AI GOVERNANCE ASSESSMENT
Find out where AI already is, what it decides, and who approved it.
Most AI in an institution arrived inside software already bought, usually more than expected. We list each use and what it touches, tier it, and name an approver where it matters.
At a glance
- You receive
- Five registers, the AI Policy among them
- Retest
- One round, on Gap List items you report closed
- Fee and duration
- In writing, before you sign
Use it when
- The board, an examiner or an auditor asked which AI is in use and who approved it.
- A vendor switched on an AI feature, and nobody knows what else has.
- A counterparty requires a written AI governance framework, as Fannie Mae and Freddie Mac have since 2026.
Five entry points; the fifth reaches a regulated decision
| Entry point | What is true of it |
|---|---|
| Adopted directly, even on a personal account | No contract, retention term or record; found by asking teams |
| Inside software you license | Arrived by release note, on by default, never approved |
| Inside a vendor's product | The vendor picks, and can change, model, version and location |
| Built or configured here | No documentation, change control or owner |
| Scores, ranks, prices, eligibility, routing | Reaches a regulated decision, core platforms included |
Six questions asked of every use
- What decision does it touch, how close to the customer?
- A drafting tool is not a pricing model.
- What data goes in, from where?
- Systems and categories, not intentions.
- What does the vendor do with the input?
- Retention, training, subprocessing, geography, in writing.
- Who can turn it off, how fast?
- A control nobody can operate under pressure is not a control.
- What record does a decision leave?
- Inputs, version, output, timestamp, human step, or none.
- Who approved it, on what basis?
- A name and a date, or a finding that nobody did.
A vendor that will not answer one of these in writing has produced a finding, and the refusal is recorded as one.
How a use is tiered
Four questions set the tier: person affected? Output reversible? Reason required? Material human review first?
| Tier | The use |
|---|---|
| Tier 1 | Assists a person's work |
| Tier 2 | Shapes a decision about a person |
| Tier 3 | Decides about a person |
A reviewer who approves a queue without the time or the information to disagree is not a control, and the tiering says so.
Eight risk domains behind the six questions
A completeness check, not a certification or a substitute for legal, technical or privacy review.
- Purpose and consequence
- Purpose, decisions touched, materiality, reversibility, failure or outage
- People and rights
- Affected groups, notice, human review, feedback, appeal, correction, remedy
- Data and provenance
- Source, permitted use, quality, representativeness, sensitivity, lineage, retention, deletion, downstream use
- Quality and evaluation
- Fitness, baseline, test population, reliability, uncertainty, thresholds, monitoring, known limits
- Security and misuse
- Access, input manipulation, exposure, model or pipeline abuse, unsafe output, supply chain, logging, incidents, recovery
- Agentic authority
- Tools and credentials held, delegated actions, approval boundaries, transaction limits, separation from production, who can stop it, trace and rollback
- Provider and operations
- Vendor evidence, model changes, subprocessors, availability, support, continuity, portability, fallback, exit, retirement
- Law, policy and accountability
- Duties, contracts, policy, records, owners, decision rights, escalation, exceptions, approvals, review dates
Five registers
| Register | What it holds |
|---|---|
| AI Use Register | Each use by entry point: what it touches, owner, switch off cost, tier |
| Vendor and Data Flow Record | Per use: the vendor, data in and out by category, where it rests |
| Consequence Tiering | Each tier with its reasoning, so it can be argued with |
| AI Policy | What is permitted, what needs whose approval, what is barred, how uses are added |
| Gap List with Owners | What is missing, ranked, each with an owner and the decision needed |
Your people maintain it. Where the institution already keeps a model inventory under its model risk program, the AI Use Register is built as an extension of it, with the same identifiers, so there is one list and not two.
Data categories: customer identifiers, account and transaction data, credit and underwriting inputs, employee data, anything nonpublic.
A gap without an owner returns next year with the same wording.
- approve
- approve with conditions
- pilot
- restrict
- defer pending evidence
- escalate
- reject
- suspend
- retire
Every decision uses one of these nine words and records the rationale, the approver, the conditions, the expiration and the trigger for reassessment.
The registers fill the board's six item packet in the post CAT board packet article.
