Across a portfolio

In plain terms. Don't run twenty separate AI programmes, and don't run one big one either. Do a few things once, properly, for everyone: how value is counted, the rules for letting software act, and the contracts with vendors. Leave the work itself with each business, because that's where their judgement lives.

The question changes when you're at the centre

Inside one business, the question is where to point AI. At the centre of a fund or a group, it's a different question: where the value is likely to be, how you would know, and what should be done once rather than twenty times.

There are two easy mistakes. The first is a central programme that picks the tools for everyone. It's simple to buy and it fits nobody's work. The second is leaving every business to get on with it. You end up with twenty pilots, twenty vendor contracts and no way to compare one result with another.

I've seen both from the inside, in businesses owned by funds and in groups deciding what to buy. The approach that works sits between them.

Share a little, and share it properly

A few things are the same in every business and are worth doing once, at the centre:

• How value is counted. The same two columns everywhere: what AI is claimed to save, and what actually reached the accounts. Same rules in every business, so the numbers can be added up and compared.
• The rules for letting software act. One register format, one set of grades, and the same rules for when a piece of AI earns more freedom or loses it.
• How to build a test set. One method for everyone, while each business still picks its own real cases.
• Buying. Model and vendor contracts negotiated once, with the right to switch written in.

What stays with each business is the part that matters most: which work to point AI at, the cases that show what good looks like, and the decisions about their people. That's where each business's judgement lives, and it doesn't travel well.

Ask every business the same nine questions

A good first step is to put the five-minute check to each leadership team. It takes an afternoon across a whole portfolio. You'll quickly see which businesses are pointing AI at real work, which can show savings in their accounts, and which have software doing more than the evidence supports.

It's a first look, not a full review. But it tells you where to spend your time first, and because every business answered the same questions, the answers can be compared.

Count it the way a buyer will

Sooner or later, a fund or a group will be asked what AI actually added. A buyer will discount whatever was claimed and pay for what was banked.

So add up the banked column across the businesses, and always show both columns side by side, never a single percentage. The register and the test sets count too. Evidence that software acting in each business's name is under control is part of what a buyer is paying for.

Watch the people question across the whole portfolio

If every business automates its junior work in the same year, the whole portfolio gets the saving at the same time. It also gets the shortage of experienced people at the same time, a few years later, possibly under the next owner.

The answer is the same as in one business: decide on purpose which work people keep because it teaches them. From the centre, the useful thing is to ask each business to show you that decision, not to make it for them. The curriculum exercise takes an afternoon per team.

One owner at the centre, one in each business

The centre stays small. In my experience two permanent people can be enough. Each business has its own champion, chosen by that business's leader rather than by the centre. The centre owns the shared rules and the counting. Each business owns its work and its results.

Where to start

Pick two businesses, not twenty: one that's keen and one that's sceptical. Run the four questions on one workflow in each, keep the two columns from the first day, and share what you learn with the rest before anyone scales anything. The first ninety days for a larger business are set out in How I run it.