Ask a UK bank, insurer or asset manager how much electricity its AI estate used last year, and the first problem is deciding who should answer. The Bank of England and FCA survey on artificial intelligence in UK financial services, published on 26 October 2026, shows where firms stand: 57% of full-survey respondents said they consider the sustainability impacts of AI, and 82% said they do not currently report those impacts externally. Of the firms not reporting, 36% planned to start within the next two years.
That gap is about to matter. The same survey found that 86% of respondents now use AI, up from 75% in 2024, and that generative AI accounts for 56% of all reported AI systems. Meanwhile the disclosure rules that sit around a financial firm are moving: new UK sustainability standards, new FCA rules for listed companies, and a client-data duty for asset managers finalised in September 2026. None of them sets an AI-specific requirement. All of them can end up containing AI.
This guide is written for the Chief Data and AI Officer who will be asked for the number, and for the CFO and sustainability lead who will have to sign the disclosure it lands in. It covers financial-services AI sustainability disclosure: what the survey found, which rules can capture AI, where the emissions sit, how to measure AI you buy, and what a defensible approach contains. It deliberately cedes per-call energy and carbon for voice workloads, and the global energy picture, to our guide on the voice AI carbon footprint, and the wider survey findings on regulatory constraints to our 2026 survey guide on AI regulatory constraints.
This guide is shipped by the team behind DATS, the five-stage AI consulting system from DILR.AI, delivered by senior practitioners who ship code, not decks. Or see the AI operating model, which covers governance, RACI and lifecycle, audit-ready by design.
What did the 2026 Bank of England and FCA survey find on AI sustainability?
The 2026 Bank of England and FCA survey on AI in UK financial services found that 57% of full-survey respondents consider the sustainability impacts of AI, while 82% do not report those impacts externally. Of the firms not reporting, 36% planned to begin within two years. The survey received 196 responses, and these sustainability figures are based on valid answers from the firms that completed the full questionnaire, which 105 did.
The denominators matter, so state them precisely when you repeat the figures internally. The survey had 196 respondents, 101 larger firms and 95 smaller ones, and firms with fewer than 100 employees could opt into a shorter questionnaire. The sustainability questions sit in section 9 of the full survey, and the report says its percentages use valid question-level responses only. The 36% figure has a different base again: it is a share of the firms that do not currently report externally, not a share of all respondents. Collapsing those into one sentence such as "a third of finance firms will report AI emissions by 2028" would misstate the evidence.
The report is also explicit that its findings describe participating firms rather than the whole UK financial sector. That caveat is a gift to a board paper. It means the right internal claim is narrow: among firms that answered, considering AI's footprint is now the majority position and reporting it is the minority one. Where your firm sits on that line is a governance fact you can establish in a week.
The same report also records that 54% of respondents named applying model risk management principles to vendor-supplied models as a governance uncertainty or implementation challenge. We read the two findings together: a firm that struggles to see inside a vendor's model for risk purposes is unlikely to hold that vendor's energy data either, which is why the measurement section below starts with supplier data. Our guide to AI agents and the autonomy ladder in financial services covers the oversight side of that same dependency.
Which UK disclosure rules could capture the footprint of AI?
None of the main UK disclosure texts sets an AI-specific requirement, but three regimes can capture the footprint of AI: the UK Sustainability Reporting Standards, the FCA's new rules for listed companies under PS26/19, and the FCA's climate rules for asset managers and other regulated asset owners. Each reaches AI indirectly, through electricity, purchased services and emissions data a financial firm already has to account for.
UK SRS. The UK government published UK SRS S1 and UK SRS S2 on 25 February 2026: S1 sets general requirements for disclosure of sustainability-related financial information and S2 covers climate-related disclosures. Neither PDF contains the words artificial intelligence or AI. What reaches AI is S2's greenhouse gas requirement: an entity discloses its absolute gross emissions classified as Scope 1, Scope 2 and Scope 3, measured under the GHG Protocol, and the electricity and purchased services behind AI sit inside those totals.
