What is AI governance? A practical guide for UK firms

AI governance is how an organisation decides which AI it will allow, sets rules for how that AI may be used, and keeps a record it can show to anyone who asks. It covers three things: knowing where AI is being used, controlling what it can do, and being able to prove both after the fact.

That definition is deliberately plain, because the term has been stretched a long way. Ask five vendors what AI governance means and you will get five answers, most of them describing whichever part of the problem that vendor happens to solve. This page sets out the whole picture first, then says exactly which part of it we work on, so you can tell what you already have covered and what you do not.

The three questions AI governance has to answer

Strip away the frameworks and the vocabulary and AI governance comes down to three questions a firm should be able to answer about itself.

Which AI is being used here? Not which AI you approved. Which AI is actually in use, by whom, on what, and at what cost. For most firms the approved list and the real list are different documents, and nobody has written the second one down.

What is it allowed to do? Which tools and models are permitted, what data may go into them, what happens out of hours, and what the organisation is prepared to spend. A rule that exists only in a policy document is a preference. A rule enforced by a system is a control.

Can you show what happened? When a client, an auditor, an insurer, a board or a regulator asks you to demonstrate that your AI use was controlled, what do you hand them? This is the question that separates firms that have governance from firms that have intentions.

What AI governance covers, in full

It is worth being clear about the scope, because two quite different disciplines share the name.

Governing how AI is built. Model development, training data provenance, bias and fairness testing, model validation and documentation, performance monitoring and drift. This matters enormously if your firm builds or fine-tunes its own models. It is a specialist discipline with its own tooling and its own experts.

Governing how AI is used. Which tools staff and systems are permitted to use, what data reaches them, who is accountable, what it costs, and what evidence exists. This is the discipline that applies to almost every firm, because almost every firm uses AI somebody else built.

Evaident works on the second. We are an accountability and evidence layer for how people and agents *use* AI. We do not do model validation or bias testing, and a page like this is exactly where a vendor is tempted to imply otherwise. If you build models, you need a model governance capability as well as this, and they are complementary rather than competing.

Why the second kind got hard

Three years ago governing AI use meant deciding whether staff could use ChatGPT. The problem has changed shape since, in three ways.

AI arrives through three doors, not one. There are the tools you approved and pay for. There are the in-house agents, scripts and API integrations your own developers have built, which call models directly and are invisible to any per-seat licence report. And there is shadow AI: the free tools, personal accounts and browser extensions nobody asked about. Each door produces a different kind of record, or none at all. We wrote a full explainer on finding shadow AI if that is the door worrying you most.

Agents removed the person from the loop. When a colleague uses a chat tool, you can reconstruct what happened by asking them. When an in-house agent runs a task end to end, calling a model dozens of times without anyone watching, there is nobody to ask. The record has to come from the system, and it has to exist before you need it.

Vendor logs do not add up to a record. Four AI vendors means four audit APIs, four export formats and four retention windows, none of which know about each other, and none of which cover the agents you built or the apps nobody approved. You can assemble something from them. You cannot assemble a single, continuous, verifiable timeline.

What UK and EU regulation actually expects

No UK regulator requires you to buy an AI governance platform. What they increasingly expect is that you can account for your AI use, and accounting for it means records rather than recollection.

The EU AI Act. Regulation (EU) 2024/1689 entered into force on 1 August 2024 and becomes fully applicable on 2 August 2026, with some exceptions. Its transparency rules come into effect in August 2026. Prohibited practices and AI literacy obligations have applied since 2 February 2025, and the obligations for general-purpose AI models since 2 August 2025.

The exceptions matter, because they are where the heaviest obligations sit. The simplification package known as the AI omnibus was given final approval by the Council of the EU on 29 June 2026 and sets a fixed timeline: rules for systems used in certain high-risk areas, including biometrics, critical infrastructure, education, employment, migration, asylum and border control, apply from 2 December 2027, and rules for AI built into products such as lifts or toys apply from 2 August 2028. Where a system is high risk, the obligations include logging activity so results are traceable, keeping detailed documentation, and putting appropriate human oversight in place. Those obligations bite on the 2027 and 2028 dates, not this August.

Note too the scope question that catches people out: a UK-only deployment may fall outside the Act's direct scope, and it is worth establishing that before assuming the whole framework applies to you.

UK GDPR. Personal data does not stop being personal data because it was typed into a chat window. If client information reaches a public model, that is a processing question with all the usual consequences.

FCA Consumer Duty and SYSC 9. Regulated firms are already expected to keep adequate records and to demonstrate that oversight is real rather than asserted. Under the Senior Managers and Certification Regime, a senior manager is expected to evidence reasonable steps, not merely describe them. We have written separately about what the FCA's Mills Review means for evidencing AI use, and there is a dedicated page for FCA-regulated firms.

The SRA Code. For law firms the pressure point is confidentiality and privilege: matter codes, client names and draft advice reaching a model outside the firm's control. There is a page for law firms that goes into this properly.

