Feature AI Risk
Retiring an AI Model Is a Control, Not a Cleanup Task
AI model retirement is a control, not a cleanup task. The decommission memo, retention surface table, and inventory states that survive an exam.
Table of Contents
TL;DR
- Retirement fails at the handoff, not the shutdown. The model stops scoring, the reason codes that produced last quarter’s adverse action notices stop being reproducible, and 12 CFR 1002.12(b) keeps the 25-month retention clock running on those notices.
- SR 26-2, issued April 17, 2026, replaced SR 11-7 and OCC Bulletin 2026-13 rescinded the OCC’s Model Risk Management handbook. Neither adds a decommissioning section. Your policy is now the standard.
- The inventory needs explicit retirement states, or the model slides into a status where it is “not in use” but still writing decisions in an open path. That gap is where examinations find things.
The kill switch gets all the attention. Retirement is where the evidence dies.
Emergency shutdown is a satisfying problem. One button, one rehearsal, a clean story for the risk committee. Planned retirement is the unglamorous version that runs for six months, involves four functions, and has no demo.
It is also where programs actually break. Not because the model keeps running, but because everything downstream of it keeps running: the reason-code text on yesterday’s denials, the score field a collections queue still sorts on, the vintage comparison in the monthly pack, the challenger that was only ever benchmarked against the retired champion.
A kill switch is built to get you back. A retirement is built to get you out — and out is much harder to evidence.
Retirement is three tests, not a date
Strip the ceremony away and “retired” is a claim that three things are true at once.
1. Out of every live decisioning path. Not out of the main application. Out of all of them, including the batch job nobody documented, the manual override spreadsheet the credit team keeps, and the vendor portal that still reads a score field.
2. Out of the active inventory, into a terminal state. SR 26-2’s inventory language is about models “under development or in use,” and an effective inventory holds “sufficient information to understand model risks.” A model that is genuinely retired leaves the active population. A model that is partly retired does not, and that distinction has to exist as a field value rather than a note in a wiki.
3. Outputs still reproducible for the full retention window. The decisions the model already made are not retired. They are pending records, complaints, reconsiderations, and possibly litigation.
Fail any one of the three and you have renamed a model, not retired it.
The April guidance change made retirement harder to bluff
For fifteen years, model risk conversations in US banks started from SR 11-7, issued April 4, 2011. That letter is gone. SR 26-2, issued jointly with the OCC and FDIC on April 17, 2026, supersedes SR 11-7 and SR 21-8 and replaces them with risk-based guidance. The OCC simultaneously rescinded its Model Risk Management handbook.
Two features of the replacement matter for retirement.
The first is the enforceability framing. SR 26-2 states that the guidance “does not set forth enforceable standards or prescriptive requirements; accordingly, non-compliance with this guidance will not result in supervisory criticism.” Some teams read that as breathing room. The practical effect is the opposite: when the guidance stops prescribing, the adequacy question moves to whether your own policy was followed and evidenced. A retirement that contradicts your written decommissioning standard is now a policy exception, and policy exceptions are what examiners can write about.
The second is scope. SR 26-2 applies its principles to “traditional statistical and quantitative models and non-generative, non-agentic AI models.” Generative AI and agentic AI are described as “novel and rapidly evolving” and sit outside the guidance, with an interagency request for information promised. So the retirement question splits:
| Model type | Governing standard for retirement | Evidence you owe |
|---|---|---|
| Traditional quant model (scorecard, PD/LGD, ALM, stress) | SR 26-2 principles plus your MRM policy | Full retirement record, reproducibility package, validation sign-off |
| Non-generative AI model (ML classifier, NLP triage) | Same as above | Same as above |
| Generative AI in a decisioning path | Internal AI governance standard, not SR 26-2 | Retirement record plus prompt/model version pinning and affected-output log |
| Agentic AI with tool access | Internal AI governance standard plus change management | Above plus revoked credentials, disabled tool scopes, and action audit trail |
The genAI row is the one most programs have not built. A retired chatbot that wrote customer-facing answers has a broader blast radius than a retired scorecard, because nobody captured reasons at the time.
Three surfaces that outlive the model
The reason retirement is a control rather than a project task: the model’s outputs are load-bearing records with their own owners, clocks, and legal hooks.
| Surface | What persists after shutdown | Clock | Owner |
|---|---|---|---|
| Adverse action notices and reason codes | The CFPB’s position that reasons must be “specific” and identify the “principal reason(s)” — you cannot answer a complaint by pointing at a model that no longer exists | 25 months from notice under 12 CFR 1002.12(b), 12 months for business credit | Credit policy / fair lending |
| Credit and capital reporting inputs | Model-estimated PD, LGD, allowance inputs, and stress results feeding prior-period filings | Life of the reporting cycle plus the retention period for the underlying report | CFO / financial reporting |
| Servicing and collections decisions | Score-based queue placement, line decreases, workout tiering | Per account record retention schedule | Servicing operations |
The CFPB made the adverse action point explicitly in Circular 2022-03 and again in Circular 2023-03: complexity is not a defense, and a creditor “must ‘disclose the actual reasons for denial… even if the relationship of that factor to predicting creditworthiness may not be clear to the applicant.’” Notice that neither circular cares whether the model is still running.
