AI Problem Definition: If You Can't Explain the Problem, AI Can't Solve It

*Part 2 of 6 in a series on the habits that make AI useful. Previously: [AI Curiosity Beats Technical Ability](/insights/ai-curiosity-beats-technical-ability/). Next: [Learning AI: The People Who Learn Fastest Expect It to Fail](/insights/learning-ai-expect-it-to-fail/).* Someone asks AI to "improve our marketing", gets back a list of platitudes about knowing your audience, and concludes the tool is useless. The gap is AI problem definition, not model quality. The tool isn't useless. The instruction was. Improve it how? More enquiries? Better quality enquiries? Lower cost per lead? Clearer positioning against the two competitors who keep beating you on price? Each of those leads somewhere completely different, and none of them was in the request. Most of the advice aimed at this problem is about prompting — templates, frameworks, prompt libraries. That helps at the margins. But a lot of what looks like a prompting problem is a thinking problem, and no template fixes that. ## AI makes vague thinking visible This is the uncomfortable and genuinely useful part. When you brief a colleague vaguely, they fill in the gaps. They know the business, they know what you meant, they know last quarter went badly and that's probably what's behind the request. You never find out how vague you were. AI doesn't do that. It takes the brief at face value and produces something exactly as unfocused as the instruction. What you get back is a fairly accurate mirror of how clearly the request was framed. Plenty of leaders have had the experience of writing a prompt, reading it back, and realising they couldn't have answered their own question either. ## Start with the outcome, not the tool We work problems first, tools second, and this is where it bites. The common opening question in an SME is *"where could we use AI?"* It sounds proactive. It usually produces a shopping list of tools and a pilot nobody uses. The better opening is *"where is work slow, repetitive, expensive or frustrating?"* Answer that first, in the language of the business, then ask whether AI has a useful role in any of it. Sometimes the answer is no and the fix is a form, a template or a conversation between two departments — which is a cheaper result than a subscription. We wrote about this at length in [Problems First, Tools Second](/insights/problem-first-tools-second/), and it remains the single biggest predictor of whether an AI project goes anywhere. ## A good AI brief looks a lot like good delegation If you have ever handed work to someone and been disappointed with what came back, you already know this checklist: - What are we trying to achieve, and why? - Who is this for? - What information is relevant, and where is it? - What constraints exist — tone, length, budget, legal, brand? - What does a good result look like? - What must it avoid? That is a delegation brief. It is also, almost word for word, what makes AI output useful. The overlap is not a coincidence: both are cases of getting something out of a system that cannot read your mind. It follows that managers who brief people well tend to get good results from AI quickly, and managers whose team quietly redoes everything they've been asked to do tend to struggle. That's a management insight disguised as a technology one. ## You can't improve a process you don't understand The other half of this is process. People frequently ask AI to automate something nobody has ever written down, and are surprised when the result doesn't match what actually happens. Before automating anything, walk the process end to end: **Input → decision → activity → output → review.** Who starts it? What triggers it? Where do they get the information? What judgement calls get made, and by whom? What happens to the output? Who checks it, and against what? Half the time this exercise finds the real problem before AI is mentioned: two people doing the same rekeying, an approval step that exists because of an incident in 2019, a report that three people receive and nobody opens. Our [Process Mapping](/services/process-mapping/) work is largely this, and it is usually where the money is. Automating a mess just produces the mess faster — [as we've said before](/insights/automate-broken-process-break-it-faster/). ## Zoom out before you zoom in Most people start small: *can AI write this email?* Fine, and often useful. The bigger question is *why are we writing this email at all?* If you send the same update forty times a month because customers can't see the status of their order, the opportunity isn't a faster draft. It's the status update they never had to ask for. Small tasks are a good place to build the habit. They are rarely where the value is. The value tends to sit one level up, in the workflow that keeps generating the task. ## Six questions that produce better AI projects When a team brings us an idea, we ask the same things: 1. What are we trying to achieve? 2. What happens today, step by step? 3. Where is the friction — time, errors, waiting, rework? 4. What information do we already hold, and is it any good? 5. What would better actually look like, in numbers? 6. Where, if anywhere, might AI help? Notice AI appears once, at the end. A team that can answer the first five has a project worth doing. A team that can only answer the sixth has a tool they'd like to justify. ## Clear thinking is the transferable skill AI literacy and business literacy are converging. The person who can define a problem, describe a process and say what good looks like will get more from these tools than someone who has memorised a hundred prompts, because they can tell the difference between an impressive answer and a correct one. That is a skill worth developing in your managers regardless of what happens to the technology next. ## FAQs ### What makes a good AI prompt? Context, a clear goal, an example of what good looks like, and explicit constraints. In practice, a good prompt is just a good brief written down. If you'd struggle to hand the same instruction to a capable new starter, it isn't ready. ### Is prompt engineering still important? Less than it was. Models handle sloppy phrasing far better than they did, so the leverage has moved to defining the problem, supplying the right context and judging the output. Basic prompting technique is worth an hour of your time, not a course. ### How much context should I give AI? More than feels necessary. The relevant background, the audience, the constraints, and a sample of previous work you were happy with. Most disappointing output comes from missing context rather than a missing technique. ### How do I identify good AI use cases? Look for work that is frequent, repetitive, text or data heavy, and where a person can quickly tell whether the output is right. Then check whether the process is understood and the information is available. Frequency plus checkability is the sweet spot. ### Which business processes are easiest to improve with AI? Typically drafting, summarising, research, classification, document comparison, internal knowledge retrieval and recurring communications. These are common across SMEs and rarely need integration work to be useful. ### Should I automate a process that isn't working well already? No. Fix or simplify it first, then automate what remains. Automation locks in whatever the process currently does, including the parts that are wrong, and makes them harder to see.

This article was written by Peter Lowe. The ideas and opinions are his own; AI was used to assist with drafting and editing.