On 19 October 2026 Jessica Rusu, chief data, information and intelligence officer at the UK Financial Conduct Authority (FCA), gave one section of a London speech the heading “Becoming an agentic regulator”. The drafted text contains two numbers. The FCA, it says, is building AI tools to oversee 60,000 firms that, under pending legislation, come into its anti-money laundering remit from 2028. It is also building agentic tools such as a “market first responder agent”, which will supervise around 9,000 small firms that submit transaction reports in wholesale markets. Six days earlier Jessica Rusu had told the House of Commons Treasury Committee that the aim was “to automate wherever we can”.

The announcement can be read as a regulator catching up with the firms it supervises. The FCA’s own documents support a more sober reading. What has been announced are tools under construction, not systems in operation, and the 60,000 firms are law firms, accountants and trust providers rather than banks. For banks and investment firms, the agent for transaction reporting matters more, because it would work on a reporting chain in which the FCA already finds errors today. So far, the only public description of how the regulator will control its own AI is one Mills Review recommendation and one line in the annual report, where its Board lists the topic as an area for strengthening. A German draft anti-money laundering bill, in the version dated 28 October 2026, takes a different route: it would set statutory limits on how the German Financial Intelligence Unit (FIU) analyses data, although on individuals rather than in the supervision of firms.

This article draws on the drafted 19 October speech, the uncorrected transcript of the 13 October committee hearing, the Mills Review of July 2026, the FCA’s annual report for 2025/26, its pages on the reform of anti-money laundering supervision and on its own approach to AI, HM Treasury’s October 2025 consultation response on AML supervision and the draft bill of the German Federal Ministry of Finance. The FCA has not published in detail what the agent will check. Where this article draws conclusions for reporting practice, they are the author’s own inferences, not statements by the regulator.

In brief

What: The FCA intends to become an “agentic regulator” and use AI agents in its own supervision. It announced the plan on 6 October, explained it to MPs on 13 October and put numbers to it on 19 October 2026

Numbers: AI tools for around 60,000 professional services firms set to move into FCA anti-money laundering supervision in phases from autumn 2028 under pending legislation, and an agent for around 9,000 small transaction reporting firms

Today: According to Jessica Rusu, a human is involved in every decision, though the FCA states an ambition to automate further

Controls: The Mills Review recommends keeping supervisory AI at advisory and preparatory levels, and the Board lists governance and controls over the FCA’s own AI use as an area for strengthening in the annual report

Comparison: A German draft bill (version dated 28 October 2026) would bar the FIU from self-learning systems that make independent risk assessments of individuals. Its subject is data on individuals, not the supervision of firms

The FCA names two groups of firms but no start date

The idea of an agentic supervisor has a short history. According to his drafted speech, Nikhil Rathi, chief executive of the FCA, told the industry body techUK on 24 June 2026 that the regulator was exploring agentic AI as its “first responder” to speed up the monitoring of wholesale markets. He repeated the line at TheCityUK on 22 September and added that it concerned more than 9,000 firms, there in the context of monitoring for market abuse rather than transaction reporting specifically. On 6 October, according to a report in Health & Protection, Jessica Rusu used the phrase “agentic regulator” at the press conference for the FCA’s Annual Public Meeting. Between the speeches came the Mills Review of July, an inquiry commissioned by the FCA itself, whose sixth recommendation calls for an “AI-enabled agentic supervisory model”.

At the Treasury Committee on 13 October, John Glen MP asked which concrete steps the FCA had taken on that recommendation. According to the uncorrected transcript, Jessica Rusu answered that AI mainly speeds up case officers’ analysis of unstructured intelligence. Then came the sentence that sums up the position: “At this point in time, we remain with a case officer, or sometimes what we call a human in the loop, with every decision, but our ambition is further than that, which is to automate wherever we can, to intervene at scale, and to shift left any harms that we can identify at the gateway.” The statement draws a line between the present, in which a person is involved in every decision, and an ambition without a date.