So a real retirement needs a frozen reproduction package, and the package is boring on purpose:
- The model object or executable, containerized and version-pinned, with the exact library versions
- A snapshot of the reason-code mapping as of each version that produced live notices
- The input extract for every decision still inside its retention window, keyed to application ID
- A signed, dated reproduction test showing the package regenerates the same reason text for a sample of at least 30 historical decisions
That last item is the one people skip. Run it, date it, keep it. A reproduction package that was never executed is a hypothesis.
The retirement decision memo
Write this before shutdown, not after. Ten fields, one page.
- Model ID, version, and every decisioning path it still touches. List systems, not teams. “Credit decisioning engine, batch 4 nightly, plus manual override sheet owned by the credit desk” — not “lending.”
- Retirement reason. Replacement, decay, vendor exit, scope reduction, cost. Be honest; “replacement” with no live replacement is really “decay.”
- Replacement model ID and validation status. Name it, confirm validation is complete, attach the date. If the replacement is not yet validated, the retirement date moves.
- Parallel run result. Both models scored the same population for a defined period, with disagreement rate and the disposition rule for each disagreement band.
- Open findings being closed permanently. Anything still open on the retired model is now un-remediable. List it and name who accepts that.
- Outputs still in flight — applications, complaints, reconsiderations, workout accounts — and how each will be reproduced.
- Retention surface and destruction date, derived from 12 CFR 1002.12(b) and the account retention schedule, whichever is longer.
- Access removal evidence. Named accounts revoked, vendor read-only links cut, score field decoupled from downstream jobs.
- Reproduction test result, with the date it ran and the sampler’s name.
- Three signatures — model owner, model risk management, accountable executive.
Field 5 is the field that changes behavior. Once someone has to sign that an open finding will never be fixed, “just retire it” stops being free.
Inventory states that stop shadow use
Two states — “in use” and “retired” — guarantee a gap, because most retirements pass through a middle phase that lasts months. Give the middle a name.
Example enum, five values, with the allowed transitions:
| Status | Meaning | Allowed next states |
|---|---|---|
Active | In a live decisioning path | Parallel, Sunset Candidate |
Parallel | Replacement live, original still scoring, shadow only | Active (back out), Sunset Pending |
Sunset Pending | Date set, reproduction package built and tested | Parallel (slip), Frozen |
Frozen | Not scoring; object preserved; outputs still reproducible | Destroyed |
Destroyed | Retention expired, artifacts destroyed with certificate | — |
The rule that makes it work: a model may only sit in Parallel while a validation of the replacement is open. No validation, no parallel status, and the model is Active — which means it keeps showing up in the active count and keeps drawing monitoring.
Add a weekly reconciliation query: every model in Parallel or Sunset Pending older than 90 days appears on the model risk committee agenda with a reason for the slip. This is the same reconciliation discipline as keeping the AI use-case register and the model inventory aligned, applied to the exit door instead of the entry door.
Parallel run exit criteria
A parallel run with no stop condition runs forever, because the person running it is never wrong. Write the exit rule before you start, and write it in numbers.
Starter thresholds, to calibrate against your own history rather than adopt:
- Disagreement rate no more than 3 percentage points above the incumbent’s own month-to-month variance on the same population.
- Approve/decline flips fully dispositioned for 100% of cases where the two models disagree across the decision boundary.
- Reason-code agreement on at least 95% of jointly declined applications, with every divergence logged.
- Population stability index on each material input below 0.10 across the parallel period, so the comparison is like-for-like.
Calibrate all four against the last three to six months of incumbent behavior before publishing them to anyone. A 3-point band borrowed from another institution’s deck is a guess with a source code.
Anti-gaming check, because a parallel run is easy to game: require the disagreement sample to be drawn by model risk management from the raw decision log, not supplied by the business line, and require the flipped-case dispositions to include the actual notice text issued. If the flips show up as “approved after manual review” with no notice, the parallel run measured something other than model behavior.
