Where should an established business actually use AI?
The short answer
AI earns its place where the input is messy and varied but the output is checkable — document extraction, classification, drafting and answering questions from your own material. It is a poor choice where a rule, a search box or a well-written page would do the job, and a dangerous one where nobody can tell whether the answer is wrong.
Start from the problem, not the technology
Most AI projects that disappoint were commissioned the wrong way round. Somebody decided the business should “use AI”, then went looking for somewhere to put it.
The useful question is narrower: where do we currently pay a person to interpret messy input and produce a predictable output? That is the shape of work language models are genuinely good at, and it is more common in an established business than people expect.
The test: messy in, checkable out
AI is a good fit when both halves are true.
Messy input means the input varies in ways rules cannot capture. Invoices from two hundred suppliers, each with its own layout. Job adverts written by different managers. Web pages that defeat ordinary parsing. If you could write a reliable rule, write the rule — it will be faster, cheaper and auditable.
Checkable output means somebody or something can tell whether the answer is right. An extracted invoice total can be reconciled. A classification can be spot-checked. A drafted reply is reviewed before it is sent.
Where the output cannot be checked, you are not automating work. You are generating confident text nobody can verify, which is worse than having nothing.
Four places it usually pays
1. Extraction from documents
Pulling structured data out of PDFs, emails and web pages. This is the least glamorous application and reliably the best value, because the alternative is a person typing.
In Yeppe, crawlers read school team pages and directories, with an AI fallback only for pages ordinary parsing could not handle. That ordering matters: cheap deterministic parsing first, the model as the exception.
2. Answering questions from your own material
When expertise is locked in guides, policies and slide decks that nobody can find, an assistant grounded in that material is genuinely useful — provided it answers only from your documents and links to the source.
Ask Adjust works this way: answers come from Adjust’s own guidance rather than the open internet, link back to the guide they came from, and follow the organisation’s house style, including terminology the sector has moved away from.
3. Classification and routing
Which team should this enquiry go to? Is this bounce permanent or temporary? Which of our forty categories does this product belong in? High volume, tolerant of occasional error, easy to sample for accuracy.
4. First drafts
Not final copy. First drafts that a person edits — proposal sections, replies to common questions, summaries of long threads. The time saved is real and the human stays in the loop by design.
Four places it usually does not
Where a rule would do. If the logic is “orders over £500 need approval”, write that. A model will cost more, run slower and occasionally get it wrong.
Where search would do. A lot of “we need an AI assistant” is really “we need to be able to find things”. A good search box is cheaper and more predictable.
Where accuracy must be total. Anything that touches payroll, regulatory submissions or safety. Use AI to prepare the work, not to make the final determination.
Where you cannot afford to be wrong in public. Customer-facing generated content, unreviewed, is a reputational risk that almost never justifies the saving.
The cost question nobody asks early enough
Model calls are cheap individually and expensive in aggregate, and the bill arrives monthly rather than up front.
The failure mode is a per-record AI call: a system that processes ten thousand records and calls a model for each one. Design it so calls are event-triggered and cached — the same input should never be sent twice — and put a hard cap on spend. Every AI call we build works this way, specifically so cost cannot surprise anyone.
How to choose your first one
Find a task where someone currently reads something and writes something, where volume is high enough to matter, and where a wrong answer is annoying rather than catastrophic.
Build that one. Measure it honestly against the manual baseline. Then decide whether there is a second.
The businesses that get the most from AI are not the ones that used it in the most places. They are the ones that found two or three tasks where it genuinely fits, and resisted the rest.