.585m. Cumulative cash turns positive after 5.84 years and stands at EUR 2.73m at year 7.
Reading. On efficiency alone, the programme does not earn the hurdle. This is the most common result of honest ESG data business cases, and the change team says it plainly rather than inflating the benefits. Three questions follow, and they are the real content of the steering committee: is there a cheaper plan, is there a benefit we have not counted, and what happens if we do nothing? Section 5.6 answers them.
PITFALL — The tempting fix is to add “revenue enabled”: the sustainable finance volumes that credible data protect. Meridian’s sustainable lending earns real margins, but attributing a share of them to a data programme is guesswork, and steering committees learn to discount such lines to zero. Keep revenue effects in the narrative and out of the NPV unless a business head signs them in the benefits register.
Question. Can the order of delivery or a quantified regulatory benefit change the answer, and what should the steering committee decide?
Three delivery plans. The phased plan of 5.5 delivers the masters first and the tools later. A big-bang plan delivers everything at the end of year 2, with no benefit before year 3: NPV −EUR 3.81m, payback 6.99 years. A wave plan reorders the same budget so that the two largest levers arrive first: the extraction tool and the KPI register in the first six months, the questionnaire and the masters next, lineage tooling last. Benefits reach 40% of run rate in year 1 and 80% in year 2, at the price of EUR 0.5m of interim interfaces that will be thrown away. NPV −EUR 0.99m, payback 5.52 years.
INTUITION — A wave plan is like renovating a house room by room while living in it: you pay for some temporary plumbing, but you start enjoying the new kitchen a year earlier. A big bang means renting elsewhere until the whole house is done.
The regulatory argument, quantified. The case for the programme is not mainly efficiency. The ECB has made RDARR remediation and full alignment with its climate expectations priorities for 2026 to 2028, the EBA guidelines have applied since January 2026, and the supervisor can respond to persistent deficiencies with binding measures, including a higher Pillar 2 requirement. A business case can price this as an expected cost avoided, with explicit and modest assumptions: a 10% probability that, without the programme, Meridian receives a Pillar 2 add-on of 5 bp of RWA. On EUR 118bn of RWA that is EUR 59m of extra capital, which at the 12% after-tax hurdle costs EUR 7.08m a year after tax, or EUR 9.44m pre-tax (7.08 / 0.75); weighted by 10%, the expected avoided cost is EUR 0.944m a year pre-tax.
FORMULA
is the expected avoided cost per year, pre-tax equivalent; the probability of the add-on without the programme; the bank’s total risk-weighted assets; the add-on in percent of RWA; the after-tax hurdle; the tax rate. The term converts the after-tax cost of holding capital into a pre-tax figure comparable with the other benefits.
With this line, run-rate benefits become EUR 5.173m. The phased plan’s NPV turns to +EUR 0.946m (payback 4.75 years) and the wave plan’s to +EUR 1.75m (payback 4.43 years). The simulator also gives the break-even: the phased plan reaches zero NPV at an add-on probability of 6.26%.
Robustness. The labour saving is the largest and the most uncertain line. If it comes in 30% below estimate (EUR 1.70m instead of 2.43m), the phased plan’s NPV falls to −EUR 3.54m and the programme does not pay back within seven years. The change team therefore attaches the labour benefit to named budget reductions and makes the wave plan’s first release, the extraction tool, a test of the largest assumption: if the pilot does not reach the 30% low-confidence share and the 1% residual error, the programme is re-planned before most of the money is spent.
What the target model does to the organisation. The 22 FTE of 5.1 are not spread evenly. The ESG data team loses most of its keying work and gains review, model monitoring and client engagement on activity data for mid-sized borrowers: fewer analysts, but more senior ones. Credit operations and loan agency lose re-keying and reconciliation and gain a KPI register to maintain. The reporting team loses period-end reconciliations and gains analytical review. Two new roles appear: a product owner for ESG data in the CDO, who decides the order of releases, and a model owner for the extraction tool, who answers to model risk management. The change team’s people plan therefore combines redeployment, non-replacement of leavers and some hiring of data skills; presenting it as a pure headcount reduction would be both inaccurate and a fast way to lose the cooperation of the teams whose knowledge the new model needs.
