AI compliance software: what it does, what it cannot, and how to choose

AI compliance software helps a firm see how AI is actually being used, enforce rules where enforcement is genuinely possible, and keep records it can show to a client, an auditor or a regulator. It supports your compliance function. No software makes you compliant, and any product that says otherwise has told you something useful about the vendor.

If you run a UK regulated firm, you have probably arrived at this page from one of three directions. A client questionnaire asked how you control AI use. A compliance consultant or board member asked what evidence you could produce if challenged. Or you simply noticed that much of the firm is using AI tools and nobody can say which ones, on what data, at what cost. All three lead to the same question: what does software in this category actually do, and how do you tell a genuine capability from a comforting claim?

This guide covers both, including the limits of our own product, because a buyer's guide that exempts its author is a brochure.

What AI compliance software actually does

Strip away the vendor language and the useful capabilities fall into five groups.

It shows you what is actually in use. Not the approved list. The real list: which AI tools people and systems are using, who is using them, on what kind of work, and what it costs. In most firms these are two different documents, and only one of them has been written down. Discovery matters most for the AI nobody asked permission for, which we cover in our guide to shadow AI. Good tools build this picture from sources you already have, such as vendor audit logs and the network logs your firewall or proxy already collects, rather than demanding new software on every laptop.

It enforces rules where rules can be enforced. Some AI traffic can be checked before it leaves the firm. If your in-house agents, scripts and API integrations are routed through a controlled gateway, requests can be tested against policy in real time: approved vendors and models only, supported UK PII stopped, customer-defined blocked terms such as matter codes or project codenames stopped, out-of-hours rules applied, a spend cap enforced. That is a control, in the sense a compliance officer means the word.

It keeps evidence, not just dashboards. A dashboard tells you what a system currently believes. Evidence is a record you can hand to someone else, in a form that shows it has not been altered since it was made. The difference sounds technical until the day someone external asks you to demonstrate what happened, which is the argument we make in full in why an AI policy is not evidence.

It maps activity to the regimes you answer to. For UK firms that usually means some combination of Consumer Duty, SYSC record-keeping expectations, the Senior Managers and Certification Regime, UK GDPR, the SRA Code for law firms, and the EU AI Act where it applies. Mapping means the evidence arrives shaped around obligations a reviewer recognises, rather than as a raw export someone has to interpret. Mapping is support, not conferral: the software points at the obligation and shows the relevant record, and a human still judges whether the firm meets it.

It makes cost visible. AI spend tends to arrive as one large invoice with no way to attribute it. Per-person and per-department cost, the share going to premium models, and a hard spend cap turn an unmanaged line item into something a finance director can actually govern. This is not strictly a compliance capability, but in practice it is often what gets the FD's attention and funds the project.

What AI compliance software cannot do

This section is the one most vendors skip. It is also the one that protects you at renewal.

It cannot make you compliant. Compliance is a property of how your firm behaves, judged by a regulator or a court, not a feature that ships in software. A product can map your AI use to obligations and produce the evidence; it cannot confer the outcome. Treat any claim of guaranteed compliance as disqualifying.

It cannot block all AI use. Be sceptical of this claim from anyone, including us. 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, browser-level capture or network log analysis, can be detected and evidenced but not blocked. Evaident's real-time blocking applies only to gateway-routed traffic; everything else we detect and evidence, and we say so plainly. A determined person with a personal phone is outside any tool's reach, which is why detection and evidence matter as much as blocking.

It is not data loss prevention. DLP tries to cover every route data can leave an organisation: email, file shares, removable media, cloud sync and more. AI compliance software governs the AI channel specifically. Evaident is an accountability and evidence layer, not a DLP replacement, and if a vendor in this category implies it covers every exfiltration path, that is a claim to test hard.

It cannot give legal advice. Software can organise the facts and reference the obligations. What your firm should do about them is a judgement for your compliance function and, where it matters, your lawyers. This page is general information, not legal advice, and the same applies to any product in the category.

It cannot recognise everything sensitive. Pattern-matching catches defined categories, such as supported UK PII and customer-defined blocked terms. It does not understand context the way a person does. Evaident, for example, does not detect legal privilege in content. Ask every vendor where their detection stops.

How to evaluate AI compliance software: seven questions

Put these to every vendor on your shortlist, in writing, and keep the answers.

1. Show me the coverage map. Which of the three routes AI enters a firm does the tool see: the approved tools you pay for, the in-house agents and API integrations your own people built, and the unapproved apps nobody declared? A tool that covers only the vendors it integrates with is always missing the part you most need to know about. The full picture of what belongs in that map is set out in our guide to AI governance.
2. Which traffic can you actually block, and which do you only observe? You want the boundary drawn precisely, in writing. Vague answers here predict overclaims everywhere else.
3. What would I hand a regulator or client, and how do they know it is genuine? Ask whether records are tamper-evident, meaning each record is cryptographically sealed so alteration is detectable, and whether an export carries an integrity certificate. If records can be quietly edited, you have management information, not evidence.
4. What do you store by default? Metadata (who used which AI, when, which model, usage and risk flags) is a very different privacy posture from storing full prompt content. Prefer metadata by default, with content capture as a deliberate, scoped, approval-led choice if you ever need it. Your own staff are data subjects too.
5. What does deployment actually involve? Read-only connections to vendor audit logs, a base URL change for your own tools, and analysis of logs you already collect is an afternoon. Software on every endpoint and network changes is a project. For a firm without a large IT function, this difference usually decides the outcome.
6. What does it cost, all in? Insist on public per-firm pricing including VAT treatment, and ask what sits in higher tiers. For reference, Evaident's pricing is public: from £99 a month excluding VAT for firms of up to around 150 staff, with the real-time gateway, customer-defined blocked terms and compliance mapping from the Control tier at £249 a month excluding VAT. Details on the pricing page.
7. Where does your product stop? The most revealing question of the seven. A vendor who answers it quickly and specifically has thought honestly about the category. A vendor who cannot is selling coverage they do not have.

Where Evaident sits, stated plainly

Evaident is an AI accountability platform built for regulated and data-sensitive UK firms, including FCA-regulated firms and law firms. It keeps one tamper-evident record of AI use across approved tools, in-house agents and shadow AI, enforces policy in real time on gateway-routed traffic, and produces stamped evidence packs on demand.

And the boundaries, in the same breath: blocking is gateway-only and everything else is detect and evidence; it is not a DLP replacement; it records metadata by default rather than prompt content; it does not detect legal privilege; and it supports your compliance function rather than making you compliant. If those limits rule us out for your problem, better to know now.

Where to start

Do not start with a shortlist. Start with facts. Until you know which AI is actually in use across the firm, you are evaluating tools against a guess, and the tool you need may be different from the tool you assumed.

Our free AI exposure check is a low-commitment way to get that first picture of your firm's AI exposure. Whatever you end up buying, from us or anyone else, it is the right first step: it turns a vague sense that the firm should do something about AI into a specific list of what needs seeing, controlling and evidencing.

---

_Evaident supports your governance and compliance functions. It does not provide legal advice. Regulatory regimes are referred to in general terms; take advice on your own obligations before relying on anything here._