Benjamin Guin
Most discussion of AI in organisations focuses on productivity: as analysis becomes cheaper, decision-making should improve. This post argues that an important consequence lies elsewhere. Where the relevant information is accessible, a wider range of people can produce professionally presented analysis and submit it to decision-makers. Responsibility for decisions, however, remains concentrated and changes more slowly. Analysis may therefore proliferate faster than those accountable for decisions can assess it. The comparative advantage of institutions may increasingly lie in deciding what evidence to trust and act on, rather than in producing more analysis. This post explores that shift in settings where many people can generate evidence but few remain accountable for decisions, such as regulators assessing consultation responses or other forms of public input.
The result is a widening gap, illustrated schematically in Chart 1 as the two blades of a pair of scissors. The falling blade is the cost of producing analysis. The flatter blade is accountable attention: the capacity of those responsible for a decision to assess the evidence, decide what to rely on and justify that choice. While analytical capacity can be expanded through staff or technology, responsibility for the final judgement remains attached to specific individuals. Formal responsibilities can change, and staff and technology can help process evidence. But accountable attention is unlikely to expand as quickly as the supply of analysis. This creates a potential gap between what can be produced and what can inform accountable decisions. This suggests a possible shift in the binding constraint on organisational decision-making, from producing analysis to deciding what evidence to trust and act on.
Chart 1
Note: Stylised illustration; not drawn to data.
The falling blade: evidence production
The production side is by now well documented. Controlled studies find AI assistance delivers productivity gains, speeding up professional writing and analytical tasks by 25% to 40%, while the cost of obtaining a given level of AI capability has fallen dramatically. The Bank of England and Financial Conduct Authority’s 2026 survey found 86% of UK financial services firms already using AI, up from 58% in 2022. Its use also extends beyond teams specialising in data analysis and modelling into compliance, legal and operational functions. Where the underlying information is available, the same shift can extend beyond the organisation: Individuals can now use general-purpose AI models to turn that information into a structured, referenced piece of analysis in minutes.
AI has sharply reduced the time and expertise needed to produce polished, well-structured analysis. Professional presentation, references and structure were all once useful, if imperfect, signals of effort and quality because they were costly to produce. As AI reduces those costs, those signals may become less informative for organisations receiving reports, applications and other material. Their challenge is therefore not simply handling more material. It is determining which claims can be independently verified, regardless of how they are produced.
The fixed blade: accountable attention
Attention, on its own, is more expandable, although constrained by resources. Committees hire secretariats, boards commission summaries and regulators expand triage teams. In the short run, accountability is harder to expand or delegate within existing governance arrangements. Under the Senior Managers and Certification Regime, responsibility in UK financial firms is assigned to named individuals through statements of responsibilities, the documents that map specific duties to each senior manager. Those responsibilities can be reassigned through formal processes, but the capacity for accountable decision-making does not automatically expand as more analysis is produced. Directors’ duties do the same for boards, while statutory objectives constrain regulators. Analysis may be produced by an ever wider set of actors, but responsibility for decisions remains attached to named individuals. The production of analysis may therefore change more rapidly than the accountability structures around it.
What is scarce, then, is not attention but accountable attention: the capacity to decide what evidence to rely on and to justify that choice. Committee structures, meeting frequencies and formal responsibilities can change, but not at the pace at which analysis can now be produced. Consider a regulator receiving 100 responses to a rule review rather than ten. Staff or AI can summarise them, but the person responsible for the decision must still decide which claims are reliable enough to influence the final judgement, and be able to explain that choice. Herbert Simon observed half a century ago that a wealth of information creates a poverty of attention. Reading and summarisation can be delegated. Ultimate responsibility for deciding which evidence to rely on cannot.
Put the two blades together and the tension is clear. The production of analysis can be redistributed quickly, while the allocation of accountability changes more slowly. Over time, specialist teams and AI tools may increasingly help screen and verify information. Even then, those responsible for decisions still need to decide which filters to trust and which evidence to rely on.
An illustration: the duty to review rules
The duty to review rules provides a useful illustration of how the mechanism described above might operate in practice. The Financial Services and Markets Act 2023 inserted into FSMA 2000 a statutory duty on the UK financial regulators to keep their rules under review (section 29 – 3RA) and to publish a statement of policy on how they do so (section 29 – 3RB). The Financial Conduct Authority finalised its Rule Review Framework in January 2024; the Prudential Regulation Authority (PRA) published its statement on the review of rules in February 2024. Both include channels through which firms, trade bodies and members of the public can submit evidence that a rule is not working as intended.