The decision. The steering committee approves the wave plan, with three conditions: the benefits register signed by the heads of ESG data, credit operations and loan agency; a stage gate after the extraction pilot; and the regulatory benefit kept as a separately reported line, because it is a probability and not a cash saving. The CFO’s comment in the minutes is the right one for a student to remember: “We are not buying efficiency; we are buying the ability to prove our numbers, and the efficiency pays for most of it.”
Question. Who controls the new chain, and how large must Internal Audit’s sample be to give assurance on a key control?
The control allocation. The target model assigns each control to a line, following the EBA guidelines (paragraphs 56 to 59) and the bank’s internal control framework.
| Control | Line | Owner | Evidence |
|---|---|---|---|
| Four-eyes validation of extracted values | First | ESG data analyst | Review log per field |
| Facility label and KPI schedule matched to signed agreement | First | Credit operations | Validation record |
| Margin reset within ten business days of certificate | First | Loan agency | Reset log with document link |
| Monthly quality rules on ESG CDEs | First (CDO) | Data steward | Dashboard, issue log |
| Independent test of key controls; challenge of proxies | Second | Risk control | Test report |
| Review of sustainability claims using the data | Second | Compliance (see E10) | Claims review file |
| Validation of the extraction model | Second | Model risk management | Validation report |
| Audit of ESG data lineage and controls | Third | Internal Audit | Audit report to Audit Committee |
The first line owns data quality; the CDO’s stewards sit in the first line in Meridian’s model because they run operational rules, while the CDO’s policy function and Risk set and test the standards. Compliance enters through the use of the data in claims (see E10). Internal Audit tests the whole, including the second line.
The audit sample. Internal Audit wants to conclude, with 95% confidence, that the deviation rate of the control “margin reset within ten business days” across the 460 labelled facilities is below 5%. In attribute sampling with zero expected deviations, the required sample size satisfies .
FORMULA
is the sample size; the confidence level; the tolerable deviation rate; rounds up to the next integer; the upper bound on the deviation rate that a sample of items with no deviation supports at confidence .
INTERACTIVE LAB — How many facilities must Internal Audit test?
Compute the attribute sample size for a control test with zero expected deviations and the upper bound on the deviation rate supported by a clean sample.
| Parameter you can move | Range | Default |
|---|---|---|
| Confidence level | 80 to 99 % | 95 % |
| Tolerable deviation rate | 1 to 20 % | 5 % |
| Sample actually tested (for the upper bound) | 5 to 300 items | 59 items |
| Population of labelled facilities | 50 to 5000 facilities | 460 facilities |
| Result at default settings | Value |
|---|---|
| Required sample size | 59 items |
| Sample as share of the population | 12.8 % |
| Upper bound on the deviation rate if no deviation is found | 4.95 % |
Try it. Lower the tolerable rate to 2%: the sample jumps to 149 facilities, a third of the population. Tight assurance on a small population is expensive, which is why audit relies on the second line’s testing where it can.
Calculation. $n = / = = $ 59 facilities, 12.8% of the population. If all 59 pass, the deviation rate is below 4.95% with 95% confidence. Had the auditor tested only 30, a clean result would support only 9.50%.
Reading. If even one deviation appears, the conclusion at 5% tolerable rate fails and the sample must be extended or the control rated ineffective. That is why the target model’s automation matters for assurance as well as for cost: a reset that is computed once by the system and derived everywhere has one point of failure to test, and Internal Audit can test the system’s logic plus a smaller sample, instead of 59 manual reset files.
