A Chief Risk Officer who sets out to hire help with AI governance meets a market with no settled shape. Strategy houses, audit firms, law firms, specialist assurance companies, platform vendors and certification bodies all offer something called AI governance, and their proposals rarely describe the same deliverable. The Department for Science, Innovation and Technology estimated in Assuring a Responsible Future for AI, published on 6 November 2024, that 524 firms supply AI assurance goods and services in the UK, and only 84 of them are specialised AI assurance companies.
This guide is for the regulated buyer: a bank, insurer, wealth manager, pharmaceutical company, law firm or other business whose AI use will one day be read by a supervisor, the reader our compliance library is written for. It compares the kinds of firm that sell AI governance support, the deliverables a regulated firm should insist on, and the questions that separate a working governance system from a well-designed report. It answers one buyer question: who should I hire, for which part of the job?
The scope is deliberate. What a governance framework has to contain for a regulated sector is covered in our enterprise AI consulting guide, and the columns of an AI inventory are covered in our post on the AI tool inventory regulators ask for first. How to interrogate an ISO 42001 certificate and its scope statement belongs to our ISO 42001 procurement guide, and what happens when a consultancy changes owner belongs to our Faculty alternative analysis. This post stays on the selection decision.
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 engagement, which covers governance, RACI and lifecycle, audit-ready by design.
What does an AI governance adviser actually deliver?
An AI governance adviser should deliver one of three distinct things: a governance system the firm runs itself, an independent assurance opinion on that system, or an accredited certification of it. These are separate jobs. A regulated firm needs the first, may need the second, and may want the third, and in our view confusing them is how a governance engagement ends with a report nobody operates.
The first job is building. A builder writes the AI policy, stands up the register of AI systems, maps controls to the rules that bind the firm, names an owner for each system and sets a review cadence that keeps the register true as models change. The output is something the firm operates every month, not a document it files. That is the work our AI operating model design is built for.
The second job is assuring. An assurer tests what the builder made and reports on it: whether the controls work, whether a model behaves as claimed, whether the evidence would satisfy a supervisor. DSIT's Introduction to AI assurance describes this family of techniques, from risk and impact assessments to bias audits, compliance audits, conformity assessments and formal verification. The value of an assurance opinion comes from the assurer's independence, so the assurer should not be the firm that built the system it is assessing.
The third job is certifying. A certification body audits an organisation's AI management system against ISO/IEC 42001, the AI management system standard published in December 2023, and issues a certificate if it conforms. Since July 2025, ISO/IEC 42006 has set requirements for the bodies that audit and certify those systems, supplementing ISO/IEC 17021-1 with AI-specific competence rules. A certificate says the management system met the standard on the audit date; it is not a governance system in itself. In our view certification is the last step rather than the first, which is the sequence this guide follows; our approach sets out how we place AI inside enterprise systems.
| Role | What the firm receives | Who typically sells it | Independence test |
|---|---|---|---|
| Build | Policy, register, control map, named owners, review cadence, run by the firm | Consultancies, boutiques, law firms (policy only), platform vendors (tooling only) | Not required; the builder works for you |
| Assure | A tested opinion on whether controls and models work as claimed | Specialised assurance companies, audit firms, some consultancies | Required; must not assure its own build |
| Certify | An ISO/IEC 42001 certificate for a defined scope | Certification bodies auditing to ISO/IEC 42006 | Required; separate from whoever built the system |
A buyer who writes one tender asking for all three from a single firm has already made the decision that matters most, and in our view made it the wrong way. The rest of this guide is about choosing well for each job, with most of the weight on the building job, because that is the one a regulated firm cannot do without. If you have not yet decided which AI uses need governing first, a mid-market AI readiness assessment comes before this decision.
Why is choosing an AI governance adviser harder in 2026?
Choosing an AI governance adviser is harder in 2026 because the supply side is broad, young and loosely defined. DSIT counted 524 UK firms supplying AI assurance, but only 84 of them are specialised assurance companies, and buyers and suppliers still use different words for the same work. A regulated buyer has to do the sorting the market has not yet done.
