Pay for the outcome

In plain terms. A saving from AI only stays with you if your pricing lets it. The larger opportunity is usually new revenue, and in a services or information business it tends to be hiding in work you already do, in data you haven't joined up, figures you haven't explained, and records of your own work. Charge for the outcome rather than the hours, and build the option to act.

Why savings don't always stay

Almost every AI plan I'm shown is a cost plan. It lists the work that can be automated, estimates the hours, and puts a number on them. That's the right place to start, and the rest of this site is about doing it well.

McKinsey made the case against stopping there in April. Their analysis is that productivity gains rarely stay with the company that makes them. Competitors buy the same tools, prices come down, and the gain ends up with customers. In their words, productivity gains "typically affect the floor of industry performance, not the ceiling". They put the lasting value in new offerings, business models and market structures.

I agree with most of it. Where I'd add something is on how the saving leaves. It goes to the customer when you charge for the activity, because a cheaper activity is a lower bill. When you charge for the outcome, you keep it for longer, and you can spend it on the things below, which is why pricing comes first.

Three places I've found revenue

None of these started as an AI project. Each started with a customer problem and a piece of work the business was already doing.

The picture nobody could see

At a global information business, we supported oil and gas traders trying to understand commodity prices. They bought data from several vendors, and each set told part of the story. The sets couldn't be lined up, because they carried no timestamp. So every trader had a partial view, however many sources they paid for.

We found a way to fuse them, then used pattern analysis and anomaly detection (machine learning) to look across the combined picture. That gave us a view no single vendor could sell, and it let us build predictive models that hadn't been possible before.

Most businesses have a version of this. They hold data from customers, suppliers and their own operations that has never been joined up, because nobody's job was to join it. Research on Chinese listed companies (Shi, Wu, Qin and Liu, 2025, International Review of Financial Analysis) found that AI adoption raised firm value, and that the firm's data assets carried about 45 per cent of the effect. It's one study in one market, but it points the same way.

Numbers into words

At the same business, a lot of what we sold existed as numbers in a spreadsheet or a database. We used natural-language generation to turn those numbers into written commentary, so a customer got the figures with the context and meaning around them. Those reports were sold to customers.

That used to be hard. It's now the easiest thing in this piece, because any language model can write fluent commentary. What's still scarce is knowing what's worth saying: which change matters, which is noise, and what a customer should do about it. That judgement is the product now.

The switch you can flick

In a PE-backed talent business, we built what was in effect an operating system for the whole organisation. Every activity was recorded in an audit trail, so we could see how work actually flowed. That record became a feedback loop. It showed us which tasks could be automated, and whether each one suited simple robotic process automation or needed an AI agent.

We didn't automate everything it found. Sometimes the timing was wrong for the business. Sometimes the decision was better left for a future owner, as part of their plan. The value was in having the switch built and knowing what flicking it would do. Having the switch is what matters.

Why I price the outcome

I learned consultative selling and value-based pricing early, and I've held to one rule since: customers should pay for the outcome, not the activity. AI makes that more important, because it makes many activities cheaper.

Take a risk intelligence platform that tells a company whether there's trouble in its supply chain. You could sell it by subscription, or charge per search, or per report. But the same report is worth very different amounts to different customers. For a company checking a critical partner, where a missed risk would be expensive, it's worth a great deal. For one looking up a small supplier that hasn't filed accounts lately, much less. Price per report and you charge both the same. And as AI takes the cost of producing that report towards nothing, the price will follow it down.

Pricing for the outcome is harder. It takes a real conversation with the customer about what they're trying to avoid or achieve. It's also the most reliable way I know to keep the saving from AI in your own margin.

What makes it last

McKinsey expects advantage to move to proprietary data, to being built into how customers work, and to learning faster than competitors. In a services or information business, the part you can own is a connected picture of your customers and work, a written record of what good looks like that software can use, and your own test cases, so you can prove the product works as the models change underneath it. That's the layer you own, and it's what makes the product sellable beyond the pilot.

Where I'd start

Four questions, for a leadership team or a fund looking at one of its businesses:

1. What do we see, across our customers or suppliers, that none of them can see on their own?
2. What do we produce as numbers that someone would pay to have explained?
3. What do we already record about our own work that would show us what to automate, and when?
4. What do customers value that we currently charge for by the hour, the search or the report?

Most of the time the answer to the first question is the most valuable, and the fourth decides whether you keep any of it.

What this rests on

The case that productivity gains are competed away and that value lies in new offerings is McKinsey's: Montard, Diedrich and Catlin, "Where AI will create value, and where it won't", McKinsey Quarterly, 29 April 2026. The finding on data assets is from Shi, Wu, Qin and Liu, "The value-creating potential of AI: a multi-dimensional analysis of effects and mechanisms", International Review of Financial Analysis, vol. 108, part B, December 2025, which studies Chinese listed companies between 2010 and 2022. The three examples are from work I led, described without names or figures. The view on pricing is mine.