AI systems for regulated work
Agentware builds retrieval, proposal, and forecasting systems for teams that have to defend the output — to a client, an auditor, or a board. Each one shows the evidence it used.
Customer bought in another region and wants warranty work done here. Are we covered?
Yes, through the inter-region route. Labour is reimbursed at your region's published rate[1], not the selling region's. Raise it as a transferred claim carrying the originating dealer code[2] — and note pre-authorisation is required above the standard threshold[3].
Inter-region service is reimbursed at the servicing region's published labour rate, not the rate in force where the vehicle was sold.
Transferred claims must carry the originating dealer code in field 04. Claims submitted without it are rejected at intake.
Pre-authorisation is required for inter-region claims above the standard threshold. Retain the authorisation reference with the job card.
Worked examples from live engagements. Documents and figures are illustrative.
What we build
Each takes a corpus nobody can read end to end and returns something a person can act on — attached to the evidence that produced it. The habit is the same across all three: never assert without a citation.
DocumentsGrounded answers
Policies, handbooks, and bulletins answered in sentences instead of ten links — and only across the documents that person is cleared to see.
An RFPA drafted SoW
An incoming RFP becomes a statement of work, drafted against your template and everything you have already won with. Your team reviews before anything is sent.
Demand + stockWhat to build
Demand projected at model and variant grain, with the drivers that actually move it — then netted against stock, so the number is what to build rather than what the market might absorb.
Beyond the three
Those are what we get asked for most. The habit underneath them — ground the output, cite the evidence, measure it before it ships — is the part that transfers, and it transfers to most of what has landed in the last two years.
Our deepest sector experience is automotive: years of work with OEMs and tier-one suppliers on warranty text, diagnostic logs, and vehicle telemetry. None of the above is specific to it.
How we work
Most failed AI projects were never evaluated — they were demoed. We invert it: build small on your real data, grade it against a test set your experts wrote, then ship.
We start on your documents, not a demo corpus. By the end of the second week there is something you can put questions to and judge for yourself.
Out: a working prototype and a candid read on feasibility
We build a test set with your subject-matter experts and score against it. You see accuracy as a number you helped define, not a claim we make.
Out: an eval harness you keep and can re-run
Your cloud, your identity provider, your retention rules. Where data residency or model choice is constrained, that constraint shapes the design from the start.
Out: a deployment your security team has actually reviewed
Runbooks, retraining schedules, and a walkthrough with the people who will own it. The goal is that you stop needing us.
Out: documentation, handover sessions, no lock-in
What we build with
Contact
The useful first conversation is about a specific decision — which clauses put us at risk, which questionnaire is due Friday, how many units to hold in Q3. Tell us that, and we’ll tell you whether this is worth building.
Worth including
The link opens a draft with these as headings. Delete whatever doesn’t apply.