The make-up of that 524 matters more than the headline. DSIT's report breaks it down as 84 specialised AI assurance companies, an estimated 225 AI developers, 182 diversified firms and 33 in-house adopters, with 60% of suppliers classed as SMEs. Together those firms generated an estimated £1.01 billion of Gross Value Added and employed an estimated 12,572 people, according to the same DSIT report. GVA is a measure of economic contribution, not of what buyers spend, so it says little about the size of any one engagement, and nothing about who will do the work, which is the question our guide to the UK AI consulting market keeps returning to.
Buyer confidence is low too. In DSIT's industry survey, only 44% of respondents agreed that they felt comfortable demonstrating that the AI systems their organisation develops or deploys comply with existing UK regulations, and 43% of firms felt they had limited understanding of both the UK's regulatory approach and those of other countries. A buyer who cannot yet say what compliance looks like will find it hard to judge a proposal that claims to deliver it.
Familiarity with the work itself is uneven. The same DSIT survey recorded high familiarity with the established mechanisms and much lower familiarity with the newer ones that matter most for generative and agentic systems.
The pattern in that chart is the procurement risk. Risk assessments and compliance audits are the mechanisms industry knows best, so a proposal built only on them will look familiar and complete. Red teaming and model cards, which tell you how a model behaves under pressure and what it was built to do, were familiar to barely a third of respondents. A regulated firm deploying large language models should check whether a proposal covers the mechanisms its systems need, rather than the mechanisms the supplier finds easiest to deliver. Our AI solutions include evaluation and observability for production agents, which is where that testing work lives once a system ships.
Terminology adds friction. DSIT's survey respondents cited the lack of international standardisation of terminology (31%) and of UK standardisation (25%) among the barriers to understanding AI assurance. Two proposals that both promise an "AI governance framework" can describe a two-page principles statement and a fully operated control system. The buyer's first task is to make suppliers describe their deliverable in the buyer's terms.
Which kinds of firm sell AI governance support in the UK?
Six kinds of firm sell AI governance support in the UK: large consultancies and the Big Four, law firms, specialised AI assurance companies, AI developers and platform vendors, certification bodies, and senior-led boutiques. Each is strong at one part of the job. A regulated firm gets the best result when it matches the firm type to the job type rather than buying the whole programme from whichever name it already knows.
Large consultancies and the Big Four. Accenture, Deloitte, PwC, EY and KPMG operate across many countries and sectors, which suits a programme that has to run in several jurisdictions at once. They can build, and their audit and risk arms can assure, which is a strength for scale and a question for independence when the same group does both. The buyer should ask who in the proposed team will write the policy and stand up the register, and how senior that person is. Our pillar on choosing an AI consultancy in the UK covers when a large firm is the right call.
Law firms. A law firm is the right source of advice on what the rules require and who they bind, from UK GDPR to the EU AI Act. Legal advice is a different deliverable from an operated register or a control system, so check what a law firm will actually build before treating it as the lead. Use it to settle legal questions the governance system then implements.
Specialised AI assurance companies. These are DSIT's 84: firms whose core offer is assessing AI systems. They are the natural assurer, especially for testing, bias audits and red teaming, provided they did not also build the system. Assessment is the work DSIT's category describes, so ask whether they also design operating models and own the change work that keeps governance alive inside a business before giving them the build.
AI developers and platform vendors. AI developers are the largest group in DSIT's count, at an estimated 225. Where a developer or platform vendor offers governance tooling, such as a register, a dashboard or an evaluation harness, that tooling can make a governance system cheaper to run, but it is not the system. A vendor assessing its own model also fails the independence test discussed below, and DSIT's public focus groups were sceptical of assurance delivered only by a profit-motivated AI developer. Our post on AI feature governance for SaaS vendors covers what a buyer can fairly ask a vendor to evidence.
Certification bodies. A certification body audits an organisation's AI management system against ISO/IEC 42001 and issues a certificate for a defined scope. It is the right choice when a certificate is a procurement or client requirement. It should not also be the firm that designed the system it certifies. Read our ISO 42001 certificate checks before relying on one.
