The Supplement classifies AI by autonomy: support, augment, automate. Architecture classifies by certainty: deterministic, probabilistic, human. Governance is where the two meet. A probabilistic component inside a deterministic control structure is governable. A probabilistic component that is the control structure is a lawsuit with a timestamp.
The rule
Three territories, one design pattern.
Deterministic: same inputs, same output, every time, provably. This is about reproducibility, not about whether machine learning is involved. A trained model with fixed weights and a pinned version is deterministic at inference; a language model sampled at non-zero temperature is not. A GLM or gradient-boosted rating model that is filed, versioned, and locked belongs here. v5.0 names GLMs as in scope, so a filed rating GLM is still an AI Model you may have to list, but at runtime it is deterministic: same inputs, same premium. Anything that has to be filed, reconciled, reproduced for an examiner, or that moves money. Rating, binding, issuance, pro-rata math, billing, reserves, payouts, bordereaux. A rating algorithm that gives different premiums for identical inputs isn’t innovative. It’s unfileable.
Probabilistic: pattern work where 95% right at a tenth of the cost is a good trade because the failure mode is a correctable inefficiency. Extraction from unstructured documents, classification, triage, drafting, summarisation, anomaly flagging, similarity search.
Human: accountability joints. Adverse decisions, exceptions, anything irreversible, anything where the examiner will ask “who decided.”
The pattern that survives every regulatory regime: the model proposes, the rule disposes, the human owns the exceptions. The probabilistic component emits a score or a structured suggestion. A deterministic rule consumes it and routes. A human handles what the rule sends them. The score, the rule version, and the human action are all logged (Control 4).
The most common failure isn’t a bad model. It’s a model whose output nothing deterministic receives. That is Nolte’s architectural reading of what has been alleged in the nH Predict litigation, which is ongoing: not that a model was wrong, but that, as alleged, no mechanism existed to catch it being wrong.
The second most common failure is spending probability where the business needed certainty. A six-figure monthly inference bill on quote-bind-issue is what that looks like in production.
The map
For each workflow: the steps, the placement, the tier from Part 3, the Exhibit A operational area, and the failure mode if placement is wrong.
Submission and intake
| Step | Placement | Why |
|---|---|---|
| Receive submission (email, portal, ACORD form, broker PDF) | Deterministic ingest | Idempotent, logged, no interpretation |
| Extract fields from unstructured docs | Probabilistic | Pattern work; wrong field is caught downstream |
| Validate extracted fields against schema and business rules | Deterministic | Required fields, ranges, state eligibility |
| Human confirmation of low-confidence extractions | Human | Confidence threshold set in config (Control 3) |
| Appetite check and routing | Deterministic rule on the extracted, validated data | The rule can consume a model score, but the rule decides |
Tier: 1 (extraction) rising to 2 if the appetite routing consumes a model score that could decline a risk. Exhibit A: Underwriting/Eligibility, Marketing. Failure if wrong: an LLM “deciding” a risk is out of appetite. That’s a declination with no rule to point at.
Quoting and rating
| Step | Placement | Why |
|---|---|---|
| Rating engine (factors, relativities, tiering, GLM outputs) | Deterministic, versioned, locked | Filed. Must reproduce. Same inputs, same premium, forever. |
| Quote assembly, forms selection, disclosures | Deterministic | Form applicability is rule-based per state |
| Quote explanation or summary for the customer or broker | Probabilistic, on top of deterministic outputs | Drafting from structured data; wrong wording is embarrassing, not unfileable |
| Referral decision (auto-quote vs refer to underwriter) | Deterministic rule, may consume a model score | The score is an input; the rule is the decision |
| Underwriter judgement on referrals | Human | Accountability joint |
Tier: 3 for the rating engine (automate, material financial impact, consumer impact). 1 for the summary. 2 for a model-scored referral rule. Exhibit A: Premium Quotes and Discounts, Ratemaking/Rate Classification. Failure if wrong: any probabilistic component between the validated inputs and the premium. It doesn’t matter how accurate it is; it cannot be filed.
