Which work should AI actually do, and how far should it be allowed to go?

Board Chair: “What are you doing about AI?”
CTO: “We’re aiming to have 50 use cases in the pipeline by Q4.”

Many AI transformations start with the wrong question: where can we use AI?

A better one: which work should AI actually do – and with how much autonomy?

In our view, two questions settle most of it:

First, can AI learn the pattern?

The sweet spot isn’t highly repetitive work – conventional automation already does that better. It’s work with high surface variation but a set of recurring cases underneath. Think hundreds of differently worded supplier emails that come down to 12 recurring intents.

Second, can we verify the output?

That determines autonomy. If the answer can be checked instantly, AI can act. If an expert can validate it in a minute, AI drafts and the expert approves. If there’s no reliable way to know whether the answer is right, the decision stays human.

That’s why one end-to-end process can need completely different solutions: workflow automation for one task, AI for the next, human judgement for the one after. And it’s why full AI automation is only one of five possible operating models – and rarely the best first move.

Decompose the process to task level. Test variability and verifiability. Then decide how much autonomy the evidence actually supports.

The attached paper sets out the framework, with a worked procurement example. We’ve applied the same approach across other functions – get in touch if you’d like to discuss yours.

The objective isn’t to put AI everywhere. It’s to put it where it actually improves the work.