Look for reading, sorting, drafting, searching
AI is at its most useful where a person currently reads something, decides what kind of thing it is, and produces a first version of a response. Those steps exist in almost every business: enquiries, invoices, applications, complaints, supplier documents, meeting notes.
They also share a useful property — the work already happens on a known volume, so the time saved can be measured rather than asserted.
Where it does not belong yet
- Decisions with legal, financial or safety consequences, without a person approving them
- Anything where you cannot describe what a good answer looks like
- Processes that are broken — AI accelerates the mess
- Work that happens twice a month and takes ten minutes
Put it inside the tool people already use
The most common reason a pilot dies is that using it required a detour: a separate tab, a separate login, a copy and paste. Where the drafted reply appears in the inbox, or the classification appears on the record, adoption stops being a training problem.
Ground it in your own material
Generic output is the fastest way to lose trust. Answers should come from your documents, your past correspondence and your policies. This is also what makes AI defensible internally — you can point at where an answer came from.
The question is never “can AI do this?” It is “what does this workflow cost us today, and what does it cost with AI in it?”
Start with one workflow
Pick a workflow with volume, a clear definition of good, and a human already reviewing the output. Test it on real historical examples before it touches anything live. If it does not beat the current process on those, it will not beat it in production.
Agree the boundary in advance
Write down what AI decides, what it drafts for approval, and what it never touches. That single page prevents most of the arguments that stall AI work later.
