There’s enormous pressure to “do something with AI.” That pressure is exactly why so many AI projects quietly fail: they start from the technology and go looking for a problem, instead of starting from a costly problem and asking whether AI is the right tool.
Start with the expensive, repetitive work
AI tends to pay back fastest on high-volume, repetitive, language- or document-heavy tasks: extracting data from documents, classifying and routing incoming requests, drafting first-pass responses, summarising long inputs. The test is simple — is a person doing something repetitive, at volume, that mostly follows patterns?
Keep humans in the loop
The reliable pattern isn’t “AI replaces the team.” It’s “AI handles the volume, people handle the judgement.” Design the workflow so AI does the heavy lifting and a person reviews or handles exceptions. That’s where you get speed and trust.
Where AI doesn’t pay back
- Low-volume tasks where setup costs more than it saves.
- Work that demands perfect accuracy with no room for review.
- Problems that are really just missing automation or integration — no AI required.
That last point matters: a lot of what people ask AI to do is actually a plumbing problem. Connect two systems, automate a handoff, and the “AI use case” disappears — cheaper and more reliable.
How to decide
Before building anything, audit your processes and rank opportunities by payback. That’s the entire purpose of an Automation Audit — and sometimes its most valuable output is “don’t build that yet.”