When these arrangements were designed, producing a credible, well-evidenced submission typically required analysts, data and drafting time. AI can reduce some of those costs substantially, although access to relevant information remains an important constraint. The cost of generating a submission has therefore fallen sharply, whether the submitter is a firm, trade body or member of the public. Similar concerns have recently been raised elsewhere in the public sector, where AI-generated complaints, appeals and claims have been argued to risk overwhelming administrative processes. By reducing the expertise and resources needed to participate, AI may enable individuals and organisations to contribute evidence and perspectives that might otherwise not have been brought forward. The regulators weighing those submissions, however, continue to operate under the same statutory objectives and accountability arrangements.
The PRA rule review framework illustrates the kind of verification filter that may become more important as credible-looking analysis becomes cheaper to produce. It asks that feedback be ‘supported by relevant and verifiable empirical evidence, where possible’, with underlying assumptions clearly set out. That is a verification filter, not a volume filter. As credible-looking evidence becomes cheaper to produce, greater weight is likely to be placed on evidence that can be verified. Where the relevant information is available, firms, trade bodies and members of the public can now produce structured and credible-looking submissions at very low cost using general-purpose AI models. The volume of submissions received, and the share meeting the verifiability standard, may provide one way of testing the framework proposed here. If the mechanism operates as described, growth in submissions could outpace growth in submissions meeting the verifiability standard.
Why the bottleneck moves: unbundling
Historically, specialist analytical functions did more than produce analysis. They challenged assumptions, reconciled competing interpretations and filtered what reached decision-makers. As analysis is produced by a wider range of actors, it does not automatically receive the same level of challenge, scrutiny and filtering. A submission drafted by a firm’s operations team, a trade body or a member of the public may contain analysis, but it has not necessarily passed through those same processes. Over time, specialist teams and AI tools may increasingly perform that filtering role. Even then, those responsible for decisions still need to decide which filters to trust and which evidence to rely on. If evidence production expands faster than verification and judgement, the bottleneck may increasingly shift from producing analysis to deciding what evidence to rely on.
Doesn’t AI solve this too?
The obvious objection is that AI can integrate evidence as well as produce it, by summarising, comparing and flagging inconsistencies. It can. Some forms of verification may become cheaper. AI can help identify sources, check consistency and synthesise submissions. Harder questions remain about how much weight to place on competing evidence and what is sufficient to justify action. The argument is not that AI cannot support or automate decisions, but that accountability for those decisions often remains assigned to identifiable individuals or institutions. AI may also broaden participation and help organisations process larger volumes of information. Canada’s recent national AI consultation generated tens of thousands of responses from over 11,300 respondents, which the government analysed using large language models. This is where the argument becomes an empirical question rather than a certainty. Everything turns on whether the cost of integrating evidence falls as quickly as the cost of producing it. If it does, institutions may be able to absorb much of the additional volume. Even if AI reduces the volume problem, judgement about what evidence to rely on, and justify, may remain harder to automate.
What this implies
If the argument is right, institutions’ comparative advantage may increasingly lie in organising evidence rather than producing it: allocating accountable attention, ensuring conclusions are challenged before they harden, and demanding evidence that can bear the weight of a signature.
The rule review frameworks can be read as an early example of this shift. They are explicit statements of how evidence for regulatory change will be assessed, including requirements around verification. The emphasis is less on generating more evidence and more on determining what evidence can reasonably support action. More broadly, if credible-looking analysis continues to become cheaper to produce, institutions may need to invest more in the frameworks, processes, people and systems used to assess it.
The familiar organisational debate, centralise or decentralise, may therefore be the wrong axis. Evidence production is becoming easier to decentralise whether institutions choose it or not. Accountability is not. The challenge for institutions is not simply handling a larger volume of evidence. It is building the frameworks, processes and capabilities needed to distinguish credible-looking analysis from evidence that can reasonably support accountable decisions.
Benjamin Guin works in the Bank’s Strategy and Policy Approach Division.
If you want to get in touch, please email us at bankunderground@bankofengland.co.uk or leave a comment below.
Comments will only appear once approved by a moderator, and are only published where a full name is supplied. Bank Underground is a blog for Bank of England staff to share views that challenge – or support – prevailing policy orthodoxies. The views expressed here are those of the authors, and are not necessarily those of the Bank of England, or its policy committees.