Bias, harm and recourse
Three separate findings; a use can pass one and fail another:
- Could a person contest the output?
- Can one decision, on one day, actually be explained?
- Is there a route to challenge that somebody must answer?
A missing explanation is named: inputs, version, the human step, or a vendor holding the only copy. Where people may be materially affected, the record names who owns notice, feedback, complaint, appeal or reconsideration, escalation, correction and remedy, with a time expectation and a record.
Five steps, in order
- Scope: areas, platforms, vendors and contacts, with the fee and duration in writing before anything is signed.
- Discovery by entry point: your identity and cloud administrator, vendor file owner, lending, underwriting or claims platform owner and a person per business area; read access to the software inventory, vendor file and contracts; any managed provider named.
- The six questions to every use, and in writing to vendors that hold the answer.
- Tiering and policy, walked through with you before either is final.
- Handover. The five registers, reviewed and signed off before release. This engagement has a finish line, and what follows it is a separate decision made with the Gap List in front of you.
What the registers feed
| Instrument | Why it matters here | Who it reaches, and who it does not |
|---|---|---|
| Freddie Mac Guide Section 1302.8 and Fannie Mae LL-2026-04 | Inventory, tiering and approvals are the framework's first parts to show. | Requires Freddie Mac Guide 1302.8, from 3 March 2026, and Fannie Mae LL-2026-04, from 6 August 2026, require a Seller/Servicer using AI or machine learning in origination or servicing to hold a written governance framework for it. Reaches Any size, through the contract. Not Anyone outside either contract. |
| NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers | The registers are what a department asks to see. | Reaches Insurers in a state whose department adopted it, with that state's variations. Status Guidance, not statute. Not Insurers in a state that has not adopted it. Whether it reaches producers, managing general agents or third party administrators is a state question. |
| Revised interagency model risk management guidance, SR 26-2 and OCC Bulletin 2026-13 | For a community bank the inventory is not going into a model risk framework. | Reaches Banks the OCC, Federal Reserve and FDIC supervise. The agencies state it is most relevant above $30 billion in total assets and place generative and agentic AI outside it; smaller banks' models remain subject to governance appropriate to size and risk profile. Not Credit unions. |
| Interagency Guidance on Third-Party Relationships: Risk Management, 88 FR 37920 | Vendor rows file into its life cycle. | Reaches Every bank the OCC, Federal Reserve and FDIC supervise, no asset size threshold. Status Guidance, not a rule. Not Credit unions; NCUA is not a party. |
A community bank's registers feed its vendor file, information security program and board record. The registers also map onto the voluntary NIST AI Risk Management Framework.
The vendor register is a subset of the security program's vendor file; a Platform Review reads the same tenants, and the Security Assessment asks the same question of the program.
This firm builds the program and the evidence to the shape the instrument requires. Counsel confirms the legal reading for your state.
Where we stop
Red Stick Solutions does not perform statistical fairness testing or independent quantitative model validation. What it builds is the governance, the documentation and the human review structure that makes an answer producible, and no register from this firm is represented as proof that an AI use is compliant.
Before you sign
| Question | Answer |
|---|---|
| Proof of compliance | No. It gives an examiner or counterparty dated evidence to reach their own conclusion. No report from this firm is represented as proof that an environment is secure or compliant. |
| Something urgent | We stop and tell you the same working day, by the escalation path and to the named person agreed in writing before the work began. |
| Retest | One round of verification of remediated findings is included in the fee as standard. Further rounds are quoted, and the proposal states the window. |
| If it stops | You keep everything produced to that point, in the form it is in that day. |
| Is the AI Policy included? | Yes. It is also sold alone; nobody buys it twice. |
| Must we buy the Program afterwards? | No. Some run the registers themselves. |
Ask for the AI Governance Assessment in this scope.
No fee is published: each is fixed against your scope, not an average one.
- Unit
- The scope on this page
- Fee
- Fixed against this scope, in writing, in the first conversation, before you sign or grant access.
- Duration
- Elapsed time and your team's hours, in writing, with the fee.
- A changed scope
- Requoted, showing which change moved the number.
- A number that proves wrong
- You hear the day we know.
The commitment in full is on Trust and Assurance.