Senior-led boutiques. A focused firm puts senior people on the work itself. Our own DATS methodology is in this group: it is delivered by senior practitioners who ship code, not decks, and its operating model engagement covers governance, RACI and lifecycle, audit-ready by design. A boutique is weaker where the work needs hundreds of consultants in several countries at once, and it should say so.
| Firm type | Strongest at | Weakest at | Fit for a regulated mid-market buyer |
|---|---|---|---|
| Large consultancy or Big Four | Scale, sector depth, multi-country programmes | Senior hands on the build; build-and-assure independence | Strong for global rollouts; check who does the work |
| Law firm | What the rules require and who they bind | Operated registers and control systems | Essential input, rarely the builder |
| Specialised assurance company | Independent testing, bias audits, red teaming | Operating model design and change | The natural assurer |
| AI developer or platform vendor | Tooling that lowers running cost | Independence on its own models | Useful tool supplier, not the governance owner |
| Certification body | ISO/IEC 42001 certification for a scope | Designing the system it certifies | Use when a certificate is required |
| Senior-led boutique | A working system built and handed over | Very large parallel programmes | Strong for building and running the first system |
The table is a starting point, not a ranking. A regulated firm that needs both a build and an independent opinion will use at least two of these types, and the decision is about which one leads.
What should a regulated firm demand as deliverables?
A regulated firm should demand deliverables it can operate and evidence: an AI policy in force, a register of every AI system with a named owner, a control map tied to the rules that actually bind the firm, a review cadence with dates, and proof that third-party models are understood. A slide deck that describes these things is not a substitute for any of them, however polished it is.
Start with the register, because a supervisor will. Each AI system needs an entry that says what it does, which data it touches, who owns it, what risk tier it sits in and when it was last reviewed. The columns that satisfy the ICO, the FCA and the EU AI Act together are set out in our post on building the AI inventory; the adviser's job is to make that register exist and stay current, not to describe it.
Third-party models need particular attention in financial services. The Bank of England and the FCA found, in Artificial intelligence in UK financial services 2024, published on 21 November 2024, that 46% of respondent firms reported only partial understanding of the AI technologies they use, against 34% reporting complete understanding, and the regulators attributed that gap largely to third-party models. The base for those two figures is firms using or planning to use AI. The same survey found that a third of all AI use cases were third-party implementations, up from 17% in 2022. An adviser who cannot show how the firm will understand and evidence a model it did not build is leaving the hardest part of the job undone.
The control map must name the rules that bind the firm, not a generic list. For a UK bank with internal model approval, that includes the PRA's SS1/23 model risk management principles, which took effect on 17 May 2024 and are relevant to UK-incorporated banks, building societies and PRA-designated investment firms with internal model approval to calculate regulatory capital. SS1/23 covers identifying and managing the risks of artificial intelligence in modelling techniques to the extent it applies to models generally. For a firm outside that scope, the map will look different, and an adviser who presents the same control list to a bank and to a law firm has not done the mapping. A layered view of how those controls stack is in our enterprise AI governance framework.
Rules that may apply later need dates, not alarm. The EU's Digital Omnibus on AI moves high-risk obligations for standalone Annex III systems to 2 December 2027 and those for AI embedded in Annex I regulated products to 2 August 2028, according to Gibson Dunn's summary of the agreed text, which notes that the changes take legal effect on formal adoption. A good adviser builds a register that can absorb those obligations when they arrive, rather than selling an urgent programme against a deadline that has moved. The Omnibus also softened the literacy duty, which we cover in our post on AI literacy under Article 4.
The same delivery discipline underpins our AI execution office, embedded delivery with production placements the client owns.
How do you keep the builder and the assurer independent?
A regulated firm keeps the builder and the assurer independent by giving the two jobs to different firms, or to clearly separated teams with separate reporting lines, and by writing that separation into both contracts. An assurer that tests a system its own group designed is grading its own work, and DSIT's roadmap is clear that who does the assuring matters to the public as much as how.
