The model is never the decider
Language models read, extract and judge. Code gates and rolls up. A person records the decision. Swap the model tomorrow and your decision logic is unchanged.
Solutions
Collect it. Calculate it. Verify it.
Three engines, one platform. Use one on its own, or chain all three into a process that runs end to end — from the first question you ask a customer to the decision you can defend two years later.
01 — Collect data
insdi Gather
Guided digital journeys that collect information from anyone — customers, brokers, employees, surveyors — and validate it at the point of capture. No fillable PDFs, no email chains, no re-keying.
- Validated at source
- Conditional logic, field-level rules and required evidence — so what arrives is already clean.
- Pre-populated
- Pull what you already know from your policy admin, CRM or member data before the first question is asked.
- Shareable any way
- A link, an email invitation, an embed on your site, or a section handed to a third party mid-journey.
- Save and resume
- Long forms survive interruptions. Multi-party journeys survive handoffs.
- Cover details
- Risk address
- Declarations
- Supporting documents
- Review & submit
pre-filled from policy admin · 14 fields
Where it is used
- Product sales & quoting
- Needs analysis & records of advice
- Risk surveys & inspections
- First notification of loss
- Broker & customer onboarding
- Group scheme enrolment
- Renewal data refresh
- KYC & compliance capture
02 — Rating & rules engines
insdi Calc
Deploy your rating and rules engines to an API using the Excel models you already price on. Same-day deployment and automated testing — no recoding of rules, instant deployment, and no version management to run yourself.
- Your Excel model is the engine
- Upload the workbook you already price on. No one re-writes it in code, so no one introduces a translation error.
- Same-day deployment
- From workbook to a live, authenticated API endpoint the same day — not the next release window.
- Automated regression testing
- Every publish is scored against your expected-result pack before it goes live. A rate that moves unexpectedly fails the build, not the customer.
- No version management burden
- Every call is stamped with the model version that produced it. Repricing a quote from nine months ago is a lookup, not an archaeology project.
Sheet1!B12 = base × zone × ncb
POST /rate/motor
- ✓ 412 test cases matched expected
- ✓ deployed — no downtime
- ✓ v13 retained for replay
Where it is used
- Premium rating
- Underwriting rules & referral triggers
- Eligibility & acceptance criteria
- Commission and fee calculations
- Benefit & sum-insured schedules
- Claims quantum and settlement
- What-if and rate impact testing
- Broker-specific rate tables
03 — AI workflows
insdi Verify
The model reads. The code decides.
A verification engine that puts a language model where it is genuinely better than a person — reading documents and spotting inconsistency — and keeps it out of the place a regulator will ask about: the decision itself.
Every AI pilot in insurance dies at the same question. Not “is it accurate?” — accuracy is measurable and improvable. It dies at: show me exactly why this was declined, eighteen months from now, to somebody who is hostile. Verify is built to answer that.
- Code Intake & correlation Webhook, monitored mailbox or upload. Attachments exploded, duplicates dropped, item matched to a case before a model is consulted.
- Model Extraction Documents become typed fields against the schema for that case type. Every field carries provenance — which document, which page, observed or inferred.
- Code Deterministic gate Anything a formula can settle is settled by a formula, before spending a token. Cover dates, premium standing, notification windows.
- Model Specialist checks, in parallel Cover, identity, consistency, quantum, fraud signals — each emitting structured findings that must cite the text they came from.
- Code Grounding & second opinion Every quoted snippet is mechanically located in the source. A reviewing model can route a case to a human, but can never change an outcome.
- Code Roll-up A declarative decision policy — not a prompt — maps check results to an outcome. You can read it, change it, and diff two versions of it.
- Human Your assessor decides Recommendation and decision are stored as separate fields. Where they disagree, a reason is required — and that disagreement rate is your calibration dashboard.
Deterministic code Model reasoning Your people
What makes it defensible
Missing is not false
Every check returns pass, fail or skipped. If the evidence was never observed, the engine says so and asks for it — it does not infer a pass.
Runs are immutable
Each run freezes its inputs, outputs, rule versions and model versions. New evidence produces run N+1; earlier runs stay readable forever.
Your experts author the checks
An underwriter uploads a checklist in Excel or writes checks in plain English. A validator, not a model’s confidence, is the gate to publishing.
Your disputes become the test suite
Any case a handler overruled can be promoted into a frozen regression example. Every future change is scored against it before it ships.
Your account, your model contract
Anthropic or OpenAI direct, or either on Amazon Bedrock inside your own AWS account. Your data, your region, your compliance posture.
One engine, four lines
Claims, underwriting, survey and sales all ask the same question: do the documents in front of me support the assertion being made? Everything that varies by line of business is configuration, not code.
Claims
Does this claim fall within cover, at a defensible amount, without fraud indicators?
Underwriting
Is this submission complete, consistent, and the risk as declared?
Risk surveys
Was the survey completed properly, and were the requirements met?
Sales files
Is this file compliant, and does the quote match what was actually sold?
Bring us one process. We will show you which engines it needs — and how fast it can be live.