Briefing
Twenty minutes. First, enough of the technology that we all use the same words. Then the numbers for an insurer of this size.
Part one — the technology
Given a sequence of text, it predicts what comes next. Everything else is built on top of that.
The loop runs once per token. What people call "reasoning" is the model spending more tokens thinking before it answers.
Part one — the technology
A chunk of text, roughly four characters. Models read, price and limit work in tokens, not words.
How much text it can hold at once. Everything it knows about your task lives there, and is gone on the next call.
The trained parameters. Open-weight models can be downloaded and run on your own hardware.
How adventurously it picks. Low for data work and code. Higher for drafting and ideas.
Part one — the technology
The right-hand column is the specification for what an agent has to solve.
Part one — the technology
A chat window gives the model words.
An agent gives it hands, memory, and a job.
Part one — the technology
You send a request. Hermes gives the model tools, reads the real results back, and repeats until the work is actually done.
It never guesses what a script returned. It runs the script and reads the result.
Part one — the technology
01
Files, spreadsheets, databases, scripts, web, email. How the model touches anything real.
02
Durable notes loaded every session, so what you explained once does not need explaining again.
03
Your procedures on disk, loaded only when the task matches. Your SOPs, not the model's defaults.
04
Scheduled and background work, so claim triage, document checks and fraud flags run overnight and are waiting in the morning.
Part one — the technology
A skill is a written procedure with a trigger. Hermes loads it only when the work calls for it.
Your value is twenty years of knowing how an underwriter wants a claim handled. A skill is that knowledge written down once, applied consistently, and no longer trapped in one person's head or lost when they are on leave.
Part one — the technology
Protea is a privacy and memory layer, used as the example of the right architecture for regulated work. It sits between a client and any model backend, so you get the frontier model and keep the data.
Named entities — people, policyholders, claims, medical terms — are replaced with stable placeholders before the request leaves your hardware, and restored in the reply. The model never sees a claimant’s name, ID number or medical reports, so you can run frontier models without an exception.
A single SQLite file per user behind an OpenAI-compatible API. An embedded local model means it runs fully offline once the model weights are pulled. No third party ever holds your corpus.
Structured recall across conversations and months — spiral-indexed by relevance and recency — rather than starting from nothing every session. Decades of knowing how this book is written stops living in one person's head.
Conductor routing passes each task to the model best suited for it; a confidence cascade escalates only when needed: local model for the routine, frontier model for the hard call. Consensus fans a delicate question to several models and synthesises the answer. Teams decompose a brief, rate deliverables, and review every result. An Open WebUI front-end gives clean per-user separation, with identity forwarded through the API.
The pitch to a policyholder: "we run AI-assisted claims handling and your personal data never leaves our environment." Very few small vendors can say that sentence truthfully.
docs/README.md; redaction pipeline docs/redaction-algorithm-research.md (local: D:\Projects\Protea\docs)Part one — the technology
Asked for something it cannot know, a model produces a fluent, plausible, wrong answer rather than nothing.
An invented variable name. A confident number with no source. A mapping that looks right and is not.
Read before answering, run before claiming, show the source. Grounding in real output — and a person signing off. Never autonomy over a regulated deliverable.
Part one — the technology
The evidence is not "experts get faster" — it is that expertise is what lets AI be used on the hard stuff at all.
The model cannot produce the question that only years of knowing claims and policy wording produces. The value is upstream of the answer.
AI is fluent and confidently wrong. Inside its capability boundary it gains ~40%; pushed past it, performance drops 19 points. Only experience says where that boundary is.
The model suggests; the domain expert decides. A 19-point penalty exists precisely because it is not obvious to a generalist which tasks AI can do and which it cannot.
A senior hand encodes their own judgement as a skill. Their expertise stops being a bottleneck and becomes repeatable.
AI compresses the skill gap on routine tasks — a beginner with AI reaches near-expert output on well-defined work. But an insurer does not win on routine. It turns on the messy, consequential, one-off judgement where a confident wrong answer costs a repudiation.
The real gain for an expert is not doing the same task faster. It is having a tireless assistant who drafts and cross-checks, while the expert spends the saved hours on the decisions, the policy wording, and the relationships no model can hold.
Part two — the numbers
The question is not whether AI is impressive.
It is how much of every claim is spent handling it, and how much leaks out the side.
Part two — the numbers
Short-term insurance claim spend in South Africa is enormous, and the industry's own crime bureau is blunt about how much of it is not honest.
R35bn / yr
Estimated annual short-term insurance claim spend in South Africa — the Insurance Crime Bureau's own figure.
R3.5bn
Lost every year. Much of it is not opportunistic — the ICB says a significant share is organised, running across multiple insurers at once.
That is the shape of the problem: a huge, repetitive, document-heavy process — with a determined adversary inside it.
