The Second Meaning of ROI
Return on Investment is close to a religious text in business. Put capital in. Get productivity, margin, or growth out. Measure the gap.
Investment means money, in one form or another, because money is the scarce thing being allocated. AI is beginning to put pressure on that assumption.
It is giving organizations serious access to another resource that has always had to be rationed: intelligence - the capacity to analyze, interrogate, model, compare, imagine, and challenge. Which raises an obvious and slightly absurd-sounding question:
If intelligence is becoming cheap enough to deploy almost anywhere, does ROI need a second meaning?
Call it Return on Inference.
My Substack - Some Assembly Required - has always been where I work things out in public. It’s where the ideas are longest, sharpest, and least edited for palatability.
This is not my phrase. I first heard Greg Shove use it during his brilliant How to Build an AI Supercompany workshop. I had already completed Section’s AI for Business Mini-MBA—still the greatest return on investment of any continuous-learning platform I have paid for—so I was listening carefully.
Return on Inference. That phrase stayed with me because, beneath the wordