The ONS AI Thematic Account Should Change Board Reporting

Model Intelligence & News

24 September 2026 | By Ashley Marshall

The UK is starting to measure AI as an economic contribution, not a novelty. Boards should take the same approach inside their own businesses.

The ONS has turned AI measurement into a board problem

The Office for National Statistics has started work on an AI thematic account for the UK economy, and that matters more than it sounds. The ONS says AI is increasingly understood as a general-purpose technology that can alter production processes, investment behaviour, labour demand, business organisation and productivity. It also says the current economic statistics framework makes it difficult to isolate AI's impact because AI is not visible enough in standard classifications. For UK leaders, the practical point is simple: the national measurement system is catching up with the kind of evidence boards should already be asking for.

This is not just a government statistics story. It changes the way AI investment should be discussed internally. If the national accounts are moving towards a clearer view of AI production, AI use, supporting infrastructure, applications and services, then business cases based only on software licences and demo screenshots will look increasingly thin. The board question becomes: can we show where AI is changing throughput, quality, cost, risk or resilience in a way that would survive external scrutiny?

The ONS article, last revised on 21 September 2026, is careful to say the work is not official statistics yet. That caution is useful. It shows the measurement problem is still live. A business does not need to wait for final national estimates before improving its own evidence. It can start by separating AI spend from ordinary software spend, mapping where AI is embedded in processes, and recording operational outcomes next to each deployed use case.

What this means in practice is that AI reporting should move out of the innovation slide deck and into the management accounts. If a sales assistant reduces research time, show the hours saved and the conversion effect. If a service assistant changes case handling, show resolution quality, escalation rate and complaint movement. If a coding assistant increases output, show release quality and rework, not just accepted suggestions. The ONS is building a macro view. Boards should build the micro version now.

Adoption is up, but depth is still the missing signal

The strongest evidence for better internal measurement comes from the ONS July 2026 analysis of AI in UK businesses. It found that self-reported AI use among UK businesses with 10 or more employees rose from around 12% in late 2023 to around 35% by June 2026. That is a large movement in less than three years. But the same release shows adoption remains relatively shallow: the average number of AI technologies used per adopting business rose only modestly, from around 1.4 to around 1.6.

This is the gap many leadership teams feel but do not yet measure. AI has spread widely enough that most firms can claim activity, yet not deeply enough for many to show durable operating change. The ONS also reported that AI use varies sharply by industry, with information and communication businesses at 58% and construction at 13%. It also found larger firms are more likely to use AI, with 49% of businesses with 250 or more employees reporting use of at least one AI technology compared with 28% among businesses with 0 to 9 employees.

Those figures should stop UK businesses treating AI adoption as a binary measure. The question is not whether the firm uses AI. It is where, how often, under whose control, with what evidence and with what operational consequence. A company with a dozen casual chatbot users may be less advanced than a company with two well-governed workflows that have clear inputs, controls, outcome measures and owners.

The counterargument is that adoption numbers are enough for now because AI is still maturing. That is partly fair. Early exploration is healthy. But shallow adoption becomes expensive when it is mistaken for transformation. Licences renew, staff time disappears into experiments, risk accumulates in untracked workflows and finance still cannot tell whether AI is moving a real business metric. Depth evidence is the bridge between experimentation and investment discipline.

The national accounts problem has a company level mirror

The ONS explains that standard economic frameworks struggle to capture AI because some AI capability is embedded inside existing software, some is developed internally, some is consumed through cloud services and some has no obvious market price. That is exactly the problem inside many firms. AI is rarely sitting in one clean cost centre. It appears inside Microsoft 365, Google Workspace, CRM platforms, support tooling, analytics tools, developer environments and custom workflow agents. It may be paid for through seats, tokens, infrastructure, consultancy, staff time or bundled product upgrades.

The ONS thematic account is intended to remain consistent with national accounts principles while providing insight into AI's contribution to the economy. Businesses need a similar consistency principle. AI costs and benefits should connect to the same operating metrics leaders already trust. That means mapping AI to business processes, not to vendor categories. A customer service workflow should track handle time, first contact resolution, escalation quality, customer satisfaction and complaint risk. A finance workflow should track cycle time, exception rate, audit evidence and manual review load. A marketing workflow should track production speed, conversion quality and compliance review, not just content volume.

What this means in practice is that the finance team, operations team and technology team need one shared AI measurement register. Each entry should name the workflow, owner, supplier or model, data touched, approval route, expected benefit, control evidence and review date. This does not need to be bureaucratic. A simple register is better than a polished dashboard that cannot explain where the numbers came from.

There is also a governance benefit. When AI is measured through workflow evidence, risk conversations become more practical. Instead of asking whether a model is good or bad in the abstract, leaders can ask whether this specific workflow has sufficient test evidence, audit logging, exception handling and fallback. That is the level at which AI creates or destroys business value.

Government policy is moving towards assurance and evidence

The ONS work sits alongside a wider UK policy shift from AI ambition to evidence. The government's AI Opportunities Action Plan: One Year On update says 38 of the 50 actions had been met by January 2026. It highlights five AI Growth Zones, Isambard-AI in Bristol, a commitment to expand UK compute capacity twentyfold by 2030, and up to GBP 500 million for the Sovereign AI Unit to support UK AI companies. It also says over one million AI upskilling courses had been delivered towards a goal of upskilling 10 million workers by 2030.

