Most AI coverage is noise. Every week brings a new claim that some category of knowledge worker is finished, counter-claims that nothing will change, and demos that bear no resemblance to real-world deployment. None of it helps you work out whether AI could make one specific thing in your organisation faster, cheaper or better.
Here's a grounded view based on what actually works in operational contexts right now — and what doesn't.
Where AI genuinely earns its keep
Summarisation and extraction. If your team reads documents, long emails or reports to pull out specific information, AI does this faster and more consistently than a tired human at 4pm on a Friday. Intake forms, case notes, supplier contracts, complaint letters — anywhere structured facts are buried in free text. A counselling service extracting referral details from GP letters, or a utilities firm pulling key dates out of supplier contracts, can turn an hour of reading into a minute of reviewing.
Classification and routing. Incoming requests, tickets and complaints that need categorising and sending to the right person are well-suited to AI-assisted triage. It doesn't need to be perfect — 'right 90% of the time, flagged when unsure' still removes most of the manual sorting.
Drafting. First drafts of routine communications — acknowledgements, follow-ups, status updates, meeting summaries — can be generated and edited rather than written from scratch. The human still reviews and sends. The blank page is what disappears.
Pattern spotting. Flagging records that look unusual, identifying trends in volumes of operational data, noticing the renewal pattern or the complaint cluster before it becomes a problem. Consistent attention over large datasets is exactly what these tools do well.
Where AI doesn't help — and won't soon
Judgement calls that depend on your specific context and carry real consequences. Anything involving accountability, client relationships or regulatory duty of care needs a human making — and owning — the decision. AI can inform; it can't be responsible.
It also doesn't act on the world by itself. AI models don't update your CRM, send the email or book the visit unless someone builds that integration — which is an engineering project, not a feature toggle.
The part nobody mentions: your data
Most operational AI failures aren't model failures — they're data failures. If the underlying records are inconsistent, incomplete or scattered across systems, AI amplifies the mess rather than fixing it. The businesses that get value from AI are almost always the ones that sorted their data foundations first. Which is, conveniently, worth doing anyway.
For organisations handling sensitive personal data — health, counselling, safeguarding contexts — there's a further question of where the data goes when it's processed. This is solvable with the right tooling and contracts, but it has to be designed in, not discovered later.
Where to start
The wrong question is 'where can we use AI?' The right one is 'what is taking our team the most time that doesn't require human judgement?' Pick one task — high volume, low judgement, clearly checkable output — and pilot it properly with a human reviewing everything. Measure the time saved and the error rate before scaling.
The useful applications are narrower than the headlines and more specific to your processes than any product demo admits. But implemented properly, on the right task, the time saving is real — and it compounds.