The 19 October speech draws the same line, and its verbs are chosen with care. For the 60,000 anti-money laundering firms the FCA “are building” tools, so they are still in development. The transaction reporting agent “will supervise” around 9,000 firms, in the future tense and with no date attached. On machine-readable data collection interfaces, the FCA says it is “also working on” them. The text names no budget, no supplier, no metric and no date for going live, and a note on the FCA website says the delivered version may differ. What has been announced is a programme. For a programme, though, the numbers are unusually specific, because they define the population against which the machine will eventually have to be measured.

The 60,000 firms come from law, accountancy and trust services

The larger number rests on a political decision that predates any of the speeches. In its consultation response published on 21 October 2025, HM Treasury decided that the FCA should become the single supervisor of anti-money laundering compliance in the professional services sector. Today 22 professional body supervisors share that task with HM Revenue & Customs (HMRC). The future supervisor, the response says in paragraph 2.15, will be able to take “a risk-based approach across a population of approximately 60,000 regulated firms”. On its reform page the FCA breaks the figure down as an estimate: around 34,000 accountancy businesses and 7,500 law firms currently supervised by professional bodies, and 18,000 accountancy and trust and company service providers supervised by HMRC.

Around 60,000 professional services firms are set to join FCA supervision Regulated firms today and professional services firms in AML supervision from autumn 2028 40,000 regulated today 34,000 accountancy (PBS) 7,500 law firms 18,000 accountancy/TCSP (HMRC) + 60,000 from autumn 2028 in phases, complete around 2030
The FCA says around 40,000 regulated firms deal with it today. From autumn 2028, in phases until around 2030, around 60,000 professional services firms are set to join for anti-money laundering supervision: around 34,000 accountancy firms supervised by professional bodies, 7,500 law firms and 18,000 accountancy and trust and company service providers supervised by HMRC. All figures are FCA estimates, and the total is a sum of two of them, not an FCA figure. The transfer depends on the Financial Services and Markets Bill. Sources: FCA Annual Public Meeting, 6 October 2026, and “AML supervisory reform”, 22 September 2026.

The year given in the speech marks a beginning rather than an end. Under the timeline the FCA published on 22 September 2026, it expects to take on supervision no earlier than autumn 2028, in phases, with the transition complete around 2030. All of this depends on legislation. The Financial Services and Markets Bill completed its stages in the House of Lords in September, and the Commons approved it at second reading on 10 November without a division. Before the committee, Nikhil Rathi gave a smaller figure than his colleague, referring to another “40,000 to 50,000” firms, while at Mansion House on 22 October he spoke of more than 60,000. None of the sources explains the gap. The group still matters to banks. Law firms, accountants and trust providers often act as intermediaries between bank and client, and their supervisor is set to change.

Supervising that many firms without machines would take an enormous number of staff. At its Annual Public Meeting the FCA referred to around 40,000 regulated firms that deal with it. Add the professional services firms and, on FCA estimates, the number of firms more than doubles, though the total is the author’s own sum, not an FCA figure. According to Jessica Rusu’s evidence to the committee, the FCA’s data, technology and innovation function has grown to more than 600 staff. In March 2025 the government also committed to cutting the administrative burden of regulation by a quarter by the end of the Parliament, and it expects greater use of AI, among other things, to make regulators more efficient. Under those conditions the regulator’s move towards automation follows from the arithmetic.

The agent could target a reporting chain in which the FCA already finds errors

For banks and investment firms with UK business, the smaller number is the more important one. The FCA does not define what counts as “small”, and Nikhil Rathi described automated queries as going to reporting firms in general. Transaction reporting is the obligation under Article 26 of UK MiFIR, the retained Markets in Financial Instruments Regulation, to report every transaction in a reportable financial instrument to the regulator. According to its discussion paper DP24/2, the FCA receives “over 7 billion transaction reports a year” in more than 20 million instruments. Its consultation paper CP25/32 adds that fewer than 8% of firms report under more than one regime, and these, often the largest, submit 85% of the data. The 9,000 small reporting firms are therefore likely to account for many senders but little volume. Anyone who has run a trade reporting process knows the errors that arise where a few people are responsible for a great many data fields.

