The first improvement usually needs no AI
On a field sales engagement that ended with a 3–4% increase in sales, most of the gain came before a model was involved. The change that mattered most was deleting a step, not automating one.
A pharmaceutical client came to us with what they described as an AI problem. Their field sales team spent too much time on administration, head office was always working from stale territory data, and they wanted to know what AI could do about it.
The honest answer, after we watched how the day actually worked, was: less than they hoped, and less than they needed. The expensive part of their process was not a task that required judgment. It was a sequence that required the same information to be handled three times.
What the process actually looked like
Representatives captured orders on paper during calls on doctors and pharmacies. In the evening, after the day's visits, they wrote up activity and entered what they could. Managers then assembled territory-level data into weekly spreadsheets, which head office read to understand what was happening in the field.
Three handlings of the same facts, each introducing delay and each introducing error. Head office was reading a picture that was several days old and partly reconstructed from memory. Representatives were losing selling time to the write-up. Managers were spending their week on data assembly rather than on coaching.
The tempting AI project here is obvious, and we have seen versions of it built elsewhere: an assistant that helps representatives complete their evening reports faster. Voice input, smart defaults, automatic summarisation. It demonstrates beautifully.
It would also have preserved the problem. If we had started with a model, we would have automated the paperwork instead of removing it.
Capture at source, once
What we did instead was redesign the daily routine around recording the work once, where it happens. Orders and visit outcomes went into the device the representative already carried, at the point of the call.
The consequences cascaded. The evening write-up no longer had a purpose, so it was deleted rather than accelerated. Territory roll-ups came from the same records, so managers stopped assembling spreadsheets and the reconciliation arguments stopped with them. Head office saw the same day's activity instead of last week's.
The client reports a 3–4% increase in sales, which they attribute to selling time recovered and to responding faster to territory-level signals. On a large field team that is a material number, and it came from changing the sequence of the work.
AI came later, for the parts that genuinely needed judgment at volume once the data was clean and current. That sequencing was deliberate: the model had something reliable to work with because the process had been fixed first.
How to tell which steps deserve AI
We work through four questions in order, and most steps are resolved before the fourth.
A surprising number of steps exist because of a system that was replaced, a person who left, or an audit that ended years ago. Nobody owns them and nobody has questioned them.
Fewer handoffs, fewer approvals, fewer places the same fact is stored. Changing who signs off is often worth more than any automation.
Often two systems already hold the data and simply are not connected. An integration is cheaper to build, cheaper to run and easier to reason about than a model.
This is where AI earns its cost: work that requires interpretation, happens often enough to matter, and tolerates a defined error rate with human review on the exceptions. If a step fails any of those three tests, a model is the wrong tool.
Why this is uncomfortable to say
We are an AI firm. Telling a client that their first and best improvement requires no AI is not commercially convenient, and there is a version of this business that never says it.
But automating a broken process makes it break faster, at greater expense, with less visibility into why. The AI project that follows a process redesign is smaller, cheaper and far more likely to be used — because by then it is solving a problem that is genuinely about judgment rather than about sequence.
The measurement discipline matters as much as the redesign. Without time, volume, error rate and cost per step, a redesign is opinion, and you cannot tell afterwards whether it worked. That is the part clients most often want to skip and the part that makes the result defensible.
Four to eight weeks gets you a measured picture of how the work runs, a redesigned target process, and a sequenced plan you can fund in stages — including the parts we advise against.
Process Re-engineering →