How we work
From a demo on your documents to a workflow in production. Fixed scope, a number agreed first, and the code is yours.
Every engagement has the same shape. It starts small and on your own data, each stage ends with a written decision, and nothing goes into production until it has been measured against how the work runs today.
The stages
Five stages, each with a written exit. You can stop after any of them.
- 01 · 1 hour, then 1–3 days
Demo on your documents
Send 10 to 20 real documents: a BOQ, a stack of POs, last month's invoices, a backlog of tickets. We run an agent on them and walk you through what it matched, what it flagged and what it refused to guess.
- Results on your own documents
- An honest view of what would and wouldn't work
- A suggested first workflow
Decides the next step: Whether it's worth a closer look.
- 02 · 2 weeks
Workflow assessment
We sit with the people who do the work, measure how it runs today and check what the agent would need to read and write. You get a shortlist of workflows with a scored case for each, and a fixed quote for the first one.
- The workflow mapped as it runs today and with agents
- A baseline taken from your own data
- A systems and data-access check
- A pilot spec with the success metric and a fixed price
- A security and data-flow note for your IT team
Decides the next step: A written go or no-go. The fee is credited to the first workflow.
- 03 · 6–8 weeks
First workflow in production
One workflow, live in your systems, on real work. We build the test set from your history first, so accuracy is measured before anyone relies on it, and part of the fee is held against the metric we agreed.
- The agent running inside your ERP, document system or repositories
- A test set built from your past work, with results
- Approval steps and an audit log
- A dashboard of the agreed metric
- Runbook, code, prompts and tests, owned by you
Decides the next step: The agreed metric, measured against the baseline.
- 04 · 4–6 weeks each
Next workflow
Each further workflow reuses the integrations, test tooling and approval patterns from the first, so it's faster and costs less.
- Same deliverables as the first workflow
- Shared integrations and test sets
Decides the next step: Its own metric, agreed up front.
- 05 · Monthly
Run and improve
We keep the agents healthy: monitoring, model and prompt updates, the test set re-run on every change, a review of the exception queue and a quarterly review of value against the baseline.
- Monitoring and alerts
- Test-set regression on every change
- Quarterly value review
- A named engineer
Decides the next step: Every quarter, against the numbers.
Governance and controls pack
An AI use policy, a control map (approvals, rollback, audit, cost limits) and a data-residency design, mapped to DPDP, PDPL or ISO 42001 where they apply.
Team enablement and handover
Training for the people who review the agent's work and the people who will change it, owner playbooks, and handover of the repository and tests.
What we need from you
Less than you'd think. But we can't do without it.
- Demo on your documents
- One owner of the workflow · A redacted sample, if you can share one
- Workflow assessment
- Two or three people who do the work · Read access to sample data · A 30-minute weekly review
- First workflow in production
- The workflow owner, an hour a week · System access through accounts you control · Users willing to try it on live work
- Next workflow
- A new workflow owner
- Run and improve
- A monthly check-in
Why pilots stall
Five reasons AI pilots stall. What we do about each.
| Reason | Evidence | What we do |
|---|---|---|
| It never touched the real systems | Integration complexity is the most-cited reason AI pilots fail to scale (55% rank it top three, Madrona 2026). | We check system access in the assessment and build inside your ERP, document system or repositories from week one. |
| Security said no | Security, privacy and compliance is the second pilot killer and the top buying criterion (Madrona 2026). | A data-flow note goes to your IT team before we build. Your cloud or region, your accounts, no training on your data. |
| Nobody agreed what success looked like | Gartner expects over 40% of agentic AI projects to be cancelled by 2027 over cost and unclear value. | The metric and the baseline are written into the pilot spec, and part of our fee is held against them. |
| The data wasn't ready | Gartner expects 60% of AI projects without AI-ready data to be abandoned. | The first week often goes on the data itself: the catalogue, the rate library, the item master. You keep that work. |
| The vendor left and nobody could change it | Gartner predicts 70% of agentic AI built by vendor forward-deployed engineers will be abandoned by 2028. | Handover is a deliverable: code, prompts, tests and runbooks in your repository, and people trained to change them. |
Before we call it done
The production checklist. Every workflow passes it before handover.
- 01Outcome checks, not just job statusWe check that records were actually created and match the source, so a job can't run green for weeks while doing nothing.
- 02No duplicates on retryEvery write has an idempotency key and an “already exists” check, so a retry or a resent email never creates a second order or invoice.
- 03Cost capsPer-run and per-month limits on model spend, with an alert before the cap, not after.
- 04Your accounts, your alertsCredentials, cloud resources and alerts sit in accounts you own, and failures page your team, not ours.
- 05A restore testWe restore from backup and replay a day of work before go-live, so recovery is something we've done, not something we hope for.
- 06A named reviewerEvery exception queue has a person's name on it and a target time to clear it.
- 07A runbook someone has usedYour team follows the runbook to change a rule and redeploy, with us watching, before handover.
What you own
When we leave, it stays.
- The code, prompts and configuration, in your repository
- The test set built from your work, and the results
- Runbooks for operating and changing the agents
- Dashboards for the metric we agreed
- Integrations, under accounts you control
Fixed price per stage
The assessment and each workflow are quoted as a fixed price once we've seen your documents and systems.
Part of the fee rides on the number
A share of the first workflow's fee is held until the metric we agreed is met against the baseline.
Running costs you can see
Run and improve is a monthly fee. Model and hosting costs are shown separately, per document or per run, before you commit.
No lock-in
No licence of ours to keep paying. If you stop working with us, the agents keep running and your team can change them.
How we build
Agents build, a different agent reviews, people approve. It's how we write the software we deliver to you.
- needs-operatorWaiting on a person: a brief to approve or a question to answer.
- needs-planThe brief is approved; the plan is being written.
- readyBrief and plan approved. An agent may start.
- workingAn agent is building, with a running log anyone can read.
- for-operatorDone and reviewed. Evidence posted for a person to check, then merge.
Questions
About working with us.
Send us ten real documents. We'll show you what an agent does with them.
A one-hour call with a founder and an engineer. Bring a BOQ, a stack of POs, last month's invoices or a backlog, and leave knowing what would work, what wouldn't, and what it would take.
170+ projects · 65 companies · founders on every engagement