Buy, build or refuse: an automation decision that survives the demo

01

Refuse is the default, not the failure case

The industry evidence on this is not subtle. Gartner predicted in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, and the three reasons it gave were escalating costs, unclear business value and inadequate risk controls. Not one of those is a technology problem. They are all operating problems, decided before a line of code ran.

MIT Media Lab's 2025 report, The GenAI Divide, was blunter still: 95% of generative AI pilots it studied produced no measurable effect on profit and loss. Not that they broke. That they worked and changed nothing.

Gartner also made a point worth keeping in your pocket for the sales call. It estimated that of the thousands of suppliers claiming agentic capability, only around 130 were genuinely doing it, and named the rest as agent washing: existing chatbots and workflow tools relabelled. So the first thing you are evaluating is not the product. It is whether the category description is true.

02

The four gates

This is the sheet I would put in front of anyone facing an automation decision on Monday morning. Four gates. A failure at any one of them means refuse, and refuse means fix the operating condition first, then reopen the question next quarter.

Gate one. The process is written down as it actually runs, not as the manual says it runs. Sit with the person who does it and record every exception they handle without thinking. Those exceptions are the automation. The happy path is the easy 20%.

Gate two. The process is measured today. If you cannot say what the current call abandon rate, no-show rate or turnaround time is, you will never be able to prove the tool helped, and you will end up defending a spend with anecdotes. Measure for four weeks before you buy anything.

Gate three. One person owns the result by name. Not a committee, not the vendor, not the person who ran the pilot and has since moved teams. Somebody whose week gets worse when the tool misbehaves.

Gate four. You can survive the supplier disappearing. Can you export your data in a form you could actually use. What is the notice period. What happens if they are acquired next year and the pricing doubles. Write the exit before you write the purchase order.

An afternoon gets you through all four. It is the cheapest afternoon in the whole project.

03

Buy the things every clinic needs

Once the gates are clear, the buy or build split is much less interesting than people make it.

Buy anything that every clinic on earth needs and none of them wins on: reminders, telephony, payments, rostering, e-signature, document storage. You will never be better at sending an SMS than a company whose entire existence is sending SMS. Building it in house means you now maintain it forever, and the person who wrote it will leave.

The MIT report found the same thing from a different direction. Systems bought from specialist suppliers reached real deployment roughly twice as often as systems built internally. Buying is not the timid option. It is usually the correct one.

The discipline in buying is not choosing well. It is choosing narrowly. One tool, one job, one owner, one measured number it is supposed to move. Bundles are how you end up paying for eleven modules to use two.

04

Build only what you would defend in a board meeting

Build when the process is the thing that makes your operation different from the clinic across the road, and when you already run it well manually. That last clause is the one people skip. Automating a process you have never made work by hand does not fix it. It industrialises the mess and adds a licence fee.

In practice that narrows the build list to almost nothing: your own quality scoring rules, your own escalation logic, the specific way you route a nervous first-time enquiry. Those are worth owning because nobody else can sell them to you.

There is also a consent question that operators treat as soft and clinicians do not. The American Medical Association's 2026 Physician Survey on Augmented Intelligence, covering nearly 1,700 doctors, found 81% using AI professionally, up from 38% in 2023. The number I find more useful is that 85% said they wanted to be consulted or directly involved in decisions about AI adoption. If the clinicians who have to live with the tool were not in the room when it was chosen, you have not bought automation. You have bought an argument.

05

What the demo is built to hide

A demo is a sales asset, and a good one is honest about the product and silent about the conditions. It will not show you what happens on the fourth exception in a row, at handover, with a partial record, in the second language your front desk actually uses.

So ask your own. Make them run your worst week, not their best case. Give them twenty real anonymised records including the three that break every system you have ever used. Ask what the tool does when it is unsure, and whether it says so or guesses. Ask who gets paged. Then ask for a customer of your size, in your market, who left them, and why.

The suppliers worth buying from answer all of that without flinching. The rest suddenly want to schedule a follow-up.

The demo is not lying to you. It is answering a question you did not ask.

06

The verdict

You do not control the supplier market. You do not control what they charge next year, which of them still exists in three years, what a regulator decides about the whole category, or whether the model behind the product gets better or quietly worse after an update you did not ask for. Treat all of that as weather. Plan for it, do not argue with it.

What you do control is the entire part that determines the outcome. Whether the process is written down as it truly runs. Whether it is measured before anyone spends money. Whose name is against the failure. What you agree to sign and how fast you can walk away. Whether the people who use it were asked.