Evaident supports your compliance function; it does not provide legal advice. Nothing on this page is a substitute for advice on your own obligations.

What an AI governance platform has to do

If you are assessing tools, these are the capabilities worth testing, whoever you buy from.

One record across every source. Approved tools, in-house agents and shadow AI in a single normalised timeline. If a tool covers only the vendors it integrates with, it will always be missing the part you most need to know about.

Real-time control where control is possible. Be sceptical of anything claiming to block all AI use, including ours. We do not, and we have not seen a tool that does. Traffic you route through a governed gateway can be checked and stopped before it leaves the firm. Traffic you merely observe, through vendor audit logs or network log analysis, can be detected and evidenced but not blocked. A tool that blurs those two is selling you a comfort you do not have.

Evidence that survives scrutiny. A dashboard tells you what a system currently believes. Evidence is a record you can hand over, that shows it has not been altered since. If a record can be quietly edited, it is management information, not evidence.

Cost visibility per person and per department. AI spend has a habit of arriving as one large invoice with no way to attribute it. Per-person cost, premium-model share and a spend cap turn that into something a finance director can govern.

Regime mapping. Usage mapped to the obligations you actually answer to, so the evidence pack arrives in a shape a compliance reviewer recognises.

A deployment that does not become a project. Read-only connectors, a base-URL change for your own tools, and log analysis of data you already collect. If it needs software on every endpoint before it tells you anything, the rollout will outlast the problem.

Where Evaident fits

Evaident is an AI accountability platform. It does three things, in this order.

Know. Every AI interaction across approved chat tools, in-house agents and unapproved apps, in one searchable record: per person, per department, per tool, including who is not covered yet. Shadow AI is discovered from the firewall, proxy, secure web gateway or SIEM logs your firm already collects, so there is nothing new to install on endpoints.

Govern. In-house agents and API tools point at the Evaident gateway instead of the vendor directly, which is a one-line change to a base URL. Policy is then enforced in real time before a request leaves the firm: approved vendors and models, supported UK PII, common credentials, customer-defined blocked terms such as matter codes and project codenames, out-of-hours rules and an organisation spend cap. To be exact about the limit, this real-time blocking applies to gateway-routed traffic. Vendor audit logs, the optional browser extension and shadow-AI discovery detect and evidence; they do not block.

Prove. Every record is sealed to the one before it with a cryptographic signature (HMAC-SHA256), so any tampering is mathematically detectable. One click produces a stamped evidence pack with an integrity certificate. This is the part that answers "show me", and it is the reason a policy document alone leaves a gap, which we set out in why an AI policy is not evidence.

There is also an AI system register with deterministic EU AI Act posture indicators, available on the Control plan and above. You inventory your AI systems and it maps your answers to cited provisions of the Act, including an EU-scope check, and it will return "potential high-risk" pending an Article 6(3) assessment rather than automatically labelling an Annex III system as high risk. It is a governance and triage aid. It is not legal advice, not a conformity assessment and not a final determination of any system's legal status.

Some honest boundaries, because they matter more than the feature list. Evaident is not a data-loss prevention replacement and does not cover every route data can leave a firm. It does not detect legal privilege in content. It records metadata by default, meaning who used which AI, when, which model, token counts and risk flags, rather than the content of prompts; full capture is optional, scoped, approval-led and per source, with PII redaction on by default. And it supports your compliance function rather than conferring compliance on you. No product can do the latter.

Where to start

Governance projects stall when they begin with a policy rewrite. Start with the facts instead.

1. Find out what is actually in use. Until you know which AI tools are being used and by whom, every rule you write is a guess. Our free AI exposure check gives you a picture without connecting a live source.
2. Bring your own agents under control first. In-house agents and API tools are the highest-risk and the easiest to govern, because routing them through a gateway is a base-URL change rather than a change management programme.
3. Then write the policy. A policy written against what you know is happening, and enforced by something, is worth more than a thorough one written against what you assumed.
4. Make sure the evidence accumulates from day one. Evidence is not something you can produce retrospectively. It either exists from the point you started, or it does not exist.

Pricing is public and starts at £99 a month excluding VAT for firms of up to around 150 staff, with the gateway, blocked terms, compliance mapping and the board report arriving at Control, £249 a month excluding VAT. The full breakdown is on the pricing page.

If you take one thing from this page, make it this. The firms that struggle when someone asks them to account for their AI use are rarely the ones without a policy. They are the ones whose policy was never connected to anything that keeps a record.

---

_Evaident supports your governance and compliance functions. It does not provide legal advice. Regulatory dates on this page were checked on 21 July 2026 against the European Commission's AI Act page (digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai, last updated 11 May 2026) and the Council of the EU press release of 29 June 2026. Some other Commission pages had not been updated to the new high-risk dates at that point, so check the current position before relying on any date for your own planning._