Binding and issuance
| Step | Placement | Why |
|---|---|---|
| Bind eligibility checks (payment, signatures, KYC, state rules) | Deterministic | Compliance gate |
| Policy number, effective dates, document generation | Deterministic | Reproducible artifacts |
| Carrier notification, bordereaux row creation at bind | Deterministic | This is the data-model decision: structure it now, not at reporting time |
| Fraud or identity anomaly flag at bind | Probabilistic score → deterministic hold rule → human review | Model proposes, rule disposes |
Tier: 3 for bind itself. 2 for the fraud score if it can hold a bind. Exhibit A: Underwriting/Eligibility, Fraud/Waste & Abuse. Failure if wrong: a probabilistic “should we bind” step. Binding is a contract formation event. It’s deterministic or it’s not binding.
Endorsements and mid-term changes
| Step | Placement | Why |
|---|---|---|
| Interpret the change request (email, call note, portal free text) | Probabilistic extraction to a structured change object | Pattern work; validated next |
| Validate the change against policy, product, and state rules | Deterministic | Which endorsements are allowed, when, with what forms |
| Re-rate and pro-rata premium calculation | Deterministic | Money math. Never probabilistic. |
| Adverse changes (coverage reduction, cancellation, non-renewal notice) | Deterministic rule → human approval | Adverse consumer outcome; human owns it |
| Endorsement document generation | Deterministic | Reproducible |
Tier: 1 for interpretation. 3 for re-rate. 3 for adverse changes. Exhibit A: Underwriting/Eligibility, Customer Service. Failure if wrong: the extraction silently changing a limit and the deterministic layer trusting it. Validation must be against the policy of record, not the extracted object.
Certificates of insurance
| Step | Placement | Why |
|---|---|---|
| Intake of COI request (broker email, portal, holder form) | Probabilistic extraction of holder, requirements, dates | Pattern work |
| Match requested coverages against policy of record | Deterministic | The certificate must reflect the policy exactly |
| Flag requests that exceed the policy (additional insured not on policy, limits above bound) | Deterministic rule → human or broker | Exceptions are where errors and E&O exposure live |
| Generate certificate | Deterministic templating from policy data | Reproducible, auditable |
| Renewal-driven reissue | Deterministic, scheduled |
Tier: 1 for intake. 2 for the whole flow because a wrong certificate is a misrepresentation to a third party. Exhibit A: Customer Service. Failure if wrong: an LLM drafting the certificate. The certificate is a statement of fact about a contract; it comes from the contract, not from a model’s reading of an email.
Claims
| Step | Placement | Why |
|---|---|---|
| FNOL intake (call, form, photos, documents) | Probabilistic extraction and classification | Pattern work; the biggest legitimate win in the stack |
| Coverage determination (is this loss covered under this policy, on this date, for this peril) | Deterministic against policy of record | Never a model. This is the contract. |
| Triage and routing (severity, complexity, fraud signal, assignment) | Probabilistic score → deterministic routing rule | The rule can be tuned; the score is logged |
| Reserve setting | Deterministic or actuarial method, versioned; model may propose | Financial statement impact (Exhibit C ref 13) |
| Adjudication: approve within authority | Deterministic rule on validated facts, possibly model-assisted | Automate only below a written authority limit |
| Adjudication: deny, partial, dispute | Human, model may summarise | Adverse consumer outcome; unfair claims settlement exposure |
| Payout calculation and payment | Deterministic | Money |
| Subrogation and salvage identification | Probabilistic flag → human | Recovery opportunity; wrong flag costs little |
| Customer communication drafting | Probabilistic draft → deterministic compliance check (required language, timelines) → human send for adverse | Timelines are statutory per state |
Tier: 3 for coverage determination, adjudication, reserves, payout. 2 for triage. 1 for FNOL extraction and comms drafting. Exhibit A: Claims/Adjudication. Failure if wrong: a model producing the coverage determination or the denial. This is exactly where the industry has already been sued, and exactly where the Supplement’s Exhibit C ref 14 points (compliance review, including unfair claims settlement laws).
