Where AI is worth using in an established business: a method, not a list.

01What "worth using" means, and why the lists get it wrong.

"Could AI do this" is the wrong question, because the answer is almost always yes. The question that decides whether money is well spent is narrower: does this process carry a cost we can name, and does it run to a standard we can capture? A process with a cost but no standard cannot be automated safely. A process with a standard but no cost is not worth automating. The places worth using AI sit where both are true.

That is why the usual lists are no help. "Summarise documents", "draft emails", "answer customer questions" describe capabilities of a model, not costs in your business. An established business does not have a document-summarising problem. It has a compliance file that takes a supervisor a day, a report a senior clinician rewrites from a junior's draft, a record that has to be found in an audit years later. Those are processes, they have owners, and they have numbers.

The standard is usually in someone's head rather than in a procedure. That is the same fact the key person risk guide starts from, seen from the other side: the person the business cannot lose is often the person whose standard is worth capturing first.

02The four signals that find the places.

Walk the business with these four. Each one points at a process where the standard is real and the cost is counted, and each has a built example behind it.

• The cost of a mistake is real and known.A fine, a rejected claim, a load that cannot leave, a rerun of a day's work. If the people who run the process can tell you what the last mistake cost, the process qualifies. At Xtreme Freight, dangerous goods paperwork took 8 hours per run and a mistake meant a load held or a breach; DGC brought it to under 30 minutes with zero paperwork errors across 120 tonnes in the first pilot.
• Checking is a queue.Work waits for one person to sign it, and that person is the bottleneck for the whole operation. The sign of it is a pile: reports waiting on the senior clinician, files waiting on the compliance lead. At Virtuosa, a clinical report took around 4 hours of a psychologist's time; drafting it to the clinic's standard brought that to about 20 minutes, with the psychologist still signing, and clinics see 25% more clients per psychologist.
• One document only one person can finish.Everyone can start it; one person has to complete it to the standard a regulator, a client or a court will accept. That person's way of finishing it is the standard worth capturing. Handover+ turned a builder's handover into a record that will still answer questions across the ten-year liability window, validated in two weeks and built in five months.
• A decision has to reach someone in seconds.The information exists, the person who needs it is elsewhere, and the gap is where the harm happens. On a mine site that is an emergency; AtSite gets the alert to the right people in under ten seconds, from the same record that produces the compliance report afterwards.

A process that shows two of the four is a strong candidate. One that shows none is a tool purchase, whatever the vendor calls it.

03The five places it usually pays, by industry.

Across the businesses we have worked with, the signals cluster in the same five places. The industry pages go into each; the checklists say what has to be true before a system is trusted with it.

1. Compliance paperwork that has to be right every run: dangerous goods declarations, water-trading approvals, safety documentation. Signal one, and often signal two. See AI in logistics and the run before it leaves.
2. Reports and claims a professional has to sign: clinical reports, assessments, funding claims. Signal two and three. See AI in healthcare and the signed report.
3. Records and evidence that must survive an audit or a claim years later: handover records, inspection evidence, gate logs. Signal three. See AI in construction, AI on mine sites and the ten-year record.
4. Dispatch and the moment: the alert, the roster gap, the Friday night the venue is full. Signal four. See the gate check and AI in hospitality.
5. The customer file: who the regulars are, what they bought, what they are owed, held in a way the business can act on. Signal three in a different form. See Friday night.

If none of the five is yours, the method in section five still applies; the five are where it has paid so far, not a boundary.

04What to leave alone, at least for now.

• Processes that are the same as everyone else's.Payroll, general bookkeeping, generic email. Software already exists; buy it, do not build it.
• Processes nobody has standardised.If three people would finish the job three ways and all three would be accepted, there is no standard to capture yet. Standardise first, then automate. A system built on an argument inherits the argument.
• Customer-facing generation nobody reviews.A chatbot that answers on your behalf with no one checking, in a business where a wrong answer costs trust or a licence. The signals say the opposite: the standard matters most where the mistake is expensive.
• The thing the vendor demonstrated.A demo shows what the model can do with the vendor's data. It says nothing about your cost of mistake or your standard, which is where the result comes from.

05The scorecard: six processes, two scores.

This takes an hour and the people who run the business, not a consultant. Take the six processes that matter most to margin or to liability. For each, score the cost of a mistake from 1 to 3 and the clarity of the standard from 1 to 3, and multiply. The top score is the first candidate for a first stage.

Two rules for reading it. A high cost with a low clarity is not a first project; it is a process to standardise, and the person who should do that is usually obvious from the last column. And a score of 9 held by one person is also a key person risk: the first stage captures the standard whether or not a system is built.

Bring the table to a first call, with whoever hires or fires on the top row. It is the whole brief.

06How the first stage turns a score into a case.

A score says where to look. It does not say whether a system would pay, what it would take, or whether the standard can be held by software at all. That is what a first stage is for, and it should be short.

1. Two weeks with the people who run the process. Not a workshop: the compliance lead, the senior clinician, the supervisor, on real jobs, mapping how the standard is actually applied and where the cost lands.
2. The opportunities ranked, with a commercial case on each. Hours removed, mistakes avoided, work no longer turned away, against what the system would take to build and run.
3. A decision. Go on, standardise first, or buy software instead. All three are good outcomes of a first stage; only a stage that ends without one has failed.
4. Then the build, sized to the case. A prototype is one sprint; larger design and build work typically runs across 5 to 12 sprints, with testing in your conditions before anyone relies on it.

If you stop after the first stage you leave with the ranked list, the cases and the standard written down, which is more than most businesses have before they start. The process page describes the stages; the guide to choosing a consultant gives the eight questions to ask whoever runs it, including us.

• <30minDGC with Xtreme Freight. Compliance paperwork per run, down from 8 hours. Zero paperwork errors across 120 tonnes in the first pilot.Read the case study
• 20minVirtuosa AI. Per clinical report, down from around 4 hours. Clinics see 25% more clients per psychologist.Read the case study
• 10yrHandover+. A builder's handover record kept for the ten-year liability window. Validated in two weeks, built in five months.Read the case study

07Questions CEOs ask before the first call.