| Figure | Value | Stage |
|---|---|---|
| ESG data effort today / target | 36.75 FTE, EUR 4.04m / 14.62 FTE, EUR 1.61m | 5.1 |
| Labour saving at run rate | EUR 2.434m | 5.1 |
| Weighted PCAF score today / target | 3.14 / 2.50 | 5.2 |
| Reported share; exposure at scores 4-5 | 38% to 55%; EUR 46.0bn to 23.0bn | 5.2 |
| First-time-right today / target | 88.54% / 97.52% | 5.3 |
| Records with an error today / target | 1,100 / 238 | 5.3 |
| Error-correction saving | EUR 1.095m | 5.3 |
| Carbon file hours manual / AI-assisted | 22 / 8.25 | 5.4 |
| Errors per carbon file manual / AI-assisted | 1.20 / 0.51 | 5.4 |
| One-off budget / run cost | EUR 14.0m / EUR 1.1m a year | 5.5 |
| Benefits at run rate, efficiency only | EUR 4.229m | 5.5 |
| NPV phased / big-bang / wave | −1.585 / −3.81 / −0.99 EUR m | 5.5-5.6 |
| Expected avoided capital cost | EUR 0.944m a year | 5.6 |
| NPV phased / wave with that benefit | +0.946 / +1.75 EUR m | 5.6 |
| Break-even add-on probability | 6.26% | 5.6 |
| Audit sample at 95% / 5% | 59 facilities | 5.7 |
A data programme is not a loan, so this section adapts the usual questions: what does the programme do to capital, to the RAROC of the loans it serves, to funding, to the accounts and to the bank’s own sustainability metrics?
Credit RWA do not move. CRR3 (Regulation (EU) 2024/1623) contains no ESG factor in the risk-weight formulas, and better ESG data change neither the F-IRB risk weight of a BBB borrower nor its EAD (see E02). Data can matter indirectly, through internal rating models that include ESG variables, through climate overlays in IFRS 9 provisions (see E01), and through the supervisor’s view of risk management in the SREP.
Operational-risk capital does not fall with fewer errors. Under the CRR3 standardised approach, operational-risk capital is the business indicator component, a function of the bank’s income, and the EU has set the internal loss multiplier to one, so the bank’s own loss history does not reduce the charge. The 1,100 data errors of 5.3 are operational-risk events and must be recorded in the loss database when they cause a loss, but avoiding them does not release capital. Business cases that claim an operational-risk capital benefit for European banks should be challenged.
Software is partly deducted from CET1. Intangible assets are deducted from Common Equity Tier 1, but Delegated Regulation (EU) 2020/2176 relieves software: only the difference between a prudential amortisation (straight line over at most three years) and the accounting amortisation already booked is deducted, and the non-deducted part is risk-weighted at 100%. A capitalised data platform therefore consumes CET1 faster than it is amortised in the accounts, which matters when the programme competes for capital with lending.
Pillar 2. The real capital channel is the one priced in 5.6: persistent RDARR or climate-risk deficiencies can lead to qualitative measures, a higher Pillar 2 requirement or periodic penalty payments. The probability is uncertain and the case keeps it as a separate line.
Question. How much does the cost of ESG data change the RAROC of a typical sustainability-linked loan, and does the programme help?
Data (case assumptions). A EUR 60m five-year drawn sustainability-linked term loan to a BBB borrower; net margin over FTP 110 bp; no fees; cost-income 40%; F-IRB with PD 0.20%, LGD 40% and M 2.5. The ESG data cost per facility is computed bottom-up from the case: today, 56 hours (22 for the carbon file, 30 for the KPI test, 4 for the review section) at EUR 68.75, plus a share of EUR 2.6m of licences over 1,100 clients; in the target model, 22.25 hours plus a share of EUR 3.0m (licences after rationalisation plus run costs). We also show the target with the programme’s EUR 14m amortised over seven years (EUR 2m a year) over the same 1,100 clients. A03 assumed a full recharge of EUR 15,000 per carbon file, which includes overheads and systems; here we isolate the direct cost that the programme changes.
INTERACTIVE LAB — RAROC of a sustainability-linked term loan and the cost of its ESG data
Compute the RAROC of a representative loan with the ESG data cost per facility today, in the target model, and in the target model including the amortised programme cost.