DSIT's Trusted third-party AI assurance roadmap, published on 3 September 2025, puts the point plainly. Third-party providers have a particular role in independently verifying AI systems, the roadmap says, especially given that "for the public the organisation assuring the AI product or service is as important as the process". The 2024 market report records the same concern from its public focus groups, where participants felt that assurance delivered only by a profit-motivated AI developer would not be trustworthy.
Independence is easiest to arrange at the start. Three practical rules help:
- Separate the contracts. The builder's statement of work excludes assuring its own deliverables, and the assurer's excludes design work on the systems it tests.
- Give the assurer access. DSIT's roadmap names information access as one of four challenges for third-party assurance, alongside quality, skills and innovation. The builder's contract should require it to give a later assurer the documentation, test results and model information that assurer needs.
- Keep the certifier apart from both. If an ISO/IEC 42001 certificate is in scope, the certification body is a third party to both the builder and the assurer, and its competence is governed by ISO/IEC 42006.
Professional standards for assurers are still forming. As at 3 September 2025, DSIT's roadmap set out a plan to convene a consortium to develop a voluntary professional code of ethics for AI assurance and a skills and competencies framework, and to work towards a future professional certification or registration scheme. Until such a scheme exists, the buyer's own diligence on an assurer's method, staff and independence carries the weight, which is why the questions later in this guide matter.
For a firm that keeps sensitive data inside its own perimeter, independence has a data dimension too. An assurer may need access to models and logs that the firm will not send to a shared cloud tool; our post on self-hosted AI for regulated institutions covers how that perimeter can be drawn.
Who stays accountable after you hire an adviser?
Accountability stays with the regulated firm after it hires an adviser. An adviser can build, test or certify a governance system, but it does not take on the firm's regulatory responsibilities by doing so. For UK banks within its scope, the PRA's SS1/23 expects responsibility for the overall model risk framework to be allocated to the most appropriate Senior Management Function holder.
Many financial firms already have a structure in place. The Bank of England and FCA survey found that 84% of firms currently using AI reported having an accountable person for their AI framework, and 82% used an AI framework, principles, guidelines or best practice. The same survey reported that 72% of firms using or planning to use AI said their executive leadership were accountable for AI use cases, while most firms reported three or more accountable persons or bodies. Split accountability is where an adviser can help or hurt: a good one makes the split explicit and documented, a weak one adds another committee.
The test for any proposal is simple. Ask the adviser to name, for each AI system in your register, the person inside your firm who will own it after the engagement ends. If the answer is the adviser's own team, the engagement is building dependency rather than governance. The decision-rights side of this, and how ownership moves when AI delivery leaves a central team, is covered in our post on the AI operating model versus a centre of excellence.
Accountability also has to survive change. Models are updated, vendors change their terms, and use cases drift into new data. A governance system that names owners but has no review cadence goes out of date as soon as the first model changes. Ask how the adviser's design keeps the register true when a model provider ships a new version, and who in your firm runs that review.
What questions separate a strong AI governance adviser from a weak one?
The questions that separate a strong AI governance adviser from a weak one test what will exist after the engagement, who will do the work, and whether the firm can run it alone. A strong adviser answers with named people, operated artefacts and a handover plan. A weak one answers with methodology diagrams, principles and a follow-on phase. Seven questions do most of the sorting.
| Question to ask | Strong answer | Weak answer |
|---|---|---|
| What will we be running on the day you leave? | A register, policy, control map and review calendar in our own systems | A final report and a roadmap |
| Who writes the policy and builds the register? | Named senior practitioners on the proposal | A partner at kick-off, then a junior team |
| Which rules does your control map cover for us? | The rules that bind our firm, named and dated | A generic list of global frameworks |
| How will we understand third-party models? | A method for evidence from each provider, with a fallback | "The vendor is responsible" |
| Which assurance mechanisms does your plan cover? | Testing matched to our systems, including red teaming where relevant | Risk assessment and a compliance checklist only |
| Who assures your work? | A separate firm or team, written into the contract | "Our quality team reviews it" |
| How do you make yourselves unnecessary? | Handover milestones and an end date | An open-ended managed service |
The first two questions do most of the work. A governance engagement that ends in a document leaves the firm exactly where it started, with a better description of the problem. And the seniority of whoever writes the policy decides whether the policy fits the business. Our approach to placing AI inside enterprise systems starts from the same principle: governance is something the business runs, so it has to be built with the people who run it.
