Separate what varies from what doesn't
Before deciding where AI goes in a piece of work, watch the work. Most of it usually doesn't change, and the part that doesn't change doesn't need a model.
October 2026
Watching before building
On one Vodafone programme, engineers were writing configuration scripts for a type of network link by hand, one at a time, at 20–25 minutes each. The obvious request was to produce them faster.
Instead, I sat with the engineer and watched the work being done. The scripts weren’t really bespoke. Most of each one was identical, and only a small amount changed from site to site. So I treated it as a template: a macro fetched the site-specific data and filled in everything else.
Scripts went from 20–25 minutes to under 5. Transcription errors went away, because the only manual step left was a single validation. That’s The Macro.
That’s not AI, and that’s the point
The macro had no intelligence in it. It didn’t need any. Once the fixed part and the variable part were separated, the fixed part was plain automation and the variable part was a lookup.
I think this is the step most often skipped in AI programmes. A team sees a slow, manual process and asks where a model fits. The better first question is: what in this work actually changes from case to case? Only that part might need judgement. Everything else is structure, and structure should be fixed, tested and boring.
Where the judgement goes
Redline splits the same way. Every review has the same shape: a quoted clause, a severity such as Severe or High, and a suggested counter-offer. That shape never changes, so a reviewer always knows what they’re looking at and where to check it.
What varies is the contract: which clauses are risky, how much harm each could do, and what a fair counter-offer looks like. That’s where the AI’s judgement is spent. Keeping the output structure fixed is what makes the variable part reviewable.
The method
Watch the real work. Find the repeating structure. Separate what varies from what doesn’t. Automate the part that doesn’t, and only then decide whether the part that does needs AI.
It’s the method I used on network scripts long before I built anything with a model, and I use it the same way now.
Don’t ask people to be reliable at repetitive work. Change the system so the error can’t happen.