| Parameter you can move | Range | Default |
|---|---|---|
| Loan amount (drawn) | 1 to 200 EUR m | 60 EUR m |
| Net margin over FTP | 30 to 300 bp | 110 bp |
| One-year PD | 0.03 to 4 % | 0.2 % |
| LGD (F-IRB senior unsecured) | 10 to 60 % | 40 % |
| Cost-income ratio on lending revenue | 20 to 70 % | 40 % |
| ESG hours per facility today | 0 to 100 h | 56 h |
| ESG hours per facility in the target model | 0 to 100 h | 22.25 h |
| Cost of one hour | 30 to 150 EUR | 68.75 EUR |
| Licences today | 0 to 10 EUR m | 2.6 EUR m |
| Licences and run costs in the target model | 0 to 10 EUR m | 3.0 EUR m |
| Programme cost amortised per year | 0 to 5 EUR m | 2 EUR m |
| Clients sharing licences and programme cost | 100 to 5000 clients | 1100 clients |
| Result at default settings | Value |
|---|---|
| F-IRB risk weight | 39.02 % |
| RWA | 23.410 EUR m |
| Allocated capital (13%) | 3.043 EUR m |
| Revenue | 0.660 EUR m |
| Expected loss | 0.048 EUR m |
| ESG data cost per facility today | 6,214 EUR |
| ESG data cost per facility, target model | 4,257 EUR |
| ESG data cost per facility, target incl. programme | 6,075 EUR |
| RAROC with today’s data cost | 8.42 % |
| RAROC in the target model | 8.47 % |
| RAROC in the target model incl. programme | 8.43 % |
| Today’s data cost in RAROC points | 15.3 bp |
Try it. Set the loan to EUR 5m: the same data cost now takes several RAROC points. ESG data costs are a fixed cost per client and hurt small tickets, which is why mid-sized borrowers get a lighter questionnaire.
Calculation, line by line (today).
| Line | EUR m | Basis |
|---|---|---|
| Revenue | 0.660 | 60 × 110 bp |
| Operating costs | −0.264 | 40% of revenue |
| ESG data cost | −0.006 | EUR 6,214 per facility |
| Expected loss | −0.048 | 0.20% × 40% × 60 |
| Pre-tax result | 0.342 | |
| After tax (25%) | 0.256 | |
| Allocated capital | 3.044 | 13% × RWA 23.41 (risk weight 39.02%) |
RAROC is 8.42% today. In the target model the data cost per facility falls from EUR 6,214 to EUR 4,257 and RAROC rises to 8.47%; once the programme’s own cost is allocated (EUR 6,075 per facility) it is 8.43%, almost unchanged. Today’s data cost takes 15.31 bp of RAROC on this loan, which is negligible. On a EUR 5m loan with the same terms the data cost takes 1.84 points of RAROC: RAROC is 6.74% today and 7.32% in the target model.
Reading. The sustainable feature changes one line of the RAROC, the data and servicing cost, and that line is a fixed cost per client. For large tickets it is noise, for small ones it is decisive. The operating model is therefore not neutral to the bank’s sustainable-finance strategy: without cheap data capture, sustainability-linked lending to mid-sized companies is uneconomic, and the bank ends up with labels only on large borrowers. The loan stays below the 12% hurdle either way; as B03 showed, such loans are justified by the relationship, and data does not change that.
The programme itself is funded through the cost budget and has no FTP. Its effect on funding comes through tags. Meridian’s treasury (see C15 and E04) issues green bonds backed by an eligible green loan portfolio and grants a small FTP incentive to qualifying loans in some business lines. Both depend on the facility label, the eligible category and the allocation register being right. A loan tagged green by mistake can end up in the eligible pool of a green bond, which is a misstatement to investors; a loan wrongly untagged deprives the business line of its incentive and shrinks the pool. The facility master is the golden source treasury reads, and the reconciliation between the green bond allocation report and the facility master is one of the key controls of the target model.
Capitalise or expense. Under IAS 38, internally developed software is recognised as an intangible asset only in the development phase, once technical feasibility, intention and ability to complete and use it, probable future economic benefits, adequate resources and reliable measurement of cost are demonstrated; research-phase costs, training and most change management are expensed. For cloud services, the IFRS Interpretations Committee concluded in March 2019 that a software-as-a-service arrangement is usually a service contract, not an intangible asset, and in its April 2021 agenda decision that configuration and customisation costs are generally expensed, when the services are received, unless they create a separate asset the bank controls. For Meridian’s budget this means the masters, the KPI register and the in-house extraction tool can be largely capitalised, while the taxonomy engine (bought as a service), its configuration, training and the programme office are expensed. The P&L therefore bears a large part of the cost in years 0 to 2, and amortisation of the capitalised part afterwards, while the business case of 5.5 is built on cash.
IFRS 9. Better counterparty data feed the climate overlays and the significant-increase-in-credit-risk assessment where the bank uses transition scores (see E01). The programme does not change the ECL model, but it changes the evidence auditors can test behind any climate-related adjustment.