Who owns what
| Activity | Owner | Artifact that proves it happened |
|---|---|---|
| Business case and replacement readiness | Model owner (1st line) | Decision memo fields 1–4, validated |
| Reproduction package build and test | Model development | Dated test output, 30-decision sample |
| Independent reproduction review | Model risk management (2nd line) | Review note with pass/fail |
| Access removal | Application owner / IAM | Revocation export with timestamps |
| Open-finding acceptance | Accountable executive | Signed risk acceptance |
| Retention and destruction | Records management | Destruction certificate at expiry |
At a bank over $30 billion, the second line is model risk management. At a fintech without one, it is the head of compliance plus an external validation vendor, and the external review note is the artifact — but the independence has to be real, not a rubber stamp on a document the first line drafted.
Three failure modes worth naming
Retirement by spreadsheet. The model is deleted, but an analyst’s override workbook still carries a hard-coded copy of its output because the batch job broke and nobody wanted to re-plumb it. Six months later the “retired” model is deciding. Detection: grep downstream job configs for the score field name after shutdown, and keep the grep output in the file.
The frozen orphan. Files archived, retention clock forgotten, destruction date never set. Three years later the storage team deletes an entire bucket during a cost sweep and the notices become unanswerable. Detection: every Frozen record carries a computed destruction date at the moment of freeze, not a hand-entered one.
GenAI retirement by vendor deprecation. The model provider sunsets a version and the program’s only action is to switch endpoints. Prompts, sampling parameters, and retrieved-document snapshots were never pinned, so nobody can reproduce anything the old version said. Detection: version-pin the endpoint, snapshot the retrieval corpus, and log prompt hashes per release — before the deprecation notice arrives.
So what
Retirement is the cheapest place in the model lifecycle to look competent, and the easiest place to look careless. It is also the only lifecycle stage where a decision made this month creates a 25-month obligation.
The work is genuinely small: a one-page memo, a five-value status enum, an executed reproduction test, and three signatures. What makes it hard is that all four are unglamorous and none of them demo well.
Start with the model your team has been meaning to retire the longest. Write the memo. If field 5 has an entry nobody will sign, you have found the reason it has not happened yet — and that is a more useful finding than another monitoring report.
The AI Risk Assessment Template & Guide includes the use-case inventory with auto-tiering, the shadow-AI register that catches decision paths outside the register, and a governance dashboard built to carry status transitions and board reporting.
Related: what SR 11-7’s rescission actually changed, the kill switch controls that pair with retirement, and reading OCC Bulletin 2026-13 clause by clause.
◆ Need the working template?
Start with the source guide.
These answer-first guides summarize the required fields, evidence, and implementation steps behind the templates practitioners search for.
◆ Related template
AI Risk Assessment Template & Guide
Comprehensive AI model governance and risk assessment templates for financial services teams.
◆ Immaterial Findings · Weekly
Sharp risk & compliance insights. No fluff.
◆ FAQ
Frequently asked questions.
Does the revised model risk guidance cover decommissioning?
How long must we be able to reproduce reasons for decisions a retired model made?
Can we just archive the model files and call it retired?
What is the difference between a kill switch and a retirement?
Who signs a retirement decision?
Author
Rebecca Leung
Rebecca Leung has 8+ years of risk and compliance experience across first and second line roles at commercial banks, asset managers, and fintechs. Former management consultant advising financial institutions on risk strategy. Founder of RiskTemplates.
◆ Related framework
AI Risk Assessment Template & Guide
Comprehensive AI model governance and risk assessment templates for financial services teams.
◆ Keep reading
Related posts.
AI Risk
Singapore's Regulator Just Made Banks Personally Accountable for AI Their Vendors Built. US Examiners Are Already Asking the Same Questions.
On October 7, 2026, Singapore's Monetary Authority of Singapore finalized AI risk management guidelines requiring financial institutions to remain accountable for all AI in their services — including systems built, operated, or supplied by third-party vendors. US bank examiners are already asking equivalent questions. Here's what the MAS framework reveals about where US AI governance requirements are heading.
Oct 9, 2026
AI Risk
AI Governance Exceptions: How to Approve a Use Case With Open Testing Gaps Without Creating Permanent Waivers
When your AI governance committee approves a use case before validation is complete, you need a real exception process — not just a to-do list. SR 26-2 leaves generative AI out of scope. The CSBS AI Supervisory Framework (September 2026) doesn't. Here's how to build an exception framework that doesn't become a permanent liability.
Oct 7, 2026
AI Risk
Your AML Transaction Monitoring Runs on AI. Your Examiner Is About to Treat It Like a Model Under SR 26-2.
SR 26-2 (April 2026) updated model risk management guidance for the first time since 2011. If your AML transaction monitoring system uses machine learning, it's a model under that guidance — with documentation, validation, and ongoing monitoring requirements your BSA team may not know exist.
Oct 4, 2026