Most conversations about AI in business start at the wrong end. They start with a model, or a tool, or a vendor — and then go looking for a problem to attach it to. The result is usually a pilot that impresses in a demo and quietly stops being used six weeks later.
The useful conversation starts somewhere less exciting: with the handoffs.
Work does not break inside a step. It breaks between them.
Take any internal process and write down what actually happens. An email arrives. Someone reads it and decides what it is. They open a second system and copy three fields into it. They attach a document that lives in a third place. They notify a colleague, who checks something, and either approves it or sends it back.
Every individual step here is fine. Nobody is doing their job badly. The cost is in the seams — the reading, the deciding, the copying, the chasing. That is where the hours go, and that is the part almost no software was designed to own, because it is the part that historically required judgement.
This is precisely what changed. A system that can read an unstructured document, decide what category it belongs to and take an action inside another system is not a better version of existing automation. It is automation applied to a class of work that could not be automated before.
Which means the question is not "where can we use AI?"
It is: where is a person currently acting as the integration between two systems?
That question produces a far better list. Invoice intake. Support triage. Lead qualification. Contract review. Onboarding paperwork. Quality checks on data that arrives in a predictable shape but never quite the same format.
These are unglamorous. They are also where the hours are.
Design for being wrong
A system that is right 95% of the time is either extremely valuable or actively dangerous, and the difference is entirely in how the remaining 5% is handled.
The systems that survive in production share a few properties. They cite their sources, so a person can verify a claim in one click rather than redoing the work. They expose confidence, and route low-confidence cases to a human instead of guessing. They log what they did, so an error can be traced rather than argued about. And they are scoped to a decision that is reversible, at least until they have earned more.
None of that is exotic. It is the same discipline any other production system gets. It is just easy to skip when the demo was impressive.
Start with one process, in production
The most common failure we see is scope. A company decides to "adopt AI", which becomes a programme, which becomes a committee, which becomes nothing.
The alternative is narrower and much more effective: pick one workflow that a real team performs every week, instrument it so you know what it costs today, build the smallest system that takes over the repetitive part of it, and put it in front of that team.
You will learn more in the first three weeks of that than in three months of evaluation. And if it works, you now have something rare inside a company: a concrete, measured example that the rest of the organisation can argue from.