Those are national scale commitments, but they point to the same operating lesson for private firms. AI adoption is no longer just about finding a tool and encouraging staff to try it. It is becoming a system of compute, data, skills, assurance, procurement, risk ownership and measurable productivity. The government update also refers to initiatives such as the National Data Library, BridgeAI and the AI Growth Lab, all aimed at moving AI from promise into usable capability.

For UK business leaders, the important policy signal is assurance. The same GOV.UK material and related AI assurance guidance emphasise testing, verification, validation and responsible deployment. If public bodies and national programmes are building evidence into AI adoption, suppliers will increasingly be asked for it, and buyers will increasingly need to understand it. This is especially true for regulated sectors, public sector supply chains and firms handling sensitive customer or employee data.

The common misconception is that assurance slows adoption. Poor assurance does. Practical assurance speeds up good decisions because it gives buyers confidence to scale what works and stop what does not. A business that can show workflow tests, permissions, outcome evidence and review logs will move faster than one that has to rediscover the same risk questions at every approval meeting.

The new metric is contribution, not enthusiasm

The ONS thematic account is useful because it reframes AI as an economic contribution question. That is the same shift boards need to make. Enthusiasm is not a metric. Licence count is not a metric. Number of prompts sent is not a metric. The useful question is whether AI is contributing to business performance in a measurable, repeatable and governable way.

That does not mean every use case needs a complex econometric model. It means every serious AI workflow should have a contribution hypothesis. For example: this assistant should reduce case preparation time by 25% without increasing error rate. This retrieval system should reduce duplicated internal research by two hours per project. This coding workflow should increase release throughput while keeping escaped defects flat or lower. This finance agent should reduce invoice exception handling time while preserving audit evidence. The numbers may start as estimates, but they should become measured observations.

A good contribution metric also includes the cost of control. If a workflow saves two hours but creates one hour of review, remediation or audit work, the net benefit is smaller than the demo suggests. If a low-cost model produces outputs that need frequent correction, its true cost may be higher than a more expensive model with fewer exceptions. This is why unit economics and assurance belong together. The cheapest workflow is not the one with the lowest token price. It is the one that completes useful work reliably at the lowest total cost and risk.

What this means in practice is that AI steering groups should stop reviewing use cases as isolated proposals. They should review a portfolio with comparable contribution evidence: cost per completed task, time saved, error movement, customer impact, staff impact, risk status and next decision. That makes AI investment legible to the board and credible to finance.

Build your own AI thematic account before the market asks for it

The practical response is not to copy the ONS methodology. Most businesses do not need a national accounting framework. They need a lightweight internal version that makes AI visible. Start with four lists: AI-enabled workflows, AI suppliers and models, AI costs, and AI outcomes. Then connect them. If a workflow uses Copilot, ChatGPT Enterprise, Gemini, Claude, a specialist vendor model or a custom agent, record the supplier, data category, owner and business process. If the workflow has a benefit claim, record the evidence standard and the latest measured result.

Next, separate experimentation from production. Experiments can be looser, but production workflows need named owners, access boundaries, logging, fallback plans and review dates. The moment a workflow touches customer data, employee data, regulated advice, financial decisions or operational commitments, it needs stronger evidence. That evidence should include test cases, failure modes, human approval points and a record of material changes to prompts, models or connected tools.

Finally, make AI visible in the regular management rhythm. A monthly AI contribution pack should be short enough to read and specific enough to act on. It should show what has gone live, what has been stopped, where spend is rising, where outcomes are improving, which risks need a decision and which workflows are ready for wider rollout. This is how AI becomes an operating capability rather than a collection of side projects.

The ONS has signalled that AI's economic contribution deserves a clearer account. UK businesses should take the hint. The firms that can explain their AI contribution in operational, financial and assurance terms will have an advantage with boards, investors, customers and suppliers. The firms that can only say they are using AI will find that answer less convincing every quarter.

Frequently Asked Questions

What is the ONS AI thematic account?

It is ONS work to develop a clearer account of AI's contribution to the UK economy, including production, use, supporting infrastructure, applications and services. The September 2026 article says the work is not official statistics yet.

Why should a private business care about a national statistics project?

Because the same measurement problem exists inside companies. AI is spread across tools, cloud services, internal workflows and staff time, so boards need a clearer view of contribution, cost and control.

What should boards measure first?

Start with AI-enabled workflows, owners, suppliers or models, data touched, expected benefit, actual outcome, control evidence and next review date. Keep it practical and tied to existing business metrics.

Is AI adoption still shallow in UK businesses?

ONS July 2026 analysis says AI use among businesses with 10 or more employees rose from around 12% in late 2023 to around 35% by June 2026, but the average number of AI technologies used per adopting business only rose from around 1.4 to around 1.6.

Does this mean every AI use case needs a formal ROI model?

No. Early experiments can stay lightweight. Production workflows should have a contribution hypothesis and measured evidence, especially when they affect customers, staff, regulated work or material cost.

How does assurance fit into AI measurement?

Assurance shows whether the workflow can be trusted at scale. Test cases, audit logs, exception handling, access controls and fallback plans should sit alongside cost and productivity evidence.

What is the biggest mistake leaders make with AI metrics?

They measure activity instead of contribution. Licence count, prompt volume and demo volume do not show whether AI improved throughput, quality, cost, risk or resilience.

What should an AI contribution pack include?

It should show live workflows, stopped experiments, spend movement, measured outcomes, risk status, material model or prompt changes, and decisions needed from leadership.