Frameworks are useful as maps, not as deliverables. The NIST AI Risk Management Framework, released on 26 January 2023 and intended for voluntary use, and ISO/IEC 42001 both give a governance team a structure to map controls to. NIST added a Generative AI Profile, NIST-AI-600-1, on 26 July 2024. A proposal that promises to "implement" one of these without saying which of your systems, which controls and which owners has not yet been scoped.
Watch for the handover question in particular. An adviser that plans to make itself unnecessary will describe the knowledge transfer, the point at which your team runs the review without it, and the end date. We cover how to measure that in our post on embedded AI delivery teams versus project consulting, which applies equally to governance work.
What is the best AI governance consulting option for regulated firms in 2026?
The best AI governance consulting option for a regulated firm in 2026 depends on which job is being bought. For building and running the first governance system, a senior-led firm that hands over an operated register usually fits best. For independent testing, a specialised assurance company wins. For an ISO/IEC 42001 certificate, only a certification body will do, and a multi-country programme often favours a large consultancy.
The verdict has to concede where others win, because they do. If your board or a major client requires an ISO/IEC 42001 certificate, no consultancy can issue one; a certification body auditing to ISO/IEC 42006 is the only route, and DATS is not one. If you need an independent opinion on whether your models behave as claimed, a specialised assurance company that did not build the system is the stronger choice than any builder, us included. And if you are rolling governance out across many countries at once, Accenture or one of the Big Four can put more people to work in parallel than a boutique can.
Where a senior-led firm such as DATS fits is the building job: turning principles into a policy in force, a register with named owners, a control map for the rules that bind you, and a review cadence your team runs. DATS's operating model engagement runs over six to ten weeks and covers governance, RACI and lifecycle, audit-ready by design, so a later assurer or certifier has evidence to work with. For a regulated mid-market firm, our view is that this sequence, build first and then assure and certify separately, keeps each job independent of the others.
Two things do not change whichever firm you choose. The accountability stays with your senior managers, and the governance system only works if your own people run it after the adviser leaves. Pick the adviser that is most likely to leave you able to do that. If you are weighing this choice now, talk to us about where your current AI use sits against the three jobs.
Frequently asked questions
Does an AI governance adviser need to be ISO 42001 certified?
An AI governance adviser does not need its own ISO/IEC 42001 certificate to help a firm build a governance system, because the standard certifies an organisation's AI management system, not its advisers. What matters is whether the adviser can design a system that a certification body could audit. If a certificate is the goal, the auditing body should meet ISO/IEC 42006 and be separate from the adviser.
Ask the adviser to show how its design maps to the clauses of ISO/IEC 42001 that matter for your scope, and leave the certificate itself to the certification body. Our ISO 42001 procurement guide explains what a certificate does and does not prove. DATS itself makes no certification claim.
Can one firm build, assure and certify our AI governance?
One firm should not build, assure and certify the same AI governance system, because each later job depends on independence from the earlier one. A builder works for the client, an assurer must test work it did not design, and a certifier must audit a system it did not create. A group that offers more than one of these jobs should be asked to show how it separates them, in writing.
DSIT's roadmap on third-party assurance says the organisation doing the assuring matters to the public as much as the process, and its 2024 report recorded scepticism in public focus groups about developers assuring their own AI. A practical sequence is to appoint a builder first and choose the assurer once there is a working system to test. Our FAQ covers common questions about how we work.
Want to see how this works in practice? Read about our DATS methodology, see how the AI execution office ships placements into production, browse the AI solutions we build, or learn about Dilr.ai and the practitioners behind it. For governance in a specific sector, see our guide to university AI governance, browse the strategy library, or simply talk to us.
Governance your team can run.
30-min scoping call · No deck · Confidential. We will tell you which of the three governance jobs you need first, and who should do each one.
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.