AI governance · AI security · Agentic AI
Your teams are already using AI — in products, in pipelines and, increasingly, as agents acting on your systems with their own credentials. The question your board, your customers and your regulators are asking is whether any of it is governed. We answer it from the CISO's chair: a governance framework people actually follow, and security controls that treat models and agents as what they are — new identities, new attack surface, and new things to monitor.
What you get
Days 1–60
Typically within sixty days: an inventory of every AI tool in use including shadow adoption, data classification and DLP applied to prompts, retrieval and code assistants, and an AI policy people can follow — so source code, customer data and trade secrets stop leaving through the side door.
Steady state
An AI Governance Council and review process sized to the real pace of adoption: intake, risk tiering, model cards and approval running as a weekly cadence rather than a quarterly bottleneck — so governance speeds the business up instead of slowing it down.
Agentic AI
The failure most organisations haven't seen yet: agents running on borrowed human credentials, with more privilege than any person would be given, acting across systems nobody scoped. We give every agent an identity, least privilege, scoped tools and a human-in-the-loop where it matters.
Alignment
Everything maps to the EU AI Act, ISO/IEC 42001, the NIST AI RMF, the OWASP Agentic AI Top 10 and — for financial services — PRA SS1/23 model risk management from day one — so the ISO 42001 certification, the customer security questionnaire and the regulator's questions are answered by the same artefacts.
The engagement
A fixed-scope engagement, typically three twenty-day blocks with their own deliverables — so you know at the outset what you'll have at the end of each. Run to that shape, start to finish, for an FX trading firm and bank under PRA and FCA supervision. Delivered by a Global CISO who has built this as the accountable executive, not a bench.
AI and agent inventory including shadow AI · risk tiering of every use case · data classification and DLP on prompts, retrieval and code assistants · interim AI acceptable-use policy · landscape mapping: which laws and standards bind you, and the roles that carry each accountability.
AI Governance Framework aligned to the EU AI Act, ISO 42001 and NIST AI RMF · AI Governance Council and AI Architecture Review Board with terms of reference and decision rights · AI policy · AI checklist for every new model and agent · model cards and AI registry · AI risk register and appetite.
IAM for AI: an identity per agent, least privilege, scoped tokens and tool allow-lists, PAM for agents · SIEM for AI and AI detection & response · guardrails and prompt-injection defences · the weekly review cadence running at two or three models a week · handover, training and the ISO 42001 readiness position.
What you're left with
“An agent with a human's credentials and no human's judgement is the most privileged user in your estate. Most organisations have several and don't know it.”
— On why agents need identities of their own
Who this is for
CTO / CIO
Engineering is shipping AI weekly. You need a governance process that keeps pace — a review board that approves in days, not a committee that meets quarterly.
Board / investors
A clear statement of AI risk appetite, an accountable executive, and evidence that the controls exist — in language the audit committee and a PE operating partner both accept.
CISO
Someone who has held the accountability, built the framework and argued it through with the CTO, legal and the DPO — to accelerate yours rather than start from a blank page.
Compliance / DPO
EU AI Act operator obligations, ISO 42001 readiness, GDPR Article 22 and sector rules — answered by one set of artefacts rather than four parallel efforts.
Secure, governed AI
Agents don't just answer; they read, decide and act. That creates a threat surface most security programmes haven't mapped yet. This is how we think about it from the CISO's chair: eight ways agentic AI goes wrong — drawn from the OWASP Agentic AI and LLM Top 10 — mapped against the six NIST CSF 2.0 functions, with the controls we build under each, drawn from the NIST AI RMF and ISO 42001 — from an AI Architecture Review Board and model cards, through an identity for every agent, to SIEM for AI and recovery that can actually roll an agent back.
The AI governance landscape
Before a framework can be designed, the landscape has to be mapped: which regulations actually bind you, which standards an assessor will measure you against, who in the organisation carries each accountability, and the seven control domains every external reviewer expects to find. This is the anonymised map we start every AI governance engagement from; the detailed capability model and crosswalks behind it are built for each client.
Registry of every model and agent, including shadow adoption
Tiering and impact assessment before deployment
Evaluations, red teaming, bias testing, model cards
Review, override and escalation of AI-influenced decisions
Agent identity, least privilege, tool permissions
Classification, lineage, DLP, retention
Logging, audit trail, SIEM for AI, incident response
Recent mandates
Tell us what's in use and what's planned. We'll give you a straight view of your exposure and what sixty days could change.