Where AI Actually Helps (And Where It Doesn't)

Not every problem needs AI. In fact, most don’t.

After 30 years working with workflow and document management systems, and three years following AI closely, the pattern is clear: the organisations that get real value from AI are the ones that understood their problems first. The ones that don’t? They bought a tool, pointed it at a mess, and got a more expensive mess.

Here’s a practical way to think about it.

Where AI genuinely helps

Finding patterns in messy information

If you’ve got years of documents, customer records, or internal knowledge scattered across systems, AI is very good at making sense of it. Not replacing it — making it searchable, sortable, and useful. Think: “find me every time we dealt with this kind of complaint” across five years of emails and notes. That used to take a person a week. Now it takes seconds.

Reducing repetitive coordination work

A surprising amount of office time is spent chasing status. “Where is this up to?” “Did anyone reply to that?” “Who’s handling this?” If the answer lives in a system somewhere, AI can surface it without you asking five people. That’s not replacing anyone’s job — it’s giving people back the hours they spend asking around.

Drafting and summarising

First drafts, meeting summaries, turning rough notes into something structured — AI is good at this. Not because the output is perfect, but because starting from a draft is faster than starting from a blank page. The key: you still need someone who knows the subject to review and refine it.

Routing and triage

Small decisions about where things should go — which inbox, which team, which priority level — are exactly the kind of pattern-matching AI handles well. Especially when you’ve got clear rules and enough examples to learn from.

Where AI doesn’t help

When the problem is coordination, not information

If your team spends half the day chasing status updates, the problem isn’t that you lack a clever search tool. The problem is that work isn’t visible. Adding AI to a broken process just gives you faster access to the same confusion. Fix the workflow first.

When you’re using it for the wrong layer of the problem

AI is probabilistic — it’s making its best guess based on patterns. That used to mean “don’t trust it for anything precise.” But the smarter approach isn’t to avoid AI on those problems — it’s to use AI at the right layer. Need a tax calculation? Don’t ask the AI to calculate it. Ask the AI to write the formula or the code that calculates it. The AI builds the deterministic tool; the deterministic tool gives you the exact answer. AI for the thinking, code for the maths. The mistake is using AI where you need precision, when you should be using AI to build the thing that gives you precision.

When your knowledge isn’t written down

This is the big one. AI needs information to work with. If everything your business knows lives in people’s heads — tribal knowledge, undocumented processes, “ask Sarah, she knows” — then AI has nothing to draw from. Step one isn’t buying an AI platform. Step one is getting knowledge out of people’s heads and into a system.

When privacy isn’t sorted

You can do powerful things with AI when you control where the data goes. But the real challenge isn’t just “don’t send data to the cloud.” It’s what happens when AI systems start talking to each other.

Think about where we’re heading: personal AI assistants, corporate AI systems, and customer-facing AI — all potentially interacting. Your personal AI knows your calendar, your notes, your thinking. The company AI knows client data, financials, internal strategy. When those systems need to work together, who decides what gets shared?

Privacy contracts between people sound nice in theory. In practice, they come with pressure — from managers, from clients, from convenience. The only reliable approach is making the boundary technical, not just policy. The computer says no. Not “the policy says you shouldn’t” — the system literally cannot share what it doesn’t have access to. Cryptographic boundaries, not trust-based ones.

Until an organisation has that sorted — clear lines between what’s personal, what’s corporate, and what’s shared, enforced by the system rather than by good intentions — AI on sensitive data is a liability.

The practical test

Before reaching for AI, ask yourself:

1. Can I describe the problem clearly? If you can’t explain what’s going wrong to a person, you can’t explain it to an AI.
2. Is the information already somewhere? AI can’t work with knowledge that only exists in someone’s head.
3. Would “good enough” be useful? If a 90% accurate answer saves hours of work, that’s valuable. If 90% accuracy causes real harm, it’s not.
4. What happens if the AI is wrong? If there’s a human reviewing the output, the risk is low. If not, think carefully.

The bottom line

AI is a tool, not a strategy. The organisations getting real value from it are the ones that understood their problems first, got their information in order, and then applied AI selectively — where it had a clear job to do.

That’s not exciting. It’s not a revolution. It’s just useful. And “useful” is what actually moves a business forward.