Ten AI use cases that work, by business function

The useful question is not what AI can do. It is which uses survive contact with real volume, real exceptions and a real audit. These ten have, across the businesses we work in. Each entry says what it needs underneath it, because that is usually where the work is.

Finance: coding and matching the routine transactions

Invoices, expense claims and statements arrive in dozens of layouts. A model reads them, applies the coding rules, matches against the purchase order, and puts only the uncertain ones in front of a person.

Needs: the coding rules written down, a tolerance for what counts as a match, and approval authority left exactly where it is today.

HR: answering policy questions from the handbook

The same twenty questions arrive every week and the answer varies with who replies. An assistant grounded in the current policy set answers in seconds, with a citation.

Needs: one current version of each policy, and a decision about what is in scope. It should answer questions and write to nothing.

Operations: turning field paperwork into data

Tickets, logs and forms captured on site get keyed into a system later, often twice. Reading them at the point of capture removes the delay and the second entry.

Needs: a defined document set, and a rule for what happens when confidence is low.

Sales: preparing the first draft of a response

Proposals and questionnaires reuse a large amount of previously written material. Drafting from an approved library is faster than starting from a blank page and more consistent than starting from the last one someone found.

Needs: a library someone maintains, and a reviewer who is accountable for what goes out.

Legal and contracts: finding the clauses that matter

Extracting terms, dates, renewal notice periods and liability caps across a contract set answers questions that are otherwise a week of reading.

Needs: the contracts in one place, and clarity that this is extraction rather than advice.

Customer service: drafting the reply, not sending it

Suggested responses grounded in your own knowledge base raise consistency and cut handling time, with the agent still deciding what is sent.

Needs: a knowledge base that is current. If it is not, this makes bad answers faster.

IT: first-line triage and knowledge

Categorizing tickets, suggesting the known fix, and answering the questions that are already documented. It takes volume off a service desk without changing who resolves anything.

Needs: a service catalogue and articles worth reading.

Procurement: comparing what you are buying

Summarizing and comparing supplier terms, pricing structures and renewal dates across an agreement set — the work nobody has time to do before a renewal lands.

Needs: the agreements gathered, which is usually the actual project.

Marketing: adapting one piece of work into many

Turning an approved source document into the formats a campaign needs, in the voice the business already uses.

Needs: the voice defined somewhere other than in one person’s head, and a review step before anything is published.

Engineering and technical documentation: answering from the manuals

Equipment manuals, standards and procedures are long, searched badly, and consulted under time pressure. Answering from them with a citation beats searching a file share.

Needs: the current revision of each document, and confidence that superseded versions are out of scope.

What these have in common

High volume, a clear notion of a right answer, and a person on the exceptions. The uses that disappoint tend to be the reverse: low volume, contested definitions of correct, and no mechanism to catch a wrong answer before it matters.

Two of these, funded and owned, will do more than all ten on a roadmap.