The Supplement leaves “high risk” for the company to define, and says regulators may ask how you define it (Materiality and Risk Assessment, p.2). Your tiering scheme is therefore itself exam evidence. This is Nolte’s. NAIC staff floated a four-level sample taxonomy at the Spring 2026 meeting as a discussion draft; it is not adopted text and this scheme does not depend on it.
Why tier
The technical build lists nine controls. Applied uniformly they cost three to four engineer-weeks per system, which is fine for a claims adjudicator and absurd for a marketing copy generator. Uniform governance is how small teams end up with none: the bar is set at the hardest case, nobody clears it, the whole program becomes a PDF.
Tiering does two jobs. It tells you where to spend. And it produces the written risk rationale Exhibit C ref 5 and 6 ask for, per system, in a form an examiner recognises.
The four inputs
Every AI system is scored on four questions. All four map to Supplement definitions.
| Input | Question | Supplement anchor |
|---|---|---|
| Consumer impact | Does the output influence a decision about a specific consumer that is subject to insurance regulation? | ”Consumer Impact”; Exhibit A “direct consumer impact” column |
| Harm severity | If wrong, what is the worst economic outcome for that consumer? | ”Degree of Potential Harm to Consumers” |
| Reversibility | Can the outcome be undone, by whom, and within what time? | Exhibit B narrative Q1g (autonomy, reversibility, reporting impact) |
| Autonomy | Does the system support, augment, or automate? | Definitions; Exhibit C ref 7 |
Financial-statement impact (Exhibit C ref 11) is a fifth input that promotes a system one tier regardless of consumer impact. Reserving and pricing models live here.
The tiers
Tier 0: Internal, no consumer or financial impact
Code assistants, internal search, document summarisation for staff, meeting notes. Test: no output reaches a consumer decision or a financial statement, even indirectly. Required: Control 1 (register entry), Control 9 (vendor register). That’s it. Trap: “internal” tools whose output gets pasted into a consumer letter. If a human copies it forward without review, it is not Tier 0.
Tier 1: Consumer-facing, support only, low harm, reversible
Chatbots answering coverage questions, quote-form assistance, FAQ, document intake that a human then processes. Test: the system informs; it does not recommend or decide. A wrong answer costs the consumer time, not money. Required: Controls 1, 2, 9, plus output content logging with review sampling (the generative addendum), and the consumer-disclosure check against the state tracker (checklist 3n). Trap: a chatbot that says “you’re covered.” That’s a decision. Promote it.
Tier 2: Augment, direct consumer or material financial impact, reversible
Underwriting triage that recommends accept/refer, claims triage and routing, fraud scoring, document extraction feeding a rater, renewal retention models, pricing model development inputs. Test: a human makes the final call, but the system’s output shapes it. Wrong outputs are catchable before they land, if someone looks. Required: Controls 1 to 4, 6, 7, 9. Control 5 in “augment” mode: the human path exists and override rate is measured. Control 8 at the level of version marker plus release checklist. Trap: augment on paper, automate in practice. If override rate is under 2% and nobody can explain why, this is Tier 3 wearing a Tier 2 label. The decision record (Control 4) is what exposes this.
Tier 3: Automate, adverse or material outcome, or irreversible
Automated declination, non-renewal, adverse rating, claim denial, payout amount determination, anything that touches reserves or the financial statements without human sign-off. Test: the system’s output is the outcome, or a human signature is a formality. Required: all nine controls, full strength. Plus: adverse decisions route to human review regardless of confidence (Control 5 asymmetry), rehearsed rollback of the model version specifically, fairness testing documented with its limitations (see the small-N section, forthcoming), and a regulatory-action log per model. Trap: believing you have no Tier 3 systems. If your rating engine is a GLM and it sets the premium, read the next section.
