Insurance LLM work is a specialist task, not a claims engine.
Insurance LLM work is a specialist task, not a claims engine. Fine-tuning is worth discussing when you have a repeatable document or correspondence job, a specialist who can judge the output, and a pack of unused cases the model will not train on. It is the wrong first step if you need a system that decides coverage or settlement on its own. The model may extract, classify, or draft. Authority stays with the people and systems you already named. This page is the review boundary and the evaluation you should demand before anyone books a run — not a promise that adaptation replaces a claims platform or inherits a handler’s pen.
By Shen Pandi, Towards AGI. Reviewed September 2026.
Who this page is for
Read this page when the brief already sits in an insurance operations, binder, or claims-support desk and someone has asked whether a language model can take a slice of the paper. Typical readers are a claims operations lead who wants faster extracts, counsel who heard “the model will decide,” and a specialist who already signs the workbench row. The page is not a tour of foundation families and it is not a claims-system replacement guide.
If you do not yet know whether training is the engagement at all, start at the fine-tuning article. If the files cannot be used yet, start at readiness. If the real constraint is where a first-notification snippet may sit, start at inference and the sovereignty checklist. Stay here when the live question is the boundary between document work and an engine that would accept, decline, reserve, or settle.
Document work we will discuss
The useful jobs are narrow and already exist on a desk. Extract agreed fields from a wording family and keep a pointer to the clause. Classify a submission into a queue a binder already runs. Draft a letter a handler still reviews and signs. Those jobs have a current process, a volume, and a definition of a good extract or a good draft. Without that definition you are asking the model to invent the desk.
A general-purpose model can draft. A specialist has to be measured on your document families, languages, and exception cases — not on a public invoice suite and not on a vendor PDF that beat a base model in someone else’s lab. Your unused pack is the evidence. A leaderboard is not. We will not accept a public suite as a stand-in for manuscript endorsements you actually handle.
We work with teams that can name the desk. If the sponsor cannot say who signs the workbench row today, there is no specialist task yet. There is a wish for automation. We will not train toward a wish. Naming the desk is not bureaucracy. It is how we know who will refuse a fluent guess on a Friday afternoon.
Policy and contract extraction with source references.
Submission classification and triage into an existing queue.
Claims correspondence assistance under an agreed review path.
The review boundary belongs in the brief
Write what the model may suggest and what it must never conclude. Illustrative workflows support specialist review. If the draft says a limit is £X because the schedule said so, a human still accepts the row. If the draft would say “this claim is covered,” that sentence is out of scope even when the wording looks obvious. Obvious to a fluent model is not authority.
Conduct and audit needs are part of readiness, not a later sticker. If procurement needs a trail from extract back to clause, that trail is a requirement on day one. If a regulator-facing file cannot contain a model-only conclusion, the output schema must make that impossible. We will not add a disclaimer footer and call it governance. A footer the handler can ignore is not a boundary.
Silence is a first-class output. Missing schedule, split clause, handwritten endorsement the scanner ate: the model should refuse and route. A fluent guess is the failure. Train restraint as hard as you train fields. If the brief has no silence route, you do not yet have a specialist task. You have a hope that the model will always find something.
Decision prompts, not scores.
When
Then
The output is a field, a queue, or a draft a named human still signs.
A specialist adaptation conversation can start after permissions and an unused comparison pack exist.
The output would accept, decline, reserve, or settle without a human conclusion.
Refuse that scope. It is a claims engine. We will not take it as a fine-tuning job.
Two specialists mark the same clause differently and nobody will own the label.
Stop. Write the task. Training copies the disagreement into production language.
The pain is last month’s circular, not the extract shape.
Retrieval or a process that opens the current file. Do not bury news in weights.
What to evaluate before anyone trains
Agree extraction accuracy in the sense your desk already uses — field match, source pointer, exception rate — and the point at which a human must take the case. Latency and operating cost matter when the workload is high. Authority for coverage or claims decisions stays where you already put it. The model does not inherit it by being fluent, and it does not inherit it because someone fine-tuned last quarter’s letters.
Vendor PDFs that show a fine-tuned mini-model beating a base model on invoices are not evidence for your book. Public suites do not contain your manuscript endorsements. Your unused pack does. If you cannot spare that pack, you are not ready to train. That is a readiness finding, not stubbornness. We will not invent a substitute pack from a demo corpus and call it your book.
Languages and document families are coverage questions. A model that is tidy on English schedule A and silent on a scanned Spanish endorsement is not “mostly done.” Write the families that must pass before anyone calls the pilot a success. Write the families that are out of scope so a later slide cannot pretend they were included. Evaluation that forgets the out-of-scope list is how a narrow specialist gets sold as a book-wide engine.
Do this writing before a training cluster is booked. Evaluation criteria that appear after a notebook looked good are a retrofit. We will not help you retrofit a win. We will help you write the bar while it can still fail.
Insurance example: extract with a pointer, not a verdict
A wording team already lifts period, named limits, and insured entities into a workbench. The senior checks exceptions. Volume is the pain; inventing coverage is not the brief. That is in-scope document work. Teach the fields and the refusal. Keep unused wordings for the final comparison so the go / no-go is not a memory test.
The moment the sponsor adds “and tell us if we should bind,” the scope has left this page. Binding is authority. We will split the brief or stop. A split that keeps the extract and drops the verdict is still a useful engagement. A split that keeps the verdict and drops the extract is not an engagement we will take.