Listed companies. The FCA finalised PS26/19 on 30 September 2026. In the FCA's words, it replaces the existing climate-reporting rules for listed issuers with requirements aligned with the UK-endorsed ISSB standards, so listed companies report against UK SRS on a comply or explain basis, for accounting periods starting on or after 1 January 2027, with first reporting in 2028. Transitional reliefs give one year for Scope 3 emissions disclosures and two years for wider UK SRS S1 disclosures. The duty binds the listed company, not its AI vendors, so where a listed bank or insurer reports Scope 2 and Scope 3 emissions, the electricity and services behind its AI sit inside its own report rather than its suppliers'.
Asset managers and asset owners. Under PS21/24, TCFD-aligned rules applied from 1 January 2022 to the largest in-scope firms and a year later to firms above the £5 billion exemption threshold. The FCA's sustainability reporting requirements page records that in-scope firms publish an annual TCFD entity report. Handbook Notice 144, published in September 2026, streamlined the product-level rules: firms must provide, at minimum, Scope 1, 2 and 3 greenhouse gas data to clients who request it to meet their own climate-related reporting obligations, once per calendar year per product. The retail-communication rules came into force on 25 September 2026 and the rules for communicating with institutional clients apply from 30 June 2027. That duty binds the asset manager, and the data concerns the products it manages rather than its own server rooms, so the AI question there is narrower than it first looks.
The pattern across all three is the same. The obligation sits with the regulated firm, it is expressed in emissions and electricity rather than AI, and AI enters the numbers only if someone has drawn the boundary to include it. That is a measurement and ownership question long before it is a reporting one, and ownership, RACI and lifecycle are what an AI operating model covers.
The EU adds one obligation that sits further up the supply chain. Under the EU AI Act, providers of general-purpose AI models must keep technical documentation listed in Annex XI. Point (e) of Section 1 requires, verbatim, "known or estimated energy consumption of the model. With regard to point (e), where the energy consumption of the model is unknown, the energy consumption may be based on information about computational resources used." That text, published on the European Commission's AI Act service desk, binds the model provider, not the bank or insurer deploying the model. It is still useful context for a buyer: a provider in scope has to document a known or estimated energy figure, possibly derived from compute rather than metering, so it is reasonable to ask a supplier whether that figure exists and whether it can be shared.
Where do AI emissions sit in a financial firm's carbon inventory?
AI emissions in a financial firm's carbon inventory sit in different scopes depending on how the AI is run. Electricity for servers the firm owns and operates falls in Scope 2. Cloud AI services, hosted models and vendor platforms the firm pays for fall in Scope 3, most naturally under purchased goods and services. A firm that runs both splits its AI footprint across scopes by default.
The GHG Protocol Corporate Value Chain (Scope 3) Standard defines 15 categories of Scope 3 activity, upstream and downstream, and the first upstream category is purchased goods and services. Scope 2 covers purchased electricity for the firm's own use. On that reading, a hosted model the firm pays for is a purchased service whose electricity is the supplier's own Scope 2, which is the uncomfortable part for a sustainability team: with generative AI now 56% of all AI systems reported in the 2026 survey, a large part of the estate may depend on supplier data the firm does not control.
For context on scale, the IEA's Energy and AI report estimates that data centres used around 1.5% of the world's electricity in 2024, about 415 TWh, and projects around 945 TWh by 2030, with AI the most important driver. That is a global data centre figure, not a figure for financial services, and our voice AI energy and carbon guide covers what it does and does not mean for an individual workload. Use it to explain why the question is on the agenda, never as a proxy for your own footprint.
For a bank, insurer or asset manager, the inventory work splits into three buckets: AI on owned infrastructure, AI on rented cloud capacity the firm configures itself, and AI consumed as a finished service inside a vendor product. Each bucket needs a different data source, and only the first is fully within the firm's own metering. Our guide to private LLMs for regulated institutions covers why some firms move workloads into the first bucket for control reasons; a side effect is that the energy figure can become measured rather than estimated.
How do you measure the energy use of AI you buy rather than build?
Measuring the energy use of AI a financial firm buys rather than builds starts with the supplier, not the meter. The firm asks each AI vendor and cloud provider for workload-level energy or emissions data, states which method it used, and fills gaps with documented estimates based on compute or spend. The disclosure then labels each figure as measured, supplier-reported or estimated.
