Every CEO I meet has bought into AI. The decks are made, the pilots are running, the budgets are signed. And yet, when I ask a simple question, “What has it returned?”, the room goes quiet.

That silence is the most important data point in enterprise AI today. We have spent two years celebrating adoption, and adoption has become a vanity metric. The number of seats licensed, the number of employees “trained”, the number of pilots in flight, none of it tells you whether the technology has changed a single outcome that your business, your customer, or your shareholder can feel.

Let me be blunt about where we are. AI is not a cyclical disruption you can ride out. It is a structural rewrite of how value gets created. And a rewrite that does not show up in the P&L is not a rewrite at all. It is theatre.

Stop Counting Pilots; Start Counting Returns

Here is the uncomfortable arithmetic. The majority of enterprise AI initiatives never escape the pilot. The model works in the lab, the demo dazzles the board, and then nothing moves for the customer, the patient, the citizen, or the operator on the floor. We have industrialised the proof of concept and forgotten that a PoC proves nothing until it is in production, at scale, doing the work.

The problem is not the technology. The problem is that we measure the wrong things. Adoption metrics flatter us because they are easy to hit and easy to report. ROI is harder, slower, and far less comfortable, which is precisely why it is the only metric that matters.

So I would put a different rule in front of every board: if you cannot draw a straight line from an AI investment to a number on your balance sheet within a defined horizon, you do not have an AI strategy. You have a science project. The companies pulling ahead are not the ones with the most pilots. They are the ones who killed the pilots that were not paying and doubled down on the few that were.

What does return actually look like? It is concrete and it is recent:

• A supply chain that responds to disruption 30% faster, measured in days saved, not slides shown.

• A network that delivers higher reliability at lower maintenance cost, with the savings visible on the line.

• A plant where unplanned downtime falls in a way operators feel every single shift.

• A clinical pathway where diagnostic delay drops and clinician capacity rises, not in one flagship hospital, but across the system.

Notice what these have in common. None of them is about how much intelligence you bought. Every one is about how much impact you extracted. That is the shift, from intelligence as capability to intelligence as return. Access to AI is no longer a differentiator; the same models, chips and clouds are available to everyone. The differentiator is whether you can convert that access into value before your competitor does.

Leaders in this era will not be judged by their demos. They will be judged by their diffusion, and diffusion is just another word for returns that compound.

Why You Cannot Do This Alone, And Shouldn’t Try

Here is the second thing boards need to internalize, and it follows directly from the first. Nobody captures ROI from AI in isolation. The returns live at the messy intersection of your domain, your data, your regulatory reality and the engineering required to put a model into live operation. Almost no organisation has all of that under one roof. Which means your ROI is, to a large extent, a function of who you choose to build with.

That makes partnership a board-level decision, not a procurement footnote. And it changes what you should be looking for. For a decade we chose partners by category, is this an “engineering services” firm, a “digital transformation” shop, a hyperscaler, an “AI specialist”? Those labels have stopped being useful. You will never experience a partner as a category. You will experience them as a capability system: can they understand your context, shape the right problem, bring the right mix of domain depth, engineering discipline and AI, and move at the speed your market now demands?

So the test for a partner is the same test you should apply to your own initiatives. Not “What is your AI strategy?” but “What have you built, for someone like us, in the last eighteen months, and what did it return?”

Push past the credentials. The partners earning trust today are not the ones boasting about how many thousands of engineers they have trained or how many models sit in their catalogue. They are the ones who can point to tangible, recent, AI-augmented outcomes and tell you exactly who built them, what broke, how they governed the risk, and how long it took to go from concept to live deployment. Ask for proof less than two years old. Ask who stays on the hook after the PowerPoint ends and reality begins. If a partner leads with tools and buzzwords rather than your constraints, they are optimising for capability, not for your return.

Niche or Scale, But Never the Muddy Middle

Speed is where partnership either pays off or quietly bleeds you. Not reckless speed, these are consequence-heavy industries, but fast enough that value does not get trapped in endless pilots and architecture debates. The partners worth having win in one of two ways. Some win by being so specialized in a narrow domain that they are simply irreplaceable. Others win by combining genuine scale with process-engineering depth, taking an idea from design to deployment across many sites and markets without losing control.

The danger is the middle: not deep enough to be irreplaceable, not disciplined enough to drive change at the speed customers now expect. When you assess a partner, do not just ask what they have done. Ask how fast they did it and how repeatable it is. What does their playbook for diffusion look like, how do they get from the first success to the tenth and the hundredth? If they cannot answer that, they cannot help you scale returns, and a partner who cannot help you scale returns is a cost, not an investment.

The Rubric I Would Put in Front Of Any Board

Strip away the noise and partner choice in the AI era comes down to five questions. I would not sign anything until I had honest answers to all of them:

• Return, Can they tie what they build directly to a number we care about, on a horizon we can hold them to?

• Understanding, Do they prove they understand our world and our definition of “effective” before they talk solutions?

• Proof, Can they show tangible outcomes from the last two years, not slideware and aspiration?

• Speed and diffusion, Do they have a repeatable way to move from pilot to scale without compromising safety or trust?

• Build power, Are their people demonstrably learning and building in our context, or just attending training?

Partners who score high on these turn intelligence into return. The rest will keep talking about AI while your competitors quietly bank the gains.

 

So here is the call to action, and it is uncomfortable on purpose. Stop reporting adoption to your board as if it were progress. Audit your AI portfolio against return, and be ruthless about what you cut. And treat the choice of who you build with as one of the most consequential decisions you will make this decade, because in the end you will not be judged on the intelligence you bought, but on the returns you and your partners had the discipline to deliver.