Checking the reports is already largely automated. The FCA’s Market Data Processor validates every incoming report against fixed rules and rejects those that fail. Under Article 15(2) of RTS 22, the regulatory technical standards in Commission Delegated Regulation 2017/590, firms must notify errors promptly, and under Article 15(3) they must regularly reconcile their trading data against extracts provided by the FCA. How the regulator reads those notifications is set out in its newsletter Market Watch 82 of July 2025. In the first quarter of 2025 the FCA received 241 breach notifications. Of these, 15% said nothing about plans for back reporting, and 31% described the firm’s own control weaknesses inadequately. The FCA is unusually frank about why it presses the point: “Delays in back reporting leave us unable to trust data and potentially limit our ability to detect and investigate market abuse.”

A case from summer 2025 shows what happens when the regulator is faster than the firm. On 1 August the FCA fined the broker Sigma Broking £1,087,300 in what its press release describes as the second enforcement action for transaction reporting failures under UK MiFIR. 924,584 reports were incorrect, covering “close to 100% of transactions handled by the firm” between December 2018 and December 2023. The FCA’s systems found the problem: “The FCA's monitoring systems identified issues with the transaction reports submitted by Sigma, which were raised with the firm in May 2023.” The regulator attributes the failure to an incorrectly configured system. An agent would not need to invent this kind of check. It could speed up one that already exists, although the FCA has not said that this is where it will start.

Nikhil Rathi described what such an agent might do when he appeared before the committee on 13 October. One of the first use cases for the FCA’s agentic tools, he said, was to clean its market surveillance data and to go back to the reporting firms, “potentially agentically”, to say: “This does not reconcile. There seems to be an error here.” The FCA might also, he added, interact with the firms’ own agents and automated platforms. No source says explicitly whether this tool is the “first responder” of the speech, and the FCA gives no date for the agent. Another change, by contrast, is fixed. Under policy statement PS26/15, a new reporting format applies from 3 April 2028, with 52 fields instead of 65 and a back reporting period of three years instead of five. The format change is a fixed rebuild date for reporting teams, and a natural point at which to reset reconciliation against FCA extracts.

At this point in time, we remain with a case officer, or sometimes what we call a human in the loop, with every decision, but our ambition is further than that, which is to automate wherever we can […] Jessica Rusu, Chief Data, Information and Intelligence Officer, Financial Conduct Authority, to the Treasury Committee on 13 October 2026

Controls on the FCA’s own machine rest on a recommendation and one line in the annual report

The strongest version of the opposing view comes from the FCA’s own documents. The Mills Review, which recommends agentic supervision, limits it in the same breath: “the FCA should develop AI to the Consultant and Approver levels, not Observer. This would mean that human supervisors remain responsible for key decisions including exercising judgement, overseeing AI usage, and engagement with firms (outside agent-to-agent communication).” Agents are to prepare actions and information requests, and people are to decide. The review also asks that the FCA’s own AI meet the standards the FCA expects of firms. In its statement on internal control in the annual report for 2025/26, which judges the control system adequate overall, the Board names as an area for further strengthening: “Strengthening governance and controls over the use of AI, including assurance, accountability, auditability and data governance”. Nikhil Rathi also pointed out on 22 October that “there are significant mechanisms to hold the FCA accountable, more than for most regulators”. That the Board lists the topic in its own annual report at least shows that it has been noticed.

Naming the issue is not the same as resolving it. As of 6 November, the speeches from June to October, the Mills Review, the annual report and the regulator’s AI page contained no published framework explaining how the FCA detects errors made by its agents, how it explains their findings to firms or who is accountable for a step an agent has prepared. That page says only that staff remain “integral” to judgement while AI pulls out facts and analyses unstructured text. For firms, the FCA has described the risk in more detail. In its 19 October collection of good and poor practice, it warns against human review that turns into a formality. The 27 October article covers that case in detail. Whether the same standard applies to its own case officers is not stated there.