Two of the large cloud providers publish customer-facing carbon tooling. Google Cloud Carbon Footprint states that it provides Scope 2 market-based and location-based emissions alongside Scope 1 and Scope 3 data, and that its data meets the GHG Protocol. The AWS customer carbon footprint tool now resolves to the AWS Sustainability Console. Firms on Azure should ask their account team what equivalent emissions data is available. Check the granularity each tool reports at before relying on it: if AI workloads are not tagged separately in the cloud estate, no report can separate them afterwards.
For AI delivered inside a vendor product, the data is thinner. Ask four questions in procurement and at renewal:
- Can you provide energy or emissions data for our usage, and at what granularity?
- Which method produced it: metered, allocated from a facility total, or estimated from compute?
- Is the electricity figure location-based, market-based or both?
- Which model sits behind the service, and has its provider documented an energy estimate you can see?
Where a vendor cannot answer, the firm estimates. ISO/IEC 21031:2024, is the Software Carbon Intensity specification, and the Green Software Foundation's SCI guide lists the functional unit among the core components of the calculation. That suits AI, because the natural unit for a financial firm's AI is a task, a document or a call rather than a server. Whatever method you choose, write it down before you calculate, because the method is what an auditor or a client will test first.
Running cost and energy use share one driver, the compute an AI workload consumes, so a firm that already tracks what its AI costs to run has a head start on the inventory. AI cost control is one of the six enterprise AI solutions we describe.
What does a defensible AI sustainability disclosure contain?
A defensible AI sustainability disclosure for a financial firm contains four things: a stated boundary saying which AI workloads are included, the method used for each figure, the split between measured, supplier-reported and estimated numbers, and the known gaps with a date for closing them. A short disclosure with those four elements survives challenge better than a precise-looking total with none of them.
Comply or explain, the approach PS26/19 applies to listed companies, suits this shape. A firm that cannot yet produce a reliable AI figure can say so, explain why, and state the steps it is taking. That is a stronger position than publishing a number built on an unlabelled estimate, which is the disclosure equivalent of a model with no validation record.
The four-step approach below is the one we would run before anyone asks for the figure.
Inventory AI workloads. List every AI system in production by owner, purpose and how it is run. A bank or insurer often holds part of this list already for model risk or data protection reasons; the sustainability version adds where the compute happens. Our guide to PRA SM&CR accountability for AI covers the ownership side of the same inventory.
Request supplier data. Send the four procurement questions above to every material AI supplier and record the answers, including refusals. A refusal is evidence too: it tells you which line items will be estimated.
Map to scopes. Assign each workload to Scope 2 or the relevant Scope 3 category, and keep the mapping in the same place as the inventory so it updates when a workload moves from vendor to owned infrastructure.
Disclose with method. Write the disclosure note in plain language, with the boundary, method and gaps stated. If the honest position is that the firm does not yet report AI separately, say that and give the date by which it will.
Keep AI in proportion when you brief the board. For a lender or investor, the climate questions attached to what it finances are a separate disclosure exercise, and nothing here suggests the AI estate is the biggest line in the report. The case for doing this work now is narrower: in the survey, 36% of non-reporting respondents planned to report externally within two years, and the inventory behind an AI figure is the same one a firm needs for AI governance anyway.
Who in a bank, insurer or asset manager should own the AI footprint number?
The AI footprint number in a bank, insurer or asset manager should be owned jointly but signed by one person. The Chief Data and AI Officer owns the inventory and supplier data, the sustainability or finance team owns the method and the disclosure, and a named senior manager signs it. Without a single signatory, nobody answers for the figure when it is challenged.
That split follows the logic of model risk management. Principle 1 of the PRA's SS1/23 on model risk management for banks expects in-scope banks to have a model inventory and a risk-based tiering approach, and in the 2026 survey 54% of respondents named vendor-supplied models as a model risk management challenge or uncertainty. An AI footprint is another attribute of the same inventory, so it belongs with the people who already maintain it, not in a parallel spreadsheet.
