Real numbers from a recent legal contract-AI build. Eight categories, 312 test cases, weighted by client priority. Same suite runs on every commit, and on every frontier model release. No green bar, no deploy.
We sell narrowly-scoped AI workflows that are measurably better than the spreadsheet they replace. No "AI strategy", no AI ethics workshop, no glossy demo that breaks on your real data. Just the thing, evaluated, in your stack.
Senior associates spending 6+ hours per MSA on first-pass review. Compliance bottleneck.
On-prem AI reviewer fine-tuned to the firm's playbook, integrated with iManage. Eval suite: 312 cases across 8 contract categories.
Four FTEs reviewing KYC files by hand, 78% first-pass error rate, audit anxiety.
Custom document-AI pipeline pulling identity, source-of-funds, sanctions evidence. Human reviews exceptions only.
Consultants reading CVs that didn't fit briefs. Half a day per consultant, wasted.
Retrieval-grounded ranking over CV history with auto-drafted outreach. Two-way Bullhorn writeback with full audit log.
Almost every AI engagement we ship lands in one of these four buckets. We say which in the scoping call - and which model, vector store and host before we charge a penny.
Four engagement shapes. Bespoke scope, bespoke pricing per business - we quote a fixed fee after a free 30-minute scoping call.
The eval is week one's deliverable. It's the contract between "looks impressive in demo" and "actually works on your real data".
Every AI build comes with the same engineering kit. The quality bar is the bit other AI vendors don't bother to write - and the bit that pays for itself the first time a frontier model release would have broken something.
"They scoped what we were trying to do in 40 minutes. The previous shop took six weeks and got it wrong."
"We were quoted £450k by a Big Four practice. AssurePath delivered the same outcome in eight weeks for £62k."
"Honest, technical, on time. They told us not to do half the things we asked. That's why we trust them."
AI is rarely the answer on its own. We redesign the process around it. Map → redesign → automate.
ExploreWhen the AI is the product, not the feature. Customer-facing SaaS, internal tools, copilots.
ExploreWe keep the AI running - frontier model upgrades, drift watch, regression guarding. Fixed monthly fee, UK engineers.
ExploreEight of the questions clients ask before commissioning an AI build.
Before we write a prompt, we write the test suite. Fifty to five hundred real cases drawn from your data, scored by your experts, weighted by what actually matters. It runs on every commit during the build, and on every new frontier model release after go-live. We don't deploy a version that regresses against it - the model that's good enough in week three has to still be good enough next year when a frontier model upgrade lands.
