AI cybersecurity consulting

Knowing what the AI Act requires of you, which security risks each use brings and which controls you need. It starts with an applicability review, system by system.

When it makes sense

When AI is already in use and nobody has decided how

In most companies AI does not arrive with a project. It arrives with a Copilot licence, a browser extension, a new feature in the CRM or someone in sales using their personal ChatGPT account. When someone asks what is being used, with what data and what the AI Act requires, nobody has the full answer.

This service brings order from a cybersecurity standpoint: what applies to you and by when, which risks each use brings and which controls you need to keep using AI without improvising.

AI is already inside and nobody has counted it

Staff use assistants on their own, the tools you pay for switch on AI features with every update, and there is no inventory of what is used or with what information.

You are being asked and have no written answer

A client sends a questionnaire about your use of AI, the board asks about the AI Act, or the Omnibus has moved dates and it is unclear which ones affect you.

You are about to build or buy AI

A chatbot, an agent connected to your systems or an AI recruitment tool. It pays to know how it will be protected and which obligations come with it before you sign or deploy.

Applicability review

What the AI Act requires of you, system by system

It is where almost every engagement starts. It begins with an inventory of what is actually in use and ends with a matrix that answers four questions for each system.

  1. Is it an AI system?The definition in Article 3 leaves out software sold as “intelligent” and takes in tools that nobody had counted.
  2. What is your role?Provider, deployer, importer or distributor. Putting your name on a high-risk system, modifying it substantially or changing its intended purpose can make you a provider.
  3. Which risk level does it fall into?Prohibited practice, high-risk under Annex I or Annex III, transparency obligations, or only the general obligations.
  4. What does it require and from when?Obligations with their dates, together with where it overlaps with the GDPR, NIS2 or the Cyber Resilience Act.
Applicability matrix Illustrative example

Generative AI assistant for drafting and summarising

Role
Deployer
Classification
Minimal risk
From
Already applies
What needs resolving
AI literacy, an acceptable use policy and which information may go into the assistant

Automated screening of job applications

Role
Deployer
Classification
High-risk · Annex III
From
2 Dec 2027
What needs resolving
Human oversight, log retention and information for candidates and staff

Your own website chatbot built on a third-party model

Role
Provider and deployer
Classification
Transparency · Art. 50
From
Already applies
What needs resolving
Telling people they are talking to an AI and protecting it against prompt injection

The actual classification depends on how each system is used and is documented case by case, with its reasoning. The second example is worked through in the guide to AI in recruitment.

What's included

Three blocks of work

They can be engaged together or separately. It almost always starts with the first, because it decides how much of the other two you need.

Applicability review

What the AI Act requires of you and how it fits with what you already comply with.

  • Inventory of AI systems and uses, including AI that arrives inside tools you already pay for
  • Your organisation's role and the risk classification of each system
  • Obligations and timeline, cross-checked against the GDPR, NIS2 and the CRA
  • Answers for clients who ask how you use AI

AI governance

Rules, owners and a process for deciding before AI turns up on its own.

  • AI acceptable use policy and rules for generative AI
  • Roles, approval of new use cases and a register
  • Assessment of AI suppliers and tools
  • Gap analysis against ISO/IEC 42001
  • AI literacy measures by role

Security of AI systems

The risks specific to AI, built into the usual security analysis.

  • Threat modelling of applications and agents: prompt injection, data leakage and excessive permissions
  • Review of permissions and data exposure before rolling out assistants such as Microsoft 365 Copilot
  • Security requirements for projects and suppliers that integrate AI
  • Logging, human oversight and incident response
How I work

How an engagement is set up

Four phases, with scope and days agreed before starting. If you only need the applicability review, the engagement ends at the third.

  1. Inventory

    Which AI is actually in use: what you pay for, what comes switched on inside other tools and what people use on their own.

  2. Applicability

    Role, classification and obligations for each system, with their dates and where they overlap with the GDPR, NIS2 or the CRA.

  3. Priorities

    Gaps ranked by risk and deadline: what must be resolved now, what can wait and who owns each item.

  4. Implementation

    Policy, controls and training put in place with your team, plus an inventory that can be kept up to date when the next tool arrives.

What you get

What you have when the engagement closes

A written answer to what applies to you, and the plan to resolve what is missing.

  • Inventory of AI systems and uses
  • Applicability matrix for each system
  • Timeline of obligations with their dates
  • Security risk analysis for each use case
  • AI acceptable use policy and approval process
  • Action plan prioritised by risk and deadline

Scope of the serviceThe work is technical and compliance-focused: what applies to you, what is missing and how to resolve it. It does not replace legal advice; when a case is in a grey area, such as whether a system is high-risk or whether a modification makes you a provider, I document it with its reasoning so that your legal adviser can confirm it. I don't issue ISO/IEC 42001 certificates or conformity assessments, I don't run penetration tests against models and I don't resell AI tools.

Frequently asked questions

What people usually ask me before hiring

Does the AI Act apply to us if we only use ChatGPT or Copilot?

Yes, although usually in its lightest form. Using a general-purpose assistant to draft or summarise is not a high-risk use, but it makes you a deployer: you have to take AI literacy measures for the people who use it and make sure nobody uses it for something that is classified, such as filtering CVs. On top of that, whatever is pasted into the assistant is still subject to the GDPR and to your confidentiality commitments.

Hasn't the Omnibus pushed everything back to 2027?

It has postponed the high-risk obligations: to 2 December 2027 for Annex III systems and to 2 August 2028 for Annex I systems. Prohibited practices have applied since February 2025 and are extended on 2 December 2026, and the AI literacy duty still applies, now worded as taking measures to support it. It is all explained in the article on the Omnibus.

Does this replace legal advice?

No. My job is to identify what applies to you, what is missing and how to resolve it with controls, processes and training. When the answer depends on interpreting the regulation in a borderline case, I document it with the arguments for and against so that your legal adviser can settle it. That tends to be a short conversation, because the technical work is already done.

Do we need ISO/IEC 42001 certification?

It is not mandatory, and being certified does not mean you comply with the AI Act either. It is a good framework for organising AI governance and demonstrating it to clients, and the applicability review is a sound basis for the statement of applicability the standard requires. You can gauge your starting point with the ISO 42001 checklist; certification, if you go for it, is issued by an accredited body.

What should we review before rolling out Microsoft 365 Copilot?

Permissions, above all. Copilot answers with whatever each user can already open, so folders shared with the whole organisation or sites without an owner end up surfacing in an answer. Before the rollout it is worth reviewing data exposure, labelling and data loss prevention policies, work that overlaps with Microsoft security and DLP.

How long does an applicability review take?

It depends on the number of systems and on whether there is in-house development. An organisation that uses third-party AI for internal tasks can be reviewed in a few days; if there are high-risk cases or an in-house product with AI, it takes weeks. Scope and days are agreed before starting, so you know the cost from the outset.

Where to go next

Let's talk

Using AI and unsure what it requires of you?

Tell me which tools you use, whether anything has been built in-house and who is asking. One conversation is enough to see whether a few days of applicability review will do or whether it makes sense to set up full AI governance.

I reply personally within 24 working hours · No commitment