Automation

Any current model can read an invoice or draft a reply. That is not the hard part and has not been for a while. The hard part is what happens at two in the morning when it reads one wrong and there is nobody watching.

So the model drafts, extracts and summarises, and it never gets to be the last thing between a number and a decision. Invoice extraction is gated on arithmetic it cannot override. Lead scoring can use a model; the routing that follows is fixed rules. A report will not render if it contains a figure that is not in the computed facts.

Five suppliers quote for 120 brackets. On the totals the suppliers stated, the cheapest is £5,946 and a spreadsheet picks it immediately.

It is a quote for 100 units. Its arithmetic is perfect and it is answering a different question, so it is excluded and a re-quote is requested rather than being scaled up to look comparable. The recommendation goes to a supplier £1,047 more expensive, and that is the correct answer.

Automation that did not check would have made the wrong decision faster, which is the usual outcome and the reason people quietly stop trusting these systems.

Website scope can be written down before work starts. Process scope cannot, because the cost is in the exceptions, and nobody can list their own exceptions from memory.

So it starts with a paid audit of the process as it actually runs, which produces a fixed quote and is useful on its own even if you never go further. Pricing an automation from a description, without looking, is how both sides end up unhappy in week three.