Govern AI · Layer 02
AI risk management frameworks that hold up
Most firms already have a risk framework. Very few have one that copes with models that change without anyone touching them, AI bought inside vendor products, or agents that act on their own.
We help fintechs, technology companies and asset managers build an AI risk framework that plugs into the one they already run: clear appetite, a proportionate assessment method, named owners, and evidence a regulator can follow.
01 · What good looks like
What is good AI risk management?
AI risk management is how your firm decides which AI risks it will accept, which it will treat, and who answers for each, across the whole life of every AI system.
Good AI risk management is not a policy or a register on its own. It is a working system with five parts that connect:
- Risk appetite and thresholds the board has agreed
- A repeatable way to assess each AI use case
- A risk register with a named owner for every risk
- Controls and treatment, matched to the level of risk
- Monitoring and reporting that triggers a re-assessment when things change
You can tell a framework is good when three things are true:
- It is proportionate. A credit model and a meeting summariser do not get the same paperwork.
- It is used. Decisions, overrides and re-assessments actually happen, and leave evidence.
- It is joined up. AI risk sits inside the enterprise risk framework and is reported in the same language as every other risk.
Every AI risk then ends in one of four decisions:
02 · Why do it
Why build an AI risk framework now
Because existing risk frameworks were built for systems that only change when someone changes them, and AI does not behave that way:
- It changes on its own. Models drift as data shifts, and vendors update the models inside their products without telling you.
- The risk sits outside the model. In the data it reads, the vendor that supplies it, and the agent permissions that let it act.
- Oversight exists on paper only. A human sign-off nobody exercises, and a risk register nobody revisits.
Source: IBM Cost of a Data Breach Report 2025.
What regulators and standards expect
UK banks with internal model approval, in force since May 2024
Model risk principles: a model inventory, risk-based tiering, independent validation, and a senior manager accountable for model risk. AI and machine learning models that meet the model definition are in scope.
The template UK supervisors already know. Extend it to AI.
US banks, most relevant above $30bn in assets
April 2026 guidance from the OCC, Federal Reserve and FDIC replaced SR 11-7. It explicitly excludes generative and agentic AI, leaving those to broader risk management.
Generative AI and agents now need a framework of their own.
UK firms
No AI-specific rules. The FCA relies on the Consumer Duty and SM&CR, and is reviewing how senior managers oversee AI.
A senior manager must be able to show how AI risk is identified and controlled.
EU financial firms, including asset managers and payment fintechs
Requires a documented ICT risk management framework, reviewed at least yearly and after major incidents.
AI systems are ICT. Their risks belong inside that framework.
US investment advisers
The 2026 examination priorities check AI-related disclosures for accuracy and whether policies and procedures supervise AI use.
Claims about AI need a risk process behind them.
Providers of high-risk AI used in the EU
Article 9 requires a risk management system run as a continuous, iterative process across the life of a high-risk system. Credit scoring is high-risk.
Risk management is a legal duty for high-risk systems, not a one-off review.
Voluntary, widely used in the US and by tech firms
Organises AI risk work into four functions: Govern, Map, Measure and Manage.
A ready-made structure for your framework.
Certifiable AI management standard, with risk guidance
ISO/IEC 42001 requires AI risk assessment, risk treatment and AI system impact assessment, repeated as systems change. ISO/IEC 23894 applies ISO 31000 risk management to AI.
Certification depends on a working risk process.
Summaries for orientation, as of October 2026. Check how each applies to your firm.
03 · How to decide
Seven decisions that shape a fit-for-purpose framework
There is no off-the-shelf AI risk framework that fits every firm. A robust one comes from answering seven questions in order. Each answer narrows the next.
Decision 1 · ScopeWhat AI are we managing?
Build from the AI inventory: in-house models, vendor AI, AI features in licensed software, and agents. If you cannot list it, you cannot assess it.
Decision 2 · Regulatory lensWho do we answer to?
Map your regulators and certifications. A UK bank under SS1/23, a US adviser under the SEC and an EU provider of high-risk AI need different evidence from the same framework.
Decision 3 · FoundationWhat can we build on?
Decide whether to extend enterprise risk, model risk, or both, and which standard gives the structure: NIST AI RMF, ISO/IEC 42001 or ISO/IEC 23894.
Decision 4 · AppetiteHow much risk will we accept?
Set AI-specific limits: which decisions AI may make alone, which data it may touch, and what error rates are tolerable for each type of use.
Decision 5 · ProportionalityHow deep should each assessment go?
Define a triage that sorts use cases into tiers, so effort goes where the risk is and low-risk tools are not slowed down.
Decision 6 · AccountabilityWho owns what?
Name a business owner for every AI system, give the risk function the challenge role, and map senior accountability to structures such as SM&CR.
Decision 7 · AssuranceHow will we know it works?
Agree the key risk indicators, re-assessment triggers and independent testing that prove the framework is used, not just written.
Decision 3 in detail
What you can build on
There are several established ways to manage AI risk, and most firms already use one or two. Each covers part of the problem. The gaps between them are where AI incidents usually start.
Extend your enterprise risk framework
Gives youFamiliar taxonomy, appetite and reporting lines. The board already understands it.
MissesAI-specific failure modes: drift, hallucination, prompt injection, agent actions.
Model risk management
Gives youStrong on validation, tiering and independent challenge for quantitative models.
MissesGenerative AI, vendor AI and agents, which many model definitions now exclude.
GRC platforms
Gives youWorkflows, registers, attestations and reporting you may already license.
