From your process to an agent in production, in ten stages.

Each stage has an observable outcome. If it is not reached, the next one does not start. That is what stops a project drifting for six months.

The full sequence

Two stages are non-negotiable: the second, because without the cost beforehand nothing can be demonstrated, and the fifth, because writing rules without having watched the agent work means guessing them.

  1. 01

    You recognise your own process in what we describe

    We arrive naming a process, a volume and an order of magnitude of cost. Not asking you where your friction is.

    Observable outcomeYou accept a meeting.

  2. 02

    Scoping

    Which process, what volume, what it costs today. We measure time per unit, loaded cost, error rate and delay.

    Observable outcomeYour figures, written down and validated by you.

  3. 03

    Decision

    Scope, price, timeline, and what stays human. All settled before a line is written.

    Observable outcomeA signed contract, with the before figure in it.

  4. 04

    Installation on your infrastructure

    Your cloud, your host or your machines. We host nothing.

    Observable outcomeThe platform runs on your own infrastructure.

  5. 05

    Observation

    The agent handles real cases and shows what it would do, sending nothing, writing nothing, paying nothing. Minutes of compute, half an hour to walk through.

    Observable outcomeIts real behaviour, documented, on your own cases.

  6. 06

    Writing the rules

    Tools, caps, rights, what gets escalated. Written from what the observation showed, not from what we imagine.

    Observable outcomeRules your business people can read, not just an engineer.

  7. 07

    Progressive go-live

    A narrow scope first. The agent acts, the hard cases escalate, we widen as evidence accumulates.

    Observable outcomeThe agent acts and nothing breaks.

  8. 08

    Measurement

    The trail exported on the agreed date, cross-read with what your teams report.

    Observable outcomeVolume handled without intervention, the cost after, the gap with stage 2.

  9. 09

    Extension

    A second workflow, by us or by you. Either suits us, and the second costs less than the first.

    Observable outcomeYou choose.

  10. 10

    Renewal

    The subscription continues because you can see what you are paying for.

    Observable outcomeYou renew without having to argue about it.

You watch the agent work before it is allowed to act.

On your real files it produces everything: the decision, the amount, the recipient. And it sends nothing. You judge the evidence, then we write the rules.

This is not in-house caution. Shadow deployment is the go-live method Chip Huyen sets out in Designing Machine Learning Systems, and Symrise validated every suggestion by hand for two years before automating. Almost no agent vendor applies it to agents.

Three people at your end, three expectations

One platform, three readings. The person who can block a deployment has to find their answers unaided.

Whoever pays

The CEO or CFO. What the process cost, what it costs now, what the agent handled, what it spent. Without opening a terminal or reading YAML.

The team that lived the process

What was done, what was escalated and why, how to take a case back. The agent removes the repetition, not the judgement.

IT and security

Where it runs, what the agent may call, how a right is revoked on the fly, what the trail says.

The first stage costs nothing.

Describe your process, out loud or in writing. We send back what it costs and what an agent can take out of it.