Renewals, cancellations, non-renewals
| Step | Placement | Why |
|---|---|---|
| Renewal eligibility and re-rate | Deterministic | Filed |
| Retention risk scoring | Probabilistic | Marketing-adjacent; wrong score costs a call |
| Non-renewal or conditional renewal decision | Deterministic rule → human | Adverse; state notice periods are statutory |
| Notice generation and timing | Deterministic | Statutory timelines |
Tier: 3 for non-renewal. 1 for retention scoring. Exhibit A: Underwriting, Customer Service, Marketing.
Bordereaux and carrier reporting
| Step | Placement | Why |
|---|---|---|
| Row creation | Deterministic, at bind and at every transaction | Structure at the source |
| Reconciliation against premium and claims ledgers | Deterministic | Must balance |
| Anomaly detection across the file before submission | Probabilistic flag → human review | Catches what reconciliation doesn’t |
| Submission to carrier | Deterministic, versioned, receipted | Evidence of what was sent |
Tier: 3 (financial reporting impact). Exhibit A: Reinsurance, Reserves/Valuations. Failure if wrong: generating bordereaux from unstructured data with a model. Bordereaux is a data-model decision made at bind time, not a reporting feature bolted on after.
Billing and payments
Entirely deterministic except: dunning message drafting (probabilistic, compliance-checked) and payment fraud scoring (probabilistic → rule → human). Tier: 3 for money movement. Exhibit A: Other.
Customer service and producer support
| Step | Placement | Why |
|---|---|---|
| Answering policy questions from documents | Probabilistic, grounded in policy of record, with citations | Wrong answer costs a follow-up |
| Anything that sounds like a coverage confirmation | Deterministic lookup, or human | ”You’re covered” is a decision |
| Complaint intake and tagging | Probabilistic classification → deterministic tag to the AI system involved (checklist 3m) |
Tier: 1, rising to 2 the moment the bot confirms coverage. Exhibit A: Customer Service.
Reading the map
Count the cells. Probabilistic placements cluster at three points: intake (extraction, classification), scoring (triage, fraud, retention), and drafting (summaries, communications). Everything between those points and the money is deterministic. Everything adverse has a human at the end.
That is not a limitation on AI. It’s where the wins are. FNOL extraction, submission intake, and endorsement interpretation are where operators lose the most time, and they’re all Tier 1. The expensive part of the stack was never the part that needed a model.
Three tests for any proposed placement:
- Does the examiner need to reproduce this output? Deterministic.
- Does this output move money or form a contract? Deterministic.
- Is this output adverse to a consumer? Human at the end, deterministic rule before. If none apply, probabilistic is allowed, provided something deterministic receives the output and Control 4 logs it.
Choosing the model
Cost and efficacy start the decision. In insurance, the tier, the data, and the evidence finish it.
The map tells you where a probabilistic component is allowed. It does not tell you which model to put there. That choice starts on two axes, and both move too fast to print: how well a model does the specific task in front of it, and what it costs per token at your real volume. We track both on BenchmarkList, a public index with a queryable data API that ranks capability and lists pricing, open-weight and frontier side by side. Read efficacy against the task you actually have (extraction, classification, triage, drafting), not a general leaderboard. A model that tops a reasoning benchmark can still be the wrong choice for pulling six fields off an ACORD form.
Then the part a leaderboard cannot score for you: whether the model fits a component an examiner will ask about. Four questions decide that.
Can the data leave? A frontier model is an API call to someone else’s servers. If the input is claims content, medical detail, or anything you would not export, an open-weight model you host inside your own boundary keeps the data where Exhibit D says it is. Provenance is not a preference here. It is a column on the exhibit.
Can you reproduce it? A reproducible output gives the same answer to the same inputs, and you have to name a last test date (Exhibit C ref 12). Frontier versions change under you, sometimes silently. An open-weight model with pinned weights and a recorded version is frozen the way the register in Control 1 and the config discipline in Control 3 expect. If the placement is deterministic, this is not a choice at all. If it is probabilistic but scored, a moving model means a moving score.