MissesThe AI risk method itself. A platform stores risks; it does not identify them.
Specialist AI governance platforms
Gives youAI-specific questionnaires, policy mapping and approval workflows.
MissesYour context and appetite. Templates can push every use case to the same depth.
Standards-based frameworks
Gives youA credible structure from NIST AI RMF, ISO/IEC 42001 or ISO/IEC 23894.
MissesHow to make them work inside your firm, at your scale.
Policy only
Gives youA quick signal of intent, often an AI usage policy.
MissesEverything after the policy: assessment, controls, monitoring, evidence.
Approaches are complementary. T3 is independent of the platform vendors in this market.
Putting the decisions together
What this usually means for your firm
There is no single right framework. The starting point depends on the risk functions you already have and who you answer to.
Fintech
Fast release cycles, lean risk team, AI in customer journeys.
- Start with
- Your enterprise risk framework, extended with an AI risk taxonomy and a light register in your GRC tool.
- Watch for
- Vendor AI in onboarding, fraud and credit decisions, and whether you can explain an outcome to a customer.
- Regulatory lens
- FCA Consumer Duty and SM&CR, DORA for EU entities, and the EU AI Act if you score credit.
Technology company
AI built into products, with internal agents and frequent model updates.
- Start with
- NIST AI RMF or ISO/IEC 42001 as the backbone, with risk gates built into your release process.
- Watch for
- Model updates shipped to customers, agent permissions, and AI commitments made in customer contracts.
- Regulatory lens
- ISO/IEC 42001, NIST AI RMF, and EU AI Act provider duties.
Asset manager
Established operational and model risk functions, many data and research vendors.
- Start with
- Your operational and model risk frameworks, extended to cover generative AI and vendor AI.
- Watch for
- AI in research and portfolio tools, Copilot with client data, and AI claims in marketing.
- Regulatory lens
- SEC for US advisers, the FCA, and DORA for EU firms.
04 · How we help
How we help, and why it stays proportionate
We do not hand you a new framework to bolt on. We extend the one you have, using NIST AI RMF and ISO/IEC 42001 as the structure, and calibrate it to your appetite and your regulators.
Then we prove it works on your highest-risk AI systems first, so the framework is tested before it is rolled out.
Extending your enterprise risk framework
Your risk team, with T3
AI risk reported in a language your board already uses
Setting AI risk appetite and thresholds
Workshops with your executive and risk committee
Decisions you can defend
Assessment method, taxonomy and first assessments
T3
Needs AI, risk and regulatory depth together
Running controls day to day
Your control owners, tested by T3
Ownership stays with you
Depth follows risk. A credit decisioning model gets a full assessment and independent challenge. An internal summarisation tool gets a short triage and a named owner.
What you get
Five deliverables
Each one is usable on its own. Together they give you an AI risk framework your board can approve and your regulator can follow.
- 01
AI Usage Policy
What staff may and may not do with AI, which tools are approved, and how to request new ones. Short enough to be read.
ISO/IEC 42001FCA SM&CR - 02
AI Risk Assessment
A repeatable method and template covering data, model, vendor, security, conduct and operational risk, scaled to the risk tier.
NIST AI RMF Map and MeasureISO/IEC 23894 - 03
AI Risk Register
Every AI risk with its rating, owner, controls and review date, held in your GRC tool and linked to the AI inventory.
ISO/IEC 42001DORA - 04
AI Impact Assessment
The effect of a system on customers, staff and markets, including fairness and explainability, for the use cases that need it.
ISO/IEC 42001EU AI Act Art. 9 and 27 - 05
Human Oversight Design
Where a person reviews, approves or can override, with escalation paths and evidence that oversight actually happens.
EU AI Act Art. 14NIST AI RMF Manage
Timing
A working framework in about eight weeks
Review your enterprise risk framework, AI inventory and existing policies, and agree scope.
Design the taxonomy, appetite, assessment method and usage policy with your risk team.
Run the first assessments on your highest-risk systems, build the register and design oversight.
After that, assessments run through the framework as new AI arrives, and we can provide independent challenge on your highest-risk systems each year.
05 · Questions
What leaders ask us
Do we need a separate AI risk framework?
Usually not. The most effective approach extends your enterprise risk framework with an AI risk taxonomy, appetite and assessment method, so AI risk is reported alongside every other risk rather than in a silo.
Our models are covered by model risk management. Is that enough?
Often not. Model risk frameworks focus on quantitative models, and the April 2026 US banking guidance explicitly excludes generative and agentic AI. Vendor AI and AI features in software also tend to fall outside model definitions.
How do we set AI risk appetite?
Start from your existing risk appetite statement and add AI-specific limits, for example which decisions AI may make without human review, which data it may use, and what error rates are tolerable for each use case.
Who should own AI risk?
Each AI system needs a named business owner, with the risk function setting the framework and providing challenge. Senior accountability should map to your existing structure, such as SM&CR in the UK.
Does the framework cover vendor and third-party AI?
Yes. Most AI risk now arrives through suppliers, so vendor AI is assessed with the same method, using supplier questionnaires and contract terms where you cannot see inside the model.
How is this different from AI compliance?
Compliance asks whether you meet specific rules. Risk management asks what could go wrong and whether it is acceptable. A good risk framework makes compliance much easier to evidence.
Get started
Talk to us about your AI risk framework
A 45-minute call with a T3 specialist. We will look at the risk framework you already run, where AI falls through its gaps, and what a proportionate next step would be.
