Start with the problem, not the tool

Someone asks for an AI agent or a new system. Underneath, there is usually something simpler: invoices go out late, enquiries sit unanswered, two systems disagree. The tool is one possible fix, and rarely the cheapest.

So the first conversation is about the problem. Who has it, what would better look like, and how will we know? The tool comes after.

Build it, if nothing on the market does the job. Connect something that already exists, if it does. Automate it, if the steps are predictable. Or hand it to a person, if it needs judgement.

Most real fixes use a bit of each. An off-the-shelf product holds the records, a small custom piece handles your unusual rule, AI reads the incoming documents, and a person approves anything risky.

A person in the process is often the right design. Someone has to own a big decision, hold the passwords or turn up on site. We say up front which parts a machine should not touch, so nothing fails quietly later.