Skip to content

AI Security

AI security and governance

In short

AI security and governance covers the policies, risk assessments and technical controls an organisation needs to deploy artificial intelligence safely and prove it did so. It spans AI system inventory, accountability for deployment decisions, model and supplier provenance, adversarial testing, and alignment to ISO 42001, the NIST AI Risk Management Framework and the EU AI Act.

The problem

The AI in your organisation is already in production. The governance is not.

The sequence is almost always the same. Engineering ships something useful with a model. Product adds a second. A vendor turns on an AI feature in a tool you already pay for. Somebody in operations starts pasting customer data into a consumer chatbot because it saves them an hour a day. None of it went through a review, because there was no review to go through.

Then a customer's security questionnaire asks which AI systems process their data and who approved them, or legal asks who signed off on an automated decision that has just been complained about. At that point the honest answer is that nobody knows, and the work of finding out takes weeks that the deadline does not allow.

Banning AI is not the answer and does not work anyway - it moves usage onto personal accounts where you cannot see it at all. The answer is a governance layer light enough that engineering does not route around it, and rigorous enough that it produces the evidence when someone asks.

Who this is for

  • Organisations where AI reached production before an AI policy existed, which is most of them.
  • Teams facing AI-specific questions inside enterprise security questionnaires for the first time.
  • Companies with EU exposure preparing for AI Act obligations, including the Annex III duties now applying from 2 December 2027.
  • Regulated firms whose supervisor has begun asking how AI-assisted decisions are governed.
  • Anyone whose board has asked who is accountable for an AI system already making decisions.

Delivery

How the engagement actually runs

Sprints inside your engineering cycle, with findings arriving continuously rather than a report arriving at the end. Durations are elapsed time for a typical scope, and get re-scoped against your estate before anything is committed.

  1. 011-2 weeks

    Discovery

    Find the AI actually in use, not the AI on the register. Interviews, tooling review, expense and SSO analysis. Shadow AI surfaces here, and it is always more than expected.

  2. 022-3 weeks

    Assess and prioritise

    Tier systems by potential harm, assess against ISO 42001 and the EU AI Act, and produce a prioritised backlog. Remediation on the highest-risk items starts before the assessment finishes.

  3. 034-8 weeks

    Build the governance layer

    Policy, accountability model, assessment methodology and inventory, built with your engineers in your tooling so it survives contact with delivery pressure.

  4. 04Ongoing

    Operate and evidence

    Assessments run on real deployments, decisions logged, the inventory maintained, and the evidence accumulating for whoever asks next - a customer, an auditor or a regulator.

Deliverables

What you actually receive

Artefacts your team owns and can maintain after the engagement, not a report that ages the moment it lands.

Deliverables for AI security and governance, with what each one contains and when it lands
DeliverableWhat it containsWhen
AI system inventoryEvery AI system in use, including shadow AI and vendor features, with owner, purpose, data touched and decision impact. Structured so your team maintains it rather than us.Sprint 1
AI policy and acceptable use standardWhat staff may and may not do with AI, written to be followed rather than to sit in a policy library. Covers consumer tools, customer data and code generation.Sprint 2
Accountability modelNamed owners for AI decisions, an approval path with the authority to stop a deployment, and the escalation route when something goes wrong.Sprint 2
AI risk assessment methodologyA repeatable assessment your teams run before deployment, sized to the system's potential harm rather than applied uniformly.Sprint 3
Supplier and model provenance registerProvenance, licensing, training-data position and change notification for every third-party model and AI vendor in scope.Sprint 3-4
Gap analysis to ISO 42001 and the EU AI ActWhere you stand against each, as a prioritised backlog with owners and dates rather than a maturity score.Sprint 4
Board-ready positionA short written statement of AI risk posture, what is covered, what is not, and what it would take to close the difference.Sprint 4

Frameworks this touches

In this practice

The individual services

FAQ

Questions people actually ask

How do we find shadow AI in our organisation?

Shadow AI is found through four routes used together: SSO and identity logs showing which AI services staff have authenticated to, expense and card data showing individual subscriptions, network or CASB data showing traffic to AI endpoints, and structured conversations with team leads about how work actually gets done. No single route is sufficient - card data misses free tools, and network data misses anything used on a personal device.

Does the EU AI Act apply to a UK company?

It can. The EU AI Act applies to providers placing AI systems on the EU market and to deployers established in the EU, and it also reaches providers and deployers outside the EU where the output of their AI system is used within the EU. A UK company with EU customers or EU users is frequently in scope. Being outside the EU is not by itself a reason to conclude the Act does not apply.

We use a third-party model rather than training our own. Are we still responsible?

Yes. Consuming a model rather than training one changes which obligations fall on you, not whether any do. A deployer is responsible for how the system is used, for human oversight, for the decisions it informs and for the data put into it. Provenance, licensing and change notification from the model provider become supply chain questions you own, and they have the longest lead time in any AI governance programme.

How long does it take to get AI governance in place?

A working governance layer - inventory, policy, accountability model and assessment methodology - typically takes eight to twelve weeks. Discovery and the highest-risk remediation start in week one rather than after a report. ISO 42001 certification, if that is the goal, takes six to twelve months in total because an auditor needs to see the management system operating over a period.

Will AI governance slow our engineering teams down?

A governance layer that slows teams down gets routed around, which is worse than not having one. The design principle is proportionality: a model ranking internal search results should pass through in minutes, while a model screening job applicants gets a full assessment. Most of the friction people fear comes from applying uniform rigour to every system regardless of what it does.

Free tools

Answer it yourself first.

Nothing is gated. Each one shows you the answer before it asks you for anything, and every date and rate carries the source it came from.

Every tool we publish lives at cloudgap.ai/tools.

Start

Three ways in, depending on how close the deadline is.

Or just email a human: hello@cloudgap.ai

We reply within24 hours