// What we won't build
Some of this doesn't work yet. Here's which parts.
This is the list we give a carrier in week two. Each item is something we have been asked to build, will not build, and can tell you what we would build instead.
Nothing on this page is about us. It is about the state of the work.
| We won’t build | What we build instead |
|---|---|
| Automated adverse decisions | The decision file, the audit trail and the human-override specification, with a named decision-maker. Responsibility for an AI decision is non-delegable — New York DFS puts it plainly, that an insurer “may not rely solely on a third-party's claim of non-discrimination or a proprietary third-party process to determine compliance with anti-discrimination laws.” |
| An ultimate-severity model for commercial auto | Fleet-level frequency prediction from telematics, because that works. Severity is being set in courtrooms by damage anchoring, jurisdiction and jury composition, none of which is in your submission data. |
| A materiality conclusion on a manuscript-form diff | The diff. Every change surfaced with its location. We build the diff; we do not build the conclusion, because a wrong answer there is worse than no answer. |
| A dashboard over unfixed ingest | The ingest. A dashboard on bad reconciliation is a confidently wrong answer with better typography. |
| Universal commercial rating | Comparative rating where the APIs exist, and a named list of where they do not. Most of the largest US commercial carriers have no quoting API at all, so every connection is a bilateral partnership — anyone promising universal commercial rating is describing a business-development outcome as a software feature. |
| A rip-and-replace of a spreadsheet that works | Automation around it. If the spreadsheet is doing the job, the build reads from it rather than requiring its replacement first, and we will say so in week two rather than in month six. |
| Anything without a named human decision-maker | The escalation path, the QA seat, and the exception queue, scoped with the build. A system nobody reviews has moved work rather than removed it. |
Swipe to see the full table →
Every one of these has cost us scope. We would rather lose it here, on a page you can read before a call, than in month four of a build.
One build, published in full.
The servicing build we ran inside our own carrier, with the definitions that make its number checkable and the parts that were hard.
Read the case study →Our Fall cohort is already 75% full. Reclaim your operations now.
Bring us the workflow you’ve already tried to fix.
A decade of running an AI-native carrier, pointed at one of your workflows for four weeks.