Almost every conversation about AI starts in the wrong place. Here’s the framework we actually use when a client asks whether it’s worth pursuing — starting with the task, not the technology.
Almost every conversation we have about AI starts in the same place: someone has read something alarming, or exciting, and wants to know what it means for their business. Usually they’ve been told they need an AI strategy. Usually what they actually need is to stop losing four hours a week to something repetitive.
That gap — between the way AI gets discussed and the way it gets used — is where most wasted budget lives. So here’s the framework we actually use when a client asks whether AI is worth pursuing.
Start with the task, not the technology
The useful question is never “how can we use AI?” It’s “which part of this job is repetitive, text-heavy, or hard to search?” Those three categories cover the overwhelming majority of genuinely good use cases in a small business.
- Repetitive — the same kind of writing, sorting, or categorising done over and over, where the judgement required is low but the volume is high.
- Text-heavy — long documents that someone has to read and summarise, or enquiries that need a first-pass response before a human refines it.
- Hard to search — information the business definitely has somewhere, but which takes too long to find when it’s needed.
If a task doesn’t fall into one of those buckets, AI is usually the wrong tool — and often an expensive way to avoid fixing a process that’s simply broken.
How to tell a real use case from a demo
A demo shows you that something is possible. A use case shows you that something is worth doing. The difference is almost always in the details, and demos skip the details on purpose.
Three questions separate them:
- Does it happen often enough to matter? A task performed twice a month isn’t worth automating, even if the automation would be impressive.
- Is the current cost measurable? If you can’t say how many hours a week it consumes, you can’t tell afterwards whether the tool helped — which means you’ll never know if it worked.
- What happens when it’s wrong? Some tasks tolerate a wrong first draft that a human corrects. Others genuinely can’t. Be honest about which one you’re looking at before you build anything.
The rule of thumb
If a human already has to review the output before it’s used, AI can usually speed that up. If the output goes out unreviewed, be much more careful — and expect to spend real effort on accuracy.
Where it tends to work well
In practice, the wins we see most often are smaller and less exciting than the headlines suggest:
- Drafting first responses to common customer enquiries, so a person edits rather than starts from a blank page.
- Summarising long documents — contracts, reports, meeting notes — into something someone can actually read on the way to a decision.
- Turning unstructured information into structured records, like pulling key details out of an email into a form.
- Answering internal questions against the business’s own documentation, so staff stop interrupting each other to ask where something is.
None of these are transformative. All of them save real hours, and cumulatively they change how much a small team can get through in a week.
Where it usually doesn’t
The failures cluster too, and it’s worth naming them because they’re the expensive mistakes:
- Replacing a decision that requires accountability. AI can inform a judgement call; it shouldn’t be the thing that makes it.
- Anything where a confident wrong answer is worse than no answer. Accuracy matters more than fluency, and fluency is the easy part.
- Tasks that are rare, high-stakes, and hard to verify — the combination leaves no room to catch errors before they cost something.
- As a substitute for a process that hasn’t been decided yet. Automating an unclear workflow just gets you to the wrong outcome faster.
What to measure after you ship it
This is the step most projects skip, and it’s the one that determines whether you learn anything. Before building, write down the specific number you expect to move. After shipping, compare against it — honestly, including when the answer is that it didn’t help.
The measures that matter are boring: hours saved per week, response time on enquiries, how many things get missed, how long it takes to find information. Not “productivity” in the abstract. If you can’t name the number before you start, you’re not ready to build.
The question isn’t whether AI can do this. It’s whether doing it this way leaves you better off than the process you have now — and whether you’d actually notice the difference.
What we do when a client asks
We audit the actual work — what gets done, how often, by whom, and what it costs in time. Then we rank the opportunities by impact against effort and tell the client plainly which ones are worth building and which aren’t.
Frequently the honest answer is that two or three specific improvements are worth doing, and the rest of the list is either premature or better solved by fixing a process. That’s a less exciting conversation than an AI roadmap. It’s also the one that produces a return.