Financed emissions. The weighted data quality score improves from 3.14 to 2.50 over three years, and the share of exposure with reported emissions from 38% to 55%. The reported total may rise or fall as proxies are replaced; the methodology note must explain changes due to data rather than to the borrowers’ real emissions, as PCAF recommends when recalculating.
GAR and Pillar 3. The facility master’s use-of-proceeds flag and the counterparty master’s CSRD status and Taxonomy KPIs are the inputs of the GAR (see E06); the sector code in NACE Rev. 2.1 and scope 3 by counterparty are inputs of the revised Pillar 3 sector template. The programme does not raise any ratio: it makes each one traceable.
Internal sustainable finance volume. The classification field set by the sustainable finance office in step 6 of the flow drives the bank’s internal target. One golden field avoids the classic problem of three volumes for the same loan (labelled, Taxonomy-aligned, financed emissions) being reported inconsistently, which B17 described.
Own operations. Model inference and data storage add electricity use to the bank’s own scope 2 emissions; for a programme of this size the effect is small relative to the bank’s buildings, but the CSRD statement covers it and the CDO should be able to estimate it.
| Risk | Example | Control | Owner |
|---|---|---|---|
| Wrong entity or group link | Emissions of a parent attached to a subsidiary | LEI and group validation at onboarding | Client onboarding |
| Untraceable figure | Published financed emissions cannot be traced to contracts | End-to-end lineage of CDEs, quarterly lineage test | CDO |
| Label without evidence | Loan tagged green, no allocation record | Label field blocked until evidence linked | Sustainable finance office |
| Stale data | Emissions of 2022 used in 2026 report | Quality rule on reporting year, ageing report | Data steward |
| Proxy presented as reported | Score 4 estimate shown as client data | Data quality score mandatory and disclosed | ESG data team |
| Margin reset error | Adjustment applied to wrong facility | Derived pricing, four-eyes, reset log | Loan agency |
| Model error in extraction | Confident wrong value accepted | Model validation, spot-check monitoring | Model risk management |
| Vendor failure or lock-in | Taxonomy engine discontinued | DORA register, exit plan, data escrow | Procurement and IT |
| Spreadsheets outside control | KPI targets held in a desktop file | End-user computing inventory and limits | First line, CDO |
| Overstated business case | Benefits never banked | Signed benefits register, post-implementation review | Programme office, Finance |
Red flags for greenwashing that start in the operating model. A reviewer, a compliance officer or an auditor should stop when they see any of these:
WHY IT MATTERS ON THE DESK — Greenwashing is usually discussed as a marketing problem. Inside a bank it is mostly a data problem: the claim was made in good faith on a figure nobody could trace. The control that prevents it is lineage, not a better slogan.
Dated facts (verified November 2026).
Orders of magnitude (qualitative). Large European banks typically employ dozens to a few hundred people on ESG data across functions, and spend millions of euros a year on ESG data licences. We found no reliable public statistic on industry-wide spending, and the case figures of this module are assumptions.
What changed recently. The 2025-2026 simplification wave reduced disclosure (CSRD scope, ESRS datapoints, GAR denominator, Pillar 3 templates) but not risk-management expectations. Language models have changed the cost of reading unstructured reports. Supervisors have moved from expectations to enforcement on both climate risk and RDARR.
Open debates.
INTERACTIVE LAB — Operating model, data and IT for ESG in the credit chain
Check the main concepts and figures of the module.
Q1. In the asset-manager analogy, what is the bank’s equivalent of the securities master?
Answer: B. The securities master holds one golden record per instrument; in a credit chain the counterparty master holds one per legal entity and the facility master one per loan.
Q2. Five manual steps each have 98% accuracy. What is the first-time-right rate of the chain?
Answer: C. 0.98 to the power 5 is 0.904. Errors compound across steps.
Q3. Where should a borrower’s scope 1 emissions be stored in the target data model?
Answer: A. Facilities inherit emissions through attribution; storing them on each loan multiplies copies and errors.
Q4. From which date do the EBA Guidelines on the management of ESG risks apply to institutions other than small and non-complex ones?
Answer: D. They apply from 11 January 2026; small and non-complex institutions have until 11 January 2027 at the latest.