Part two — the numbers
65.2%
Claims paid as a share of premiums (2023). Every point of handling cost or leakage comes straight off the margin.
48% / 43%
Santam's own data: insurance crime splits motor (48%) and non-motor (43%). The document-heavy half is the vulnerable half.
R1bn
In fraudulent claims identified for investigation over three years — plus 2,031 claims worth R65m rejected outright. Detection pays.
You do not need all of it. You need the insurer that handles claims faster and leaks less — and can show the board the number.
Part two — the numbers
Annual cost of claims handling plus fraud leakage, ZAR. Modelled on a mid-size non-life book at ~R1.2bn gross written premium.
Part two — the numbers
Underwriting result, ZAR millions. Waiting is not neutral: leakage and handling cost keep compounding whether or not you act.
Part two — the numbers
| Path | Handling + leakage | Claims touched by hand | Underwriting result | Combined ratio | Cumulative result 2026–31 |
|---|---|---|---|---|---|
| Start now (2026) | R65m | 40% | +R53m | 95% | R212m |
| Wait until 2029 | R80m | 60% | +R30m | 98% | R109m |
| Do not act | R134m | 85% | −R24m | 106% | −R54m |
R103m
Cumulative underwriting result foregone by starting in 2029 instead of 2026. The capability is reachable later. That money is not.
≈R1.2m
First-year programme: platform, integration to the claims system, and the time to write the procedures down properly. A rounding error against the leakage.
Part two — the numbers
Every scenario keeps the team growing. The gain is the volume one assessor or administrator can carry, not a smaller department.
The mechanical half. Keying the same fields, reading the same documents, chasing the same outstanding information.
The decision to pay, decline or investigate. Underwriting judgement. Who signs off on a repudiation.
The claims and policies the same team can handle — more volume, faster settlement, and a reason to pay them more.
An insurer does not win by employing fewer assessors. It wins by settling good claims faster and catching the bad ones sooner.
Part three — where to point it
Read the FNOL, classify the claim, set the reserve, route to the right handler. Hours off the front of every claim.
Police reports, invoices, assessor notes, medical reports — pulled into structured fields, with a person confirming.
Flag the claim against known syndicate patterns and prior claims across the book. Investigate the flagged, pay the rest.
First-pass risk summaries and wording comparison against the schedule — an underwriter's draft, not their decision.
Compare cover, find the exclusion that applies, draft the client-facing explanation in plain language.
Status updates, renewal notices, decline letters — consistent tone, human sign-off, nothing sent unseen.
Assemble the file and the timeline for a complaint — the regulator-facing paperwork nobody has spare hours for.
Conduct reporting and the audit trail, drafted from the system of record and checked by the compliance officer.
Part three — where to point it
The interesting part is not handling the same claims cheaper. It is what becomes possible once the marginal claim is cheap to review.
Automated first-pass checks on 100% of claims, instead of a manual audit of five percent.
Score at notification, not at the end of the chain — before the money is out the door.
Low-value, clean claims settled in minutes. The handler only sees the exceptions.
New niche products priced and administered cheaply enough to be worth writing at all.
Give brokers the same triage and document tools — stickier relationships, cleaner submissions.
Claimant names, IDs, medical details. Local models mean the benefit without the data leaving the insurer.
Part three — where to point it
Pick one claim type — motor windscreen or household — and time the handling honestly for two weeks first.
Write that procedure down properly as a skill. This is the real work, and it pays off even if the AI part fails.
Run it against live claims beside the handler. Compare decisions. Keep the human sign-off on every payment permanently.
Measure handling hours and leakage caught. If they are real, widen to the next claim type. That is the whole thesis.
Mid-size non-life book, ~R1.2bn gross written premium, claims ratio 65.2%. Handling + leakage modelled at a share of claims consistent with the ICB's 10%-fraud assumption and insurance KPI benchmarks. Repetitive handling absorbed over ~4 years.
Combined ratio improves as handling cost and leakage fall; in the "do not act" path it drifts above 100%, where underwriting turns into a loss.
Sources: short-term claim spend R35bn and 10%-fraud → R3.5bn (Insurance Crime Bureau, SAICB); crime split motor 48% / non-motor 43% (Santam); claims ratio 65.2% and FSCA fines (SA insurance industry statistics, 2023 — secondary aggregator); RAF R1bn fraudulent claims identified and 2,031 claims / R65m rejected (Daily Maverick, Nov 2024); SA market structure (SARB Prudential Authority; SAIA). Premium and ratio figures are 2023/24 vintages.
Close
The technology is ordinary. It predicts text, and a framework gives it tools and memory.
The opportunity is not. Billions spent handling claims, in an industry that writes off every wasted rand.
Nothing here requires betting the company. It requires writing down how you handle claims, and starting.