Four ways to build or buy AI

“Should we build AI or buy it?” is the wrong question, because it hides two better options. Any workflow has four moves: build, buy, change the process or hire a person. The right one varies workflow by workflow, and the cheapest is often the one without AI.

Start with the workflow, not the technology

Most build-or-buy debates start from a tool someone saw. Reverse it. List the recurring work that expensive people do weekly in roughly the same shape. Estimate the hours each takes and rank the list by them.

For each workflow on the list, one of four moves applies. Take them in cost order, cheapest first.

Option 1: change the process

The cheapest move, and the most overlooked. Much of what looks like an automation problem is a process problem. The same data typed into two systems, a queue nobody owns, an approval that adds a day and catches nothing.

AI layered on a broken process automates the mess. It does the wrong thing faster. Fix the process first, and two things happen: some of the AI case disappears, and what remains gets sharper. A clean process is also far easier to automate later.

Option 2: buy an off-the-shelf tool

Buy when your problem is common. Invoice capture, meeting notes, email triage and document search are categories vendors already solved. Vendors have spent years on the edge cases you have not met yet. Buying a solved problem beats rebuilding it, every time.

The buy test has three parts:

• —Is the problem common? If other businesses your size have the same workflow in the same shape, someone sells the answer.
• —Is the category mature? Several vendors, stable pricing, and customers who have renewed. A category of demos is not a category yet.
• —Can you leave? Check what happens to your data when you cancel. A tool you cannot exit is a tool you did not buy; it bought you.

Option 3: hire a person

Unfashionable in an AI article, and often right. Hire when the work is judgement, relationships, or accountability. A person can be responsible for an outcome; software cannot. If the value of the work grows with the person doing it, a machine caps that value.

You rarely need to choose between the person and the machine. The strongest version is a hire whose admin you automate. The hours you pay for then go on the part only a person can do.

Option 4: build

Build when the workflow is particular to you and an off-the-shelf shape would break what makes it work. That happens. A firm’s intake process, house document style or pricing method can be the business itself.

Go in with eyes open, because building is a commitment. Someone must own the system, maintain it and fix it when a supplier changes an API. Before building, ask who that someone is. If the answer is nobody, buy, or scope the build so breakage is an inconvenience and never an outage.

A test you can run this week

Pick your three most time-hungry workflows and score each on three axes:

• —Value — hours saved or capacity gained if this improved. Be concrete: whose hours, doing what instead?
• —Effort — what it takes to change, including the disruption while people adjust, not only the invoice.
• —Reversibility — how cheaply you can undo the change if it disappoints. Favour reversible moves early. A pilot you can abandon teaches more than a commitment you must defend.

High value, low effort, easily reversed: do it now, whichever of the four moves it is. High value and high effort is where outside advice earns its fee, because a wrong call costs you in both directions.

Getting it decided

You can run this framework yourself, and for many businesses that is enough. The hard part is the middle, where workflows are partly judgement and partly repetition. There the answer is a split, and that allocation with reasoning per line is what a structured audit produces. We have written up what a week-long AI audit involves if you want to see the shape of one. And if the question is what advice like that should cost, see our guide to AI consultancy pricing in the UK.