The practical RACI looks like this:
| Activity | Responsible | Accountable |
|---|---|---|
| AI workload inventory | Chief Data and AI Officer's team | Chief Data and AI Officer |
| Supplier energy and emissions data | Procurement and vendor management | Chief Data and AI Officer |
| Method and scope mapping | Sustainability or finance team | CFO |
| External disclosure note | Sustainability or finance team | Named senior manager signing the report |
Our AI execution office is embedded delivery with production placements the client owns, and the DATS methodology puts the operating model stage before pilot to production for a similar reason: a number with no owner is hard to defend. For wealth and private markets firms specifically, our guide to AI for wealth, PE and hedge funds maps where AI pays in that sector, and our ongoing advice evidence guide shows the same evidence-first discipline applied to advice delivery.
Training matters here too. A reviewer who does not know how AI is bought and run has little basis for questioning a vendor's headline figure. Dilr Academy, our AI tutor that builds interactive courses on demand, is one option where AI literacy is the gap.
What is the best way to start AI sustainability reporting in 2026?
The best way to start AI sustainability reporting in 2026 is to build the AI inventory and supplier-data request first, publish a short method-led statement second, and only then attempt a quantified AI line. For a UK financial firm, the first useful output is usually an honest boundary and method note, not a carbon total, because a total depends on supplier data the firm may not yet hold.
The choice of partner depends on what you are missing. If the gap is assured, firm-wide sustainability reporting across financed emissions and operations, a large audit and advisory firm such as Deloitte, PwC, EY or KPMG is the more natural fit, because sustainability assurance is a service those firms provide and DATS does not. If the gap is AI-specific, compare AI-focused and engineering-led consultancies, Accenture and Slalom among them, on the three tests below. We concede the first case to the Big Four plainly: DATS is not a sustainability assurance provider.
Where DATS fits is the AI side of the problem. The placement diagnostic within DATS AI consulting produces a ranked roadmap of where AI belongs and where it does not, which starts from knowing what AI the firm runs. The operating model stage then covers governance, RACI and lifecycle, which is where ownership of an AI inventory belongs. We hold ourselves to a delivery discipline of three shippable placements a year, with a named owner per ship, which keeps the work small enough to finish. Our enterprise AI consulting guide explains how to compare consultancies on these criteria.
Three tests sort any proposal, whoever writes it:
- Does it produce an AI inventory you can keep, or a one-off report?
- Does it state the method for every figure, including estimates?
- Does it name who signs the number?
Do UK asset managers have to report AI emissions today?
UK asset managers face no FCA climate rule written specifically for AI emissions today. Firms in scope of the FCA's TCFD-aligned rules publish an annual entity report, and product rules finalised in September 2026 require Scope 1, 2 and 3 data for clients who request it for their own reporting. AI enters those figures only through the boundary and data the asset manager uses.
The duty to provide product-level greenhouse gas data applies once per calendar year per product, and the rules for communicating with institutional clients apply from 30 June 2027, according to the FCA's Handbook Notice 144. Check with your compliance team which products and clients are in scope before assuming the data request will reach AI at all.
Is AI energy use material for a financial firm?
Whether AI energy use is material for a financial firm is a judgement each firm has to make and document, not an assumption to skip. The 2026 Bank of England and FCA survey found that 57% of full-survey respondents already consider the sustainability impacts of AI, so a documented view, either way, puts a firm alongside the majority of respondents rather than behind them.
Scale is the reason to keep AI in proportion: UK investment management alone held £10.0 trillion of assets under management in 2024, according to the Investment Association, and the climate questions attached to those assets are a separate disclosure exercise from an AI estate. Proportion is not a reason to ignore AI, though. Our guide to the evidence on AI and headcount takes the same proportionate approach to another AI number boards ask for.
Want to put an owner and a method behind your AI footprint? See the DATS system, read about our approach to placing AI inside regulated firms, browse more industry AI guides, or learn about Dilr.ai and the practitioners behind it.
Know your AI estate before anyone asks.
30-min scoping call · No deck · Confidential. We will tell you whether DATS fits, and what an owned AI inventory would take.
Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.