Instrument
How to tell whether a process suits AI
Most businesses start with the question of what artificial intelligence can do. The more useful question is the other way round: which of your routines has the properties that carry an automation. Five of them you can check in an afternoon.
Last updated: July 2026
01
The five questions
- Does it repeat?
- A routine that occurs twenty times a month carries the effort. One that occurs four times a year does not - however irritating it is. The benefit comes from repetition, not from intelligence.
- Is there a rule behind it, even if nobody wrote it down?
- Experienced people decide by patterns they can rarely articulate. That is no obstacle - models learn from examples. But if there is no pattern at all and every decision really is taken individually, there is nothing to take over either.
- What does an error cost?
- A misfiled email costs two minutes. A wrongly costed quotation costs a margin. The more expensive the error, the more checking has to be built in - and the smaller the saving becomes.
- How long does it take today, and how often?
- Multiply. A process taking twelve minutes across forty cases a week ties up a good person-week a month. Only that figure turns an irritation into a business case.
- Is the routine reversible?
- Sorting, matching and preparing can be undone. A rejection already sent cannot. Reversible steps may run independently early on, irreversible ones get a release - and that decides the scope.
02
The question you cannot answer yourself
Can we reach the data the process needs - in sufficient quality, in a format that can be read, with permissions that can be reproduced cleanly?
That cannot be settled from a desk, only by trying. Which is exactly why the access check sits in the audit and not in the quotation. It is the part where projects tip over.
Four of the five questions you can answer without us. The fifth is the one it fails on.
03
A rule of thumb for the first process
The first automated process in a business should not be the most valuable one but the clearest. It has to show that this works, and quickly enough that the workforce can still see the connection between project and effect.
In practice that means: high repetition, low error cost, reversible, data in one system rather than five. A system integration or the sorting of a shared mailbox meets that more often than quotations do - even though quotations move more money.
Frequently asked
How many cases does a process need before automation pays?
As a rough lower bound: the process should tie up at least one person-day a month. For invoice processing the threshold sits, in our experience, at around 300 documents a month; below that the upkeep costs more than the saving.
Does our process have to be documented first?
No. We record it in the audit as it runs. Existing documentation helps but is frequently out of date - what counts is the routine as lived.
What if our data is untidy?
That is the normal case and rarely the obstacle. A model copes with filing that grew over time considerably better than a classic interface does. It becomes a problem when the data contradicts itself - then data quality is the actual project.
Tell us where your time goes.
In the audit we record the process, work out what is worth automating, and you get a plan with fixed prices - credited against the build if you go ahead.