Five phases. Inherited from ERP. Re-tuned for AI.
Most AI projects fail because they're run like software pilots. We run them like ERP implementations — discovery, fit-gap, configure, train, hypercare — because that's the playbook that's already known to work in the room AI has to fit into. The differences are in what gets measured.
Discovery. The week where we stop assuming.
A round of 1:1 interviews with the people who actually do the work — finance, ops, IT, and whoever owns the chart of accounts. The output is a system inventory, a process map for every candidate workflow, and a list of the disagreements between what your software is supposed to do and what your team actually does.
Fit-gap. Where ERP discipline meets AI scope.
For every candidate workflow, three questions: does the existing system already do this (Fit), does it almost-but-not-quite (Gap), or is it a true greenfield (Missing). Only Gaps and Missings go forward into the scorecard. The Fits get a one-line "don't build this" — that answer alone usually pays for the Assessment.
Configure. The Build, run like an implementation.
The Foundation goes in first — Hermes deployed, retrieval indexed, the messaging gateway wired, Langfuse observing. Then the vertical workflows stack on top, one at a time, each shipped to production behind a feature flag, each with its eval suite in place before traffic. The discipline is straight out of ERP go-lives.
Train & eval. People and models, in that order.
Two parallel tracks. Your team learns the operator — what it can and can't do, how to handle reviews, where the audit log lives. Meanwhile the eval suite is locked in: a regression set for every workflow, scored on accuracy, latency, and cost, gating each release. We don't ship a workflow without it.
Hypercare. The first 90 days, run by us.
After cutover, we're on every standup for 30 days, on the channel for 90, and named owners on the runbook for the duration. Hypercare ends when the system has cleared an eval regression cycle without our intervention. After that, you're on an Operate retainer — or off our books entirely.
The methodology is also what we won't do.
Most of what makes an AI engagement go off the rails happens in the gaps between the steps above. So we name the disciplines we don't carry over from the rest of the consulting industry.
No 90-day "discoveries."
The Assessment is two to three weeks. If we need longer, scope is wrong — and the wrong scope kills a project faster than the wrong model.
No demo-driven sales.
We won't build a pretty PoC to win the room. The Assessment is the sales conversation, and it's billed.
No quiet routing to frontier models.
If a workflow leaves your walls for a frontier API, you'll know — it'll be in the architecture diagram, the contract, and the audit log.
No account-handoffs.
The principal who scoped your engagement is the one who's on the hypercare standup. We're sized around that constraint — not around scaling headcount.
Five phases, run by the same person, from week one.
Start with the Assessment. Three weeks, fixed-fee, refundable against Build.