There's enormous pressure to 'add AI' to everything right now. Some of it is genuinely transformative. A lot of it is an expensive way to do worse what a simple rule already did well. Here's how to tell them apart.
- AI earns its place on messy, language-shaped, judgement-light work — reading unstructured text, drafting, classifying, summarising.
- If a deterministic rule or a plain integration already solves it reliably, adding AI usually makes it slower, costlier, and less predictable.
- The best pattern is often AI drafts, a human approves. You get most of the speed and keep accountability where it belongs.
I get asked to "add AI" to things a lot at the moment. Sometimes it's exactly the right instinct and the result genuinely changes how a business runs. Sometimes it's a solution in search of a problem — an expensive, less predictable way to do something a five-line rule already did perfectly. The hard part, and the most valuable thing a technical partner can offer right now, is telling those two apart before the money is spent.
So here's my honest working model of where applied AI pays off, and where reaching for it is theatre.
What large language models are genuinely good at
Strip away the marketing and current AI has a real, specific superpower: it handles language and ambiguity at a level nothing before it could. That makes it transformative for a particular shape of work — work that used to require a human to read something messy and exercise light judgement.
- Reading unstructured text. Pulling the delivery date, amount, and reference out of an email or a PDF that every sender formats differently. This used to be either manual or a brittle mess of regexes; a model does it well.
- Drafting a first version. A reply, a summary, a description, a first-pass categorisation. Something a human then reviews and approves rather than writes from scratch.
- Classifying and routing. "Which team should this ticket go to?" "Is this feedback a bug, a request, or praise?" Fuzzy judgements over natural language.
- Answering from a body of content. A grounded assistant over your own documentation — retrieval-augmented, cited, honest about what it doesn't know.
Notice the common thread. These are all jobs that are messy, language-shaped, and tolerant of the occasional miss — especially when a human stays in the loop. That's the natural habitat of a language model.
AI is at its best where the alternative was a human reading something and making a small judgement call. It's at its worst where the alternative was a computer following a clear rule.
What they're a bad tool for
The failures are just as patterned, and they're expensive because they look plausible right up until they don't.
Anything that must be exactly right, every time, with no reviewer. Calculating a total. Deciding whether a payment clears. Enforcing a permission. These are deterministic problems with correct answers, and a language model — which is fundamentally probabilistic — is the wrong instrument. It will be right almost always, and "almost always" is precisely the failure mode you can't accept for money or access.
Anything a simple rule already handles. If "route invoices over £10,000 to the finance director" is the whole requirement, that's an if statement. Wrapping it in an LLM makes it slower, costlier, harder to test, and less reliable. Reaching for AI here isn't sophistication; it's adding a probabilistic component to a problem that had a certain answer.
Anything where a wrong answer is expensive and invisible. If a mistake is costly and nobody's checking, the maths on automation stops working — the rare bad output does more damage than all the saved time was worth.
The test I actually apply
When someone asks whether AI fits a task, I run it through a few honest questions:
- 01Does this involve understanding messy, natural language or unstructured input? If no, AI probably isn't the tool — a rule or an integration likely is.
- 02Is an occasional wrong answer survivable, or can a human catch it? If every output must be exactly right with no reviewer, be very cautious.
- 03Would a deterministic rule solve most of it? If yes, do that first. You can always add AI for the genuinely fuzzy remainder.
- 04Can I measure whether it's working? If there's no way to tell good output from bad at scale, you can't operate it responsibly — build the evaluation before the feature.
Only tasks that pass all four are strong AI candidates. It's a smaller set than the hype implies, and the work that clears the bar is genuinely worth doing.
The pattern that quietly wins: AI drafts, a human approves
The most durable AI automations I build usually aren't fully autonomous, and that's a feature. The model does the heavy lifting, a person makes the final call.
messy input ─▶ ┌──────────┐ draft/suggestion ┌───────────┐ ─▶ action
│ AI │ ─────────────────▶ │ human │
└──────────┘ │ approves │
└───────────┘The support agent whose reply is drafted and who edits and sends it. The bookkeeper who sees the AI's extracted figures pre-filled and confirms them. The manager who gets a suggested ticket routing and overrides it when it's wrong. You capture most of the time saving — the blank page is filled, the reading is done — while accountability stays with a human. It's less impressive in a pitch than "fully autonomous," and it's far more likely to still be running, trusted, a year later.
Boring is allowed to win
The unfashionable truth is that a lot of "AI automation" wins are really just automation — a clean integration, a sensible rule, a removed manual step — that never needed a language model at all. And a lot of genuine AI wins are modest, human-in-the-loop tools that quietly save an hour a day rather than headline-grabbing autonomous agents.
None of this is anti-AI. Applied well, to the jobs it's actually suited for, it's one of the most useful capabilities we've had in years. But the value comes from matching the tool to the task honestly — which sometimes means the most valuable advice is that the impressive-sounding AI feature should be a plain rule, and the money is better spent elsewhere. That answer wins no awards. It just happens to be right more often than the alternative.