Nobody has the full list of AI in use
Models in the product, copilots adopted by teams, agents wired into internal systems. Discovery almost always finds more than the org chart predicted.
By industry
Governance for models and the identities they use
Governance for models and the identities they act through.
The models are already in production, acting through credentials nobody inventoried, on data nobody classified for that purpose. Governance is arriving after the fact — the work is catching it up without stopping the roadmap.
The regulatory picture
What this sector has to satisfy
A certifiable management system, the security work underneath it, and the data protection rules that apply the moment training data contains people.
Where you are
The constraints at this stage
What tends to be true when companies come to us at this size.
Agents act through identities with no lifecycle
Service accounts, API keys and tokens created to let something act on a user's behalf, then never rotated, scoped down or revoked.
Buyers have started asking, and you have no answer
Enterprise security reviews now include AI questions — training data, retention, human oversight — and there is no artefact to point at yet.
What we run for you
The work itself
Delivered by our specialists, not handed to you as a template pack.
AI SecurityDiscovery of AI in use, threat modelling for model integrations and agentic workflows, and guardrails that survive production.AI IdentityThe non-human identities AI creates — inventoried, least-privileged, and given issuance, rotation and revocation.ISO 42001An AI management system you can certify, giving buyers and regulators a recognised standard to assess you against.AI Agents & Agentic WorkflowsWhere agents act on your systems, we design the guardrails and the human-in-the-loop points — and build them.GRC ProgramAI governance inside the same programme as everything else, rather than as a parallel discipline nobody owns.
How the engagement works
What actually happens
The same four beats every time, scoped to what this stage needs.
Discover what is already running
Models, copilots, agents and the integrations behind them, including the ones adopted outside IT. The inventory is usually the finding.
Threat model the workflows
Prompt injection, data leakage, output handling and what an agent can reach if it misbehaves — assessed per workflow, not in the abstract.
Control the identities and the data
Least privilege for agents and workloads, credential lifecycles, and lawful basis and retention for whatever the models were trained on.
Certify if it helps
Where buyers or regulators want independent assurance, ISO 42001 turns the programme into something assessable.
Proof
Track record
Across every engagement we have run, at every size.
0+
Companies Secured
Zero
Failed Audits
4-8
Weeks to Compliance
0+
Team Certifications
What you walk away with
The artefacts
Everything below is a thing you own at the end, not a promise about the process.
- An inventory of AI systems, models, copilots and agents in use
- Threat models for each production AI workflow
- Guardrails, monitoring and defined human-in-the-loop control points
- A non-human identity register with least privilege and rotation
- Training and inference data mapping with lawful basis and retention
- AI policy and acceptable use, written to be followed rather than filed
- ISO 42001 management system, certified where that is wanted
Related
Where to go next
The problems this stage usually arrives with, and the services behind them.
Questions
What people ask at this stage
It is worth it when someone is asking — enterprise buyers, a regulator, or a board that wants independent assurance. If nobody is asking, the underlying governance work still matters and the certificate can follow later.
Yes, and often more urgently. You inherit the model's behaviour without controlling it, so the controls sit in what you send, what you do with the output, and what the integration is allowed to reach. That is squarely your responsibility.
Agents act through credentials. Those identities need the same governance human ones get — an owner, least privilege, rotation, expiry and review — except there are usually far more of them and they are created programmatically.
If it contains personal data, all of it applies: lawful basis, purpose limitation, retention, and data subject rights. Rights are the hard part, because deletion requests against a trained model are considerably harder than against a database row.
Govern what is already in production.
Tell us what you have shipped and what your buyers have started asking. We will start with discovery, because that is where the surprises are.