One passage of the Mills Review leaves a question open. The requirement that people remain responsible for engaging with firms explicitly excludes “agent-to-agent communication”. That is precisely the scenario Nikhil Rathi sketched as one of the first use cases, though he qualified it twice with “potentially”. Read literally, the bracket leaves open who is responsible when the regulator’s agent queries a firm’s agent and the firm’s agent replies. Practitioners have commented on the role of people in supervision in more general terms. Nathan Wilmott of the law firm Ashurst Perkins Coie told Insurance Times on 4 November that the FCA “cannot allow AI to become in practice the decision-maker at any stage of the regulatory process”. Jonathan Fong of the Association of British Insurers (ABI) called the move “rational”, provided the FCA follows the approach it has set out for firms.

Germany’s draft bill sets limits on the authority’s machine in statute

Germany approaches the same question from a different angle. The difference lies less in the technology than in where the rule is set. Section 65(6) of the Federal Ministry of Finance’s draft bill implementing the sixth EU Anti-Money Laundering Directive, in the version dated 28 October 2026, governs automated data analysis by the FIU. It would bar self-learning and automated systems that can independently assess how dangerous an individual is, as well as any decision based solely on automated processing that has legal effects on a person. For the first ban the explanatory memorandum cites a ruling by the Federal Constitutional Court on 16 February 2023, under which the essential features of such methods must be laid down by law. The second, it says, implements the “human in the loop” principle. Here the limit would be drawn by Parliament rather than by the authority itself.

The comparison is not exact. The FIU analyses suspicious activity reports about individuals, whereas the FCA supervises firms, and the German text is a draft that, according to Haufe, the cabinet intends to adopt before the end of the year. Meanwhile Germany keeps anti-money laundering supervision of the professions fragmented, the very thing the UK is consolidating. Under Section 50 of the current Anti-Money Laundering Act (GwG), which the draft is to replace when the EU Anti-Money Laundering Regulation applies from 10 July 2027, supervision of lawyers, tax advisers and auditors lies with their professional chambers, supervision of notaries with the presidents of the regional courts and supervision of the rest of the non-financial sector with the federal states. When the Financial Action Task Force (FATF) assessed Germany in November 2021, it counted around 337 supervisors at state and district level, who were responsible for the non-financial sector and certain insurers. The draft creates a central coordination office at BaFin which, under Section 81(3), does not supervise obliged entities. The UK intends to replace 22 professional bodies and HMRC with a single supervisor, while Germany is coordinating the ones it has.

At European level, the Anti-Money Laundering Authority (AMLA) must begin its first selection of directly supervised financial institutions by 1 July 2027 under Article 13(4) of Regulation (EU) 2024/1620. Direct supervision of at most 40 groups in the first round begins six months after the list is published, which means mid-2028 at the earliest. Its work programme for 2026 to 2028 lists machine learning, anomaly detection and governance to guard against model drift, but, in the version reviewed, does not mention agents. ECB Banking Supervision uses assistants such as Athena and Heimdall. Pedro Machado, a member of the Supervisory Board, said in October 2025: “Heimdall does not decide.” Among the EU authorities reviewed here, only the European Securities and Markets Authority (ESMA) describes an agent of its own, in its data report published on 29 May 2026, and that agent checks the methodologies of credit rating agencies. By that measure the FCA announces the more extensive use, and its published rules for it are thinner.

Recommendations

Make the RTS 22 Article 15(3) reconciliation a continuous process

A firm that reconciles its trading data against FCA extracts only on a sample basis may hear about a systematic error from the regulator first. In the Sigma Broking case, almost all of the firm’s transactions over five years were misreported, and seven months passed between the FCA’s first contact in May 2023 and the end of the period under review in December 2023. Reconciling continuously, with documented thresholds for what counts as a discrepancy, is the precondition for answering an automated query from the firm’s own figures.