The control matrix
| Control | T0 | T1 | T2 | T3 |
|---|---|---|---|---|
| 1 Register | ● | ● | ● | ● |
| 2 Prompts in VCS | ● | ● | ● | |
| 3 Config as code | ● | ● | ||
| 4 Decision record | ○ output log | ● | ● | |
| 5 HITL authority model | ● augment | ● with adverse-routing | ||
| 6 Validation + observability | ○ sampling | ● | ● + fairness + scheduled re-run | |
| 7 Data provenance | ● | ● | ||
| 8 Release control | ○ version + checklist | ● + rehearsed model rollback | ||
| 9 Vendor register | ● | ● | ● | ● |
● required · ○ reduced form
Rough cost: Tier 0 an hour, Tier 1 two days, Tier 2 one to two weeks, Tier 3 three to four weeks. Most operators have one or two Tier 3 systems and a long tail of Tier 0 and 1; a vendor selling into underwriting or claims may be Tier 3 by its customers’ use even if it never sees a consumer. Spend accordingly.
How to run the tiering exercise
This is the method Nolte uses. It is a process, not a stack; tooling decisions follow the audit, never precede it.
- Enumerate from evidence. Outbound network logs, billing, environment variables, package manifests, and the vendor register. Not memory. Every model-provider hostname is a candidate system. Production AI dependencies that appear in no document the client holds are the norm, not the exception.
- Score each system on the four inputs, in writing, with a rationale. One paragraph per system. This paragraph becomes Exhibit C ref 5 and 6.
- Check the label against the log. Where a decision record exists, override rate and human-action distribution tell you the real autonomy. Where it doesn’t, that absence is the first finding.
- Write the tier, the required controls, and the gap. Baseline and target per control, as a table. Nolte specs every delivery this way: a metric, its value today, its value when done. “0 configuration stores catalogued → 100%.” It makes the gap unarguable and the work billable. The reference for what “done” means is the pending NAIC Supplement and its state pilot.
- Change nothing in production during the audit. State this as a constraint in the spec. It is what makes the audit safe to run on a live book and what makes the findings credible to a carrier.
- Sequence by tier, then by dependency. Tier 3 first. Within a tier, Controls 1 to 3 before 4 to 6 before 7 to 9. Do not replace an operating process you have not yet run yourself; document it, run it, then earn the change.
Your rating engine is probably in scope
Founders hear “AI” and think language models. The Supplement’s definition of AI System covers any machine-based system generating “predictions, recommendations, content, or other output influencing decisions,” and its definition of Predictive Model covers “mining of historic data using algorithms and/or machine learning to identify patterns and predict outcomes.” A generalized linear model fitted on loss history to set rating factors is a predictive model by that definition.
Whether GLMs are explicitly named in the final Supplement was still under discussion in early 2026 (Fenwick, March 2026, lists “the inclusion of generalized linear models” among issues under review). Nolte’s read: the definitional language already captures them, the discussion is about whether to say so out loud, and anyone with a rating engine, whether they built it, rent it, or sell it, should plan as if the answer is yes. Colorado’s regulation already reaches predictive models on external consumer data; New York’s Circular Letter 7 reaches external consumer data and information sources in underwriting and pricing.
What this means practically:
- Your rating engine goes in the register (Control 1) with
model_type: GLMor whatever it is, an implementation date, and a version. - If it sets premium without human review per policy, it is Tier 3 by autonomy and financial impact, even though it is not “AI” in the way your team uses the word.
- Exhibit D applies. Geocoding, geo-demographics, credit-based scores, vehicle data, loss experience: the rating engine’s inputs are precisely the categories Exhibit D lists, and Exhibit D column 5 wants the vendor name for each external one.
- Validation (Control 6) for a GLM is actuarial review plus outcome monitoring, not an LLM eval suite. The artifact is different; the requirement to produce a dated one is the same.
- The rating engine is also the system most likely to be rented. If a vendor runs it, Exhibit B Q3b applies and you need their validation procedure in writing, not their marketing.
To be precise about what is and isn’t new here: rating models already carry actuarial governance. ASOP 56 covers model documentation and validation, and rate filings impose reproducibility. None of that is being questioned. What is new is the register: the Supplement asks for the model by name, version, and vendor in an AI inventory, and most operators’ actuarial documentation lives somewhere the AI program never looks.
The contestable version, for the record: your rating engine already has actuarial governance. It does not have an entry in your AI register, and the examiner will look in the register.