Who owns the validation? With a frontier model you inherit a vendor’s attestation and supplement it with your own testing (Q3b asks exactly this). With an open-weight model you own the eval end to end. Neither is free. Pick the one you can actually evidence.
What tier is it? Part 3, Tiering sets the eval and logging bar, and it rises with consumer impact, automation, and money. A higher tier does not ban a frontier model, but it narrows where a black box you cannot reproduce is acceptable, and it raises what you log around whatever you pick (Control 4).
| Dimension | Frontier | Open-weight |
|---|---|---|
| Efficacy ceiling | Highest on hard reasoning and drafting | Closing fast; often sufficient for bounded tasks |
| Cost at scale | Per call, grows with volume | Fixed hosting; cheaper at high volume |
| Data control and residency | Leaves your boundary | Stays in your VPC |
| Reproducibility | Versions move under you | Pin and freeze the weights |
| Validation ownership | Vendor attestation, you supplement | You own the eval |
| Exam fit | Strong where quality is the point and the data can leave | Strong where reproducibility, residency, or cost decide |
The map already sorts the work into where this choice happens. High-volume intake extraction (FNOL, submissions, endorsement requests) usually leans open-weight: volume makes cost dominate, the data is sensitive and high-count, and the output is caught by a deterministic validator, so “good enough, corrected downstream” clears the bar. Scoring (triage, fraud, retention) leans toward whatever is cheapest with honest calibration, because the score feeds a deterministic rule and is logged, not trusted raw. Drafting (summaries, customer communications) is the one place a frontier model can earn its price on quality, and it is also where the data and eval bar is highest, so it sits behind a deterministic compliance check and a human send for anything adverse.
Current picks
Today’s picks by task, ranked on BenchmarkList by capability against price. Models move fast, so treat this as a snapshot: re-run the query as the field changes, and drill into the benchmark that matches your task rather than a single composite score.
| Task, and where it sits in the map | Frontier | Open-weight | Why |
|---|---|---|---|
| Field extraction: submission, FNOL, endorsement intake | Gemini 3.8 Flash | GLM 5.3 Flash, or Qwen3.8-Flash-Next | High volume, sensitive data, and a deterministic validator catches errors. Cost and residency decide, so open-weight usually wins. |
| Classification and scoring: triage, fraud, retention | Gemini 3.8 Flash | DeepSeek V4 Flash, or Qwen3.8-Flash-Next | The score feeds a deterministic rule and is logged. Calibration and cost matter more than a raw ceiling. |
| Drafting and summarization: quote and claim summaries, customer communications | GPT-5.6 Terra, or Gemini 3.8 Flash | Qwen3.8-2.4T-A95B, or Kimi K3 | Quality earns its price on customer-facing text. Keep it behind the compliance check and a human send for anything adverse. |
| Reasoning-assist: coverage-adjacent analysis, adjudication summaries, exceptions | Claude Opus 5, or GPT-6 Astra for the ceiling | Kimi K3 | The hardest reasoning, where the model only proposes. A human owns the decision; the model is never the decision. |
Snapshot pulled from BenchmarkList on 2026-09-08.
The cheapest, most capable model that you cannot reproduce, cannot keep the data inside, or cannot evidence is the wrong model for a component the examiner will ask about. BenchmarkList tells you cost and efficacy. The map and the tier tell you the rest.
Orientation, not endorsement. Nolte does not endorse any model or vendor; current capability and price live on BenchmarkList, which moves faster than this page.
What we build this way to protect
In Nolte’s own delivery experience across a full MGA stack in all fifty states, the hard part was never the code. It was drawing this line and holding it under pressure from every vendor demo and every “just let the model handle it” suggestion. The rating engine was deterministic, versioned, locked. Quote-bind-issue was deterministic. Bordereaux was deterministic. The probabilistic tools lived at the intake edge where being wrong was cheap.
That line is what the Supplement now asks you to draw in writing. Exhibit A’s AI Models by type under Direct Consumer Impact and Material Financial Impact, Exhibit C ref 7 (risk classification), and Exhibit B Q1g’s autonomy-and-reversibility question are all the same question: where did you allow the guessing, and what catches it.