What does it cost to put one AI agent into one real workflow, and how do we know it worked?

1. 01
   Select and baselineConfirm the workflow, usually the one an Opportunity Map named first. Measure how it runs today: volume, time per item, error rate, cost. No baseline, no Pilot.
2. 02
   DesignOne page, signed by your sponsor: what the agent reads, what it may do, the permissions it gets (the fewest that work), where a person checks it, what it must never do, what is logged, and a cost ceiling.
3. 03
   BuildAgainst a copy or sandbox of your systems first. Before the agent is written, a test set is built from your real past items with known correct outcomes.
4. 04
   Test and hardenRun the test set. Measure accuracy, cost per item and speed. Add guardrails for the mistakes seen. Set the confidence level below which items go to a person.
5. 05
   Shadow run, at least two weeksThe agent works on live items, but a person reviews every output before anything happens. The results are compared with the baseline.
6. 06
   Result and handoverThe numbers, the runbook, the failure register, and your decision: scale it, keep it in shadow, or stop.

Workflows that tend to suit a Pilot have a clear trigger, clear inputs, a definable "done", and a person who can check the result. For example:

• qualifying inbound enquiries
• handling supplier documents such as order acknowledgements, invoices and delivery notes
• triaging purchase-order exceptions: short shipments, price mismatches, late deliveries
• an internal assistant that answers from your own documents, with references
• entering orders from unstructured emails into the ERP, with a review step

Where the human sits. During the shadow run, a person approves every output. After it, items below the agreed confidence level still go to a person. You decide at the end whether it goes live.

• Access to the systems involved, or a copy. The agent is built against a copy or sandbox before it touches anything live.
• Two weeks of historical items with known outcomes, to build the test set.
• A reviewer for the shadow run, about two hours a week.
• A sponsor who signs the design and makes the decision at the end.
• If there is no API, the agent works from exports, a mailbox, or a read-only copy of the database. The route is written into the design.
• If there is no baseline, measuring how the work runs today comes first. Where the data to measure from exists, that measurement is inside the price. Where it does not, the proposal says how long the measuring takes and prices it separately.

This is about how the Pilot is delivered, not about the agent the Pilot builds for you, whose permissions are set in the signed design. The Pilot is led by Sam Naseem, and this section is in his words.

• Done by me: the signed design and its permissions, the route into each of your systems, the guardrails and never-do list, the cost ceiling, and every step that moves the agent from the copy or sandbox to live items. What stays with me is chosen by risk and reversibility.
• Or by a trusted engineer, under my review: part of that work may be done by an engineer I have worked with, named in the proposal before they have access to your systems.
• Done by AI agents I have trained and direct, under my review of every design decision and every deliverable: the engineering against a copy or sandbox of your systems, building and running the test set, and first drafts of the runbook and reports.
• Who checks it: I review every design decision and every deliverable. During the shadow run, your reviewer approves every output of the Pilot's agent, and you decide whether it goes live.
• What AI agents helping with the work are never allowed to touch: your live systems and their credentials.
• How your data is handled: the test set is built from your real past items. It is kept on AI UVD's systems on AWS in London, or in your own environment if you prefer. Where an AI service helps, it is a business or API plan that has a data processing agreement and does not train on your data. Working material is deleted or returned within 30 days of handover.

• Price: £12,000, fixed. A Pilot starts from a completed Opportunity Map, or from a workflow that is already that clear. Prices exclude VAT.
• Baseline measurement: inside the £12,000 where the data to measure from exists — that is, your systems, records or spreadsheets can show volume, time per item, error rate and cost. Where they cannot, measuring how the work runs today is quoted separately in the proposal, and the four to six weeks start when that measurement is finished. A Pilot that starts from a completed Opportunity Map always has its baseline, because the Map names it, so the separate quote is the exception.
• Model and hosting costs: passed on at cost, under a ceiling agreed in the design.
• Duration: four to six weeks, of which the shadow run is at least two.

Why the price can be fixed and the timeline short: AI agents I have trained and direct do part of the engineering, drafting and testing, under my review of every design decision and every deliverable, so fewer engineering hours go into each Pilot. The design, the route into your systems and every step towards live work stay with me.

These durations and hours are the method as designed, not an average of Pilots already delivered. Where a figure on this page is measured, it is shown with the date it was measured.

Not included: a second workflow, going live without the shadow run, ongoing operation, model costs beyond the agreed ceiling, and — where your systems and records cannot show how the work runs today — the period spent measuring it, which the proposal prices separately and which the four to six weeks start after. Ongoing operation is your team with the runbook, or the Fractional AI Architect retainer, within its days each month.

The boundary, in plain words: one workflow. A second one worth doing gets its own Pilot. The Pilot ends with a runbook and a number.

• A signed one-page design: permissions, where the person sits, the never-do list, the cost ceiling.
• The test set and its results.
• The agent: code and deployment notes.
• The guardrail and permission specification.
• The shadow-run report: time per item, error rate and cost per item, against the baseline.
• A runbook your own people can use to operate, monitor, pause and roll back the agent.
• The failure-mode register, updated from what actually happened.
• A recorded handover session.

You don't get: a second workflow, going live without the shadow run, ongoing operation, model costs beyond the agreed ceiling, or a measuring period where there is no data to build a baseline from — that is priced separately in the proposal.

• Businesses without a clear workflow yet. Start with an Opportunity Map.
• Anyone who wants several workflows at once.
• Anyone who wants it live without a shadow run.
• Businesses with nobody who can review outputs for about two hours a week.
• Businesses with their own team already building agents.