Q5. Why does a 30% fall in the labour benefit matter so much in the case?
Answer: B. The business case is most sensitive to its largest line; the programme makes the extraction pilot a stage gate for that reason.
Q6. An AI tool extracts emissions from corporate sustainability reports. Under the AI Act as amended in 2026 it is:
Answer: C. Annex III covers creditworthiness of natural persons; corporate ESG extraction is not in it, but model risk and DORA still apply.
Q7. What does the value-chain cap introduced by the Omnibus I directive mean for the bank’s ESG questionnaire?
Answer: D. From FY2027 a CSRD reporter may not require more than the voluntary standard from smaller value-chain companies.
Q8. Does avoiding operational errors reduce Meridian’s operational-risk capital under CRR3?
Answer: A. The standardised approach in the EU depends on the business indicator; the loss history does not reduce the charge.
Q9. In the case, why does the efficiency-only business case have a negative NPV?
Answer: B. Run-rate net benefits are EUR 2.35m after tax, but EUR 14m of one-off cost and two years of partial benefits push break-even beyond what a 12% discount rate rewards.
Q10. An auditor wants 95% confidence that a control’s deviation rate is below 5%, expecting no deviations. Minimum sample?
Answer: C. ln(0.05)/ln(0.95) is 58.4, rounded up to 59.
Q11. Under IFRS, configuration costs of a software-as-a-service taxonomy engine are generally:
Answer: D. This follows the IFRS Interpretations Committee agenda decision of April 2021.
Q12. Which is a greenwashing red flag rooted in the operating model?
Answer: A. The label changed in one place and not in another; the published volume overstates labelled lending.
Q13. Why do ESG data costs matter more for a EUR 5m loan than for a EUR 60m loan?
Answer: C. In the case the same cost takes 15 bp of RAROC on EUR 60m and about 1.8 points on EUR 5m.
Q14. As data quality scores improve, financed emissions:
Answer: B. Sector averages can understate or overstate a given borrower; the methodology note must explain data-driven changes.
A bank’s chain for 5,000 facility records has four manual steps with accuracies of 97%, 98%, 95% and 99%. Today 70% of errors are found before publication at EUR 300 each; the rest cost EUR 3,000 each. A redesign brings each step to 99.5% and detection to 90%. Compute the first-time-right rates, the error volumes, the annual costs and the saving.
Solution. Today: FTR = 0.97 × 0.98 × 0.95 × 0.99 = 89.40%; errors = 5,000 × 10.60% = 530; cost per error = 0.7 × 300 + 0.3 × 3,000 = EUR 1,110; annual cost = 530 × 1,110 = about EUR 0.588m. After the redesign: FTR = 0.995 to the fourth power = 98.01%; errors = 99; cost per error = 0.9 × 300 + 0.1 × 3,000 = EUR 570; annual cost = about EUR 0.057m. Saving about EUR 0.532m a year. Note that the cost per error falls by half because of detection, while the volume falls by more than four fifths because of accuracy.
A programme costs EUR 3.0m in year 0 and EUR 2.0m in year 1, with run costs of EUR 0.4m a year from year 1. Benefits at run rate are EUR 2.0m pre-tax, with 50% achieved in year 1 and 100% from year 2. Tax is 25%, the hurdle 12%, and the horizon years 0 to 5. Compute after-tax cash flows, NPV and payback.
Solution. Year 0: −3.0 × 0.75 = −2.25. Year 1: (1.0 − 2.0 − 0.4) × 0.75 = −1.05. Years 2 to 5: (2.0 − 0.4) × 0.75 = 1.20 each. NPV = −2.25 − 1.05/1.12 + 1.20 × (1/1.12² + 1/1.12³ + 1/1.12⁴ + 1/1.12⁵) = −2.25 − 0.9375 + 1.20 × 2.7128 = about +EUR 0.067m. Cumulative cash: −2.25, −3.30, −2.10, −0.90, +0.30; payback = 3 + 0.90/1.20 = 3.75 years. The NPV is barely positive: a one-year delay in benefits would turn it negative.
