How We Build an Automation System, Step by Step

Phase one: we watch the work happen

We do not start from a requirements document. We start by watching somebody do the task, because what people describe and what people do are reliably different, and the difference is where projects fail.

That means a screen recording, a shadowing session, or sitting with whoever handles the awkward cases. One hour of watching produces more usable specification than a day of meetings.

Phase two: a written scope with a fixed price

You get the workflow written as numbered steps, the exception list, the systems it touches, a phased plan and a fixed price for phase one. You own that document and can take it to another supplier.

The exception list is the part we spend the most time on. A scope with fewer than five exceptions in it is not finished, and we will keep asking until we have them.

Phase three: build, with something running every fortnight

You see working software every two weeks, not a status report. Early demonstrations are rough on purpose — the point is that you can click through the actual path and tell us where it is wrong while changing it is cheap.

• One named developer, whom you meet before signing
• Fortnightly demonstration with your team, not just the sponsor
• Changes logged and priced, including the ones we absorb
• Code in your repository from day one

Phase four: parallel running, then cutover

For two to four weeks the automation runs alongside your manual process and the outputs are compared. It feels like duplicated effort and it is the cheapest insurance available — it finds the 5% of cases nobody mentioned while they still cost an afternoon rather than a customer.

Only when the comparison is clean do we cut over, and the manual route stays available for a defined period afterwards.

What we need from you

1. One person who can make decisions, about half a day a week
2. Questions answered within a day or two
3. Real data, including the awkward records, for testing
4. Attendance at the demonstrations with actual feedback

Projects where the client does these finish two to three weeks earlier than identical scopes where they do not. It is the largest variable we do not control.

Frequently asked questions