Align breach notifications with the template set out in Market Watch 82

The FCA expects a breach notification to state the issue, its root cause, the exact number of affected reports, the period concerned, a date for back reporting and a description of the control weakness. A template that fixes these fields in advance and assigns them to a responsible function shortens the response noticeably. That matters once queries arrive faster, and possibly from a machine.

Name a responsible person and keep records of automated query handling

Nikhil Rathi considers it possible that the FCA will interact with firms’ agents. A firm that uses its own agents to answer queries should decide who approves their answers and record which data, model versions and escalations were involved. That matches the documentation the FCA describes as good practice in its collection. It is not yet required for regulatory queries, but it keeps a person involved on the firm’s side, so that machines are not the only parties to the exchange.

Use the switch to the 2028 reporting format to rebuild data lineage and reconciliation

UK subsidiaries and branches that report under UK MiFIR have to rebuild their reporting by 3 April 2028 anyway. If they also re-engineer, within the same project, how every reporting field is traced to its source and how reports are reconciled against FCA extracts, they build once instead of twice. The FCA has not linked the format change to the agent. The 2028 rebuild is simply a convenient moment.

Timeline: From the Mills Review to agentic supervision
Decisions, speeches and deadlines on AI in UK supervision, with the German approach alongside
21 October 2025
HM Treasury decides on a single supervisor
The FCA is to become the single anti-money laundering supervisor of the professional services sector, covering around 60,000 firms.
24 June 2026
“First responder”
Nikhil Rathi says the FCA is exploring agentic AI to monitor wholesale markets.
6 July 2026
Mills Review
Recommendation 6 calls for an agentic supervisory model, with people at advisory and preparatory levels.
9 July 2026
Annual report 2025/26
The Board lists governance and controls over the FCA’s own AI use as an area for strengthening.
August 2026
PS26/15
The FCA publishes the new transaction reporting format with 52 fields instead of 65.
13 October 2026
Treasury Committee
Jessica Rusu: a person in every decision today, with the ambition “to automate wherever we can”.
19 October 2026
“Becoming an agentic regulator”
Drafted speech names AI tools for 60,000 anti-money laundering firms and an agent for around 9,000 reporting firms.
28 October 2026
German draft bill
Section 65(6) would bar the FIU from self-learning systems that independently assess how dangerous a person is.
3 April 2028
New reporting format in force
Back reporting period of three years instead of five.
Mid-2028
AMLA begins direct supervision
Earliest start for at most 40 financial groups in the first round.
Autumn 2028 to around 2030
Transfer to the FCA
The FCA takes over anti-money laundering supervision of professional services in phases, provided the bill passes.

Glossary

Transaction reporting: the obligation under Article 26 UK MiFIR to report every transaction in a reportable financial instrument to the FCA, using prescribed fields. The data are used above all to detect market abuse.

Market Data Processor: the FCA system that validates incoming reports against fixed rules, rejects those that fail and provides firms with extracts for reconciliation.

Breach notification and back reporting: a firm that finds reporting errors must notify the FCA promptly under Article 15(2) RTS 22 and resubmit the corrected reports.

Professional body supervisors: the 22 professional bodies that currently supervise law firms’ and accountants’ compliance with anti-money laundering rules in the UK. Under the bill, this role is to pass to the FCA.

FIU: Financial Intelligence Unit, the German authority that receives and analyses suspicious activity reports on money laundering.

Christian Schablitzki

Christian Schablitzki – the agentic banker

Strategy & Management Consultant · Agentic AI expert for financial institutions

More than 20 years in investment banking and derivatives trading, followed by over 10 years advising financial institutions. Certified as Claude Partner – Claude Code (Anthropic), as well as in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.

LinkedIn profile → · agenticbanker.ai →
newsletter
the agentic banker

Keep reading – every 14 days in your inbox.

Capital-markets insights, regulatory updates and AI trends. Concise, well-founded, free.

GDPR-compliant. Unsubscribe at any time.

← Back to overview