Schedule extract under a human sign-off
insurance
Name the fields, the source-pointer rule, and the exception queue. State that a blank beats a guessed limit. State that no output row may contain a coverage verb. Those sentences go in the proposal before a single file is transferred. If a product owner later wants a “recommended bind” column, that column is a new brief and a new refusal until a human owns the recommendation as their own.
If evaluation would require live claim files you cannot permit, use a permitted historical pack or stop. “We will anonymise later” is not a permission. See the readiness page for that gate. This page only needs you to know that a live file you cannot use is not a reason to skip evaluation. It is a reason to wait.
Retail contrast: the same boundary, a different desk
Retail is not this page’s industry, but the same boundary shows up when a shop wants a model to decide a refund. We will not take that as an engine either. Attribute fill and returns routing under a human policy are the retail analogue of document work. Price and live stock stay in systems. The insurance lesson is the same sentence: specialist task, named review, no inherited authority.
If you came here from a retail brief, read this boundary, then use the fine-tuning and SLM pages for catalogue jobs. Do not copy an insurance extract scope onto a refund decision and call it the same engagement. The industries share a refusal. They do not share a schema.
Why a refund engine is the same refusal
retail
A returns desk already codes reasons; a supervisor already applies the refund policy. Teaching the codes is a candidate job. Teaching the model to issue the refund is an engine. We refuse the engine in retail for the same reason we refuse it in insurance: authority is not a fine-tune. A fluent “yes, refund” is still a decision the firm has not assigned to weights.
The useful crossover is editorial. Both desks need a silence route and an unused pack. Neither desk is a reason to publish a score we have not measured. If a conference talk needs a number, commission an evaluation. Do not borrow an invoice-demo percentage and change the nouns to claims or returns.
Correspondence is still document work, until it is not
A letter that restates facts the file already holds, in a house style a handler already edits, is document work. A letter that would tell the customer the firm’s position on liability is not. The difference is not tone. The difference is whether the output can be read as the firm speaking a decision. If it can, the model is the wrong author for that sentence.
We will discuss tone and restraint for in-scope drafts. We will not discuss a generator that “handles complaints end to end.” End to end is how an engine sneaks in through the correspondence door. The review path has to name the human who sends, not only the human who glances.
Data and inference still sit under the claim file
Policy PDFs and first-notification snippets are not generic text. Who owns them, who may use them for training, and where a prompt may travel are readiness and inference questions. A tidy extract on a shared endpoint can still be a file the counsel team has not accepted. We will not treat those as afterthoughts because the task is “just documents.” Just documents is how claim files leave the building.
The sovereignty checklist is the artefact when residency or operator access is the real constraint. This page will not reprint that list. It will say: if the claim file cannot leave a named region, write that before you pick a base model. Training will not move the endpoint. A specialist that learned your letter style on a forbidden hop is still a forbidden hop.
In scope versus out of scope. Not a menu of products.
Ask
This practice
Extract named fields with a clause pointer
In scope after readiness
Route a submission into an existing queue
In scope after readiness
Draft a letter a handler signs
In scope with a written review path
Decide coverage or settlement
Out of scope
Replace the claims platform
Out of scope
Publish a customer accuracy number from a demo
Out of scope
How an insurance engagement actually starts
Bring one document family or one correspondence type, the current desk, and the constraint that matters — review, residency, or volume. We will say whether the first paid week is readiness, a focused pilot, or a stop. Model artefacts and any later support are written before work begins. We do not start a training cluster to look busy. Looking busy is how a claims engine gets a purchase order.
Other industries can use the same shape. The insurance-specific rule is the engine refusal. If your book is specialty, marine, or personal lines, the rule does not change. The fields change. The authority does not move to the weights. A marine schedule extract is still an extract. A marine coverage verdict is still an engine.
If procurement needs a longer envelope, we will write a later option that can be cancelled when the unused pack fails the bar. We will not write a year of training so that a platform replacement can hide inside a document-work statement of work. That hide is how this page would stop being citable.
Questions we hear first
How do you measure success?
We agree task-specific criteria before training, then compare the adapted model with a baseline on held-out examples. Quality, latency and running cost can all form part of the evaluation.
Is test data used for training?
No. The held-out test set is reserved for final evaluation, separate from training data and validation data used during development. If test results are used to guide further changes, a fresh independent evaluation is needed before claiming final performance.
Does fine-tuning make our AI sovereign or secure?
Fine-tuning alone does not establish sovereignty or security. Those depend on where data is processed, who operates the infrastructure, access controls, licensing and operational practices. We scope these requirements alongside model adaptation.
Can the model pre-fill a reserve recommendation?
Only if a named human still sets the reserve and the output cannot be read as the firm’s decision. If the brief wants the model to set the number, that is engine territory. We will not take it.
Do you work only with carriers?
We work with teams that own the desk — carrier, MGA, broker operations — when the job is document work under a review path. The entity type matters less than whether someone can judge a good extract.
Will you train on live claims this quarter?
Only if permissions, minimisation, and the unused pack are already true. Live files are often the wrong first corpus. A permitted historical family is usually the honest start. The readiness review is where we say which.
Is submission triage the same as a claims decision?
No. Triage into a queue a binder already runs is classification under review. A claims decision is authority. We will discuss the first after readiness. We will not relabel the second as triage to make a training week possible.
Can we start with a vendor invoice demo as the insurance baseline?
No. An invoice suite is someone else’s paper. Your unused wordings are the baseline conversation. A demo can show that adaptation exists as a technique. It cannot show that your book is ready.
Start with one task and a written scope.
Bring the workflow, the constraints and where inference should run. We use that brief to decide whether LLM fine-tuning, an SLM, retrieval or a readiness review is the right first step.