A bank produces 600 carbon files a year of 80 data points each. Manual keying takes 10 minutes per point plus 5 hours of analysis. With AI-assisted extraction, 20% of points are low confidence and reviewed at 4 minutes each; the rest are spot-checked at half a minute each. Compute the hours per file and the annual saving. Internal Audit then wants 90% confidence that the deviation rate of the review control is below 4%: what sample is needed with no expected deviation?
Solution. Manual: 80 × 10 / 60 + 5 = 18.33 hours. AI-assisted: 80 × (0.2 × 4 + 0.8 × 0.5) / 60 + 5 = 1.6 + 5 = 6.6 hours. Saving per file 11.73 hours, or 7,040 hours a year (4.4 FTE at 1,600 hours). Sample: ln(0.10) / ln(0.96) = 56.4, rounded up to 57 files.
1. What is a golden source and why does it matter for ESG data? A golden source is the one system whose value of a field is authoritative; all other systems copy from it. Without one, the same counterparty can carry different emissions or CSRD status in the CRM, the rating tool and the finance warehouse, and nobody can say which was reported. For ESG data this matters because the same attribute feeds pricing, risk, and published figures. The supervisor and the auditor will ask which source was used, and lineage from the report back to it.
2. How would you explain BCBS 239 to a business manager? It is a set of fourteen Basel principles from 2013 that ask a bank to be able to add up its risks accurately, completely and quickly, and to report them clearly, including in a crisis. In practice it means owned definitions, controlled data quality, documented lineage and limited reliance on spreadsheets. The ECB has made it a supervisory priority with on-site inspections and possible penalties. ESG data falls under it as soon as it is used in risk management or reporting.
3. Where does ESG data enter the credit chain? At origination through the client questionnaire, at onboarding through identifiers and sector, in the carbon file built by the ESG data team, in the credit memo and classification opinion, at booking through the label and KPI schedule, during servicing through compliance certificates and resets, and at reporting where it is aggregated. Each entry point needs an owner and a control. The most damaging errors are early ones, such as a wrong group link, because they propagate to every facility.
4. Your business case shows a negative NPV. Do you kill the programme? Not necessarily: I would first check whether the plan can be resequenced to bring benefits forward, whether there are benefits not yet counted, and what doing nothing costs. Regulatory expectations on ESG risk management and risk data apply regardless, so the real choice is often between ways to comply. I would quantify the regulatory benefit explicitly, as a probability-weighted avoided cost, and report it separately. Then I would make the riskiest assumption a stage gate, so that little money is spent before it is tested.
5. How would you use AI in the ESG data process, and what controls would you require? I would use a language model to extract values from sustainability reports and compliance certificates, with a page reference and a confidence score for each value. A human reviews all low-confidence values and spot-checks the rest. The model is validated before use, its error rate is monitored monthly on spot-checked values, and the evidence is kept for auditors. It is not a high-risk system under the AI Act, but it is a model, and if hosted externally an ICT third-party service under DORA.
6. Who owns data quality in the three lines of defence? The first line owns it: the people and teams who capture and maintain the data, with data stewards running quality rules. The second line, Risk and Compliance, sets standards, tests key controls independently and challenges proxies and classifications. Internal Audit gives independent assurance on the whole framework, including the second line. The EBA guidelines on ESG risks follow the same allocation.
7. Does better ESG data improve the bank’s capital ratio? Not directly in the EU: credit risk weights do not use ESG data, and operational-risk capital under CRR3 does not fall with fewer losses because the internal loss multiplier is set to one. The capital effect is indirect, through avoiding supervisory add-ons for RDARR or climate-risk deficiencies. Capitalised software also consumes some CET1 through the prudential amortisation rule. I would not put a capital release in a data business case.
8. Why might financed emissions rise after a data quality programme? Because proxies based on sector averages per euro of revenue or exposure can understate the emissions of particular borrowers, especially heavy industrial ones. Replacing them with reported figures changes the total without any change in the real world. PCAF asks banks to explain such recalculations. The steering committee and the communications team should know this before the first report.
Standards and principles
EU legislation
Supervisory texts
Commentary on 2026 developments (consulted)
Textbook
Course modules: A03 (carbon file and PCAF), B02 and B03 (labels and SLL mechanics), B17 (loan agency and tagging), E01 (credit risk and climate), E02 (capital), E06 (sustainability reporting), E07 (financed emissions), E10 (greenwashing), E12 (governance and transition plan).