The engagement
Every engagement starts with an Inference Audit: one workflow, mapped, and a decision on where AI belongs. From there the path is fixed. Prove it in weeks against a baseline agreed up front. Place it on the infrastructure the audit chose. Run it with the records your reviewers will ask for.
Map · Prove · Place · Run
Short, fixed scope
We map one workflow — the systems holding its records, who decides what, and what it costs when the work goes wrong — then recommend where AI belongs, and where it doesn't.
Produces A decision on where AI runs, agreed success measures and a measured baseline. The map is yours either way — act on it with us or without us.
Four to six weeks
We build the agent into the workflow itself: it reads your documents, drafts the work, and stops at your approval gates. Then we measure it against the baseline agreed in the audit. No demonstration data. Your records, your environment, the actual job.
Produces A working agent, measured results, and a keep-or-stop number.
The pilot ends at a decision. If the numbers say stop, you stop, and you keep the map, the measures and the record of what was tried.
Sized by the audit
The proven agent moves to production. Where depends on its records: on-premise or air-gapped hardware first, sovereign cloud where local hardware isn't warranted, approved cloud only where you allow it. The permissions, approval gates and audit records move with it — nothing loosens between pilot and production.
Produces The workflow running on the placement decided in the audit, alongside the AI you already run.
Ongoing, reviewable
Quality, cost, policy and model behaviour are watched as the workflow and its source material change, so the system stays inside its rules after the novelty wears off.
Produces A standing audit record your clients and principals can inspect — and the next workflow mapped when you're ready.
After the engagement
An agent doing the job inside the workflow it was measured in, with its evidence and approvals visible.
See what an agent is →Each workload assigned to the placement its risk requires — air-gapped or on-premise by default, cloud only where you approve it — with the reasoning on record.
See where AI can run →Who approved it, what it drew on, what happened next — kept as the work happens, so a client or principal review is answered from the record, not reconstructed from memory.
Where the records can never leave your site, the same engagement delivers to the airon appliance, on your own hardware.
See the appliance →After the first workflow
The first workflow leaves behind the infrastructure the agent runs on, the connection into your records, and the controls that governed it. The second workflow starts on that foundation; the third starts further ahead again. Estimating is a common first move — then the work around it: tendering, contracts, document control, procurement, project controls, HSEQ. Each addition is scoped like the first, measured like the first, and passes its own keep-or-stop gate. Nothing rolls over because the last one worked.
The gains compound the same way — hours back on the desk, more work priced and checked by the same team, evidence ready before the review asks for it.
The shape of it
Setting these boundaries up front saves both sides time.
Start
Bring the workflow that is costing you — the bid desk, the safety response, the report that takes a week. The audit maps it, prices the failure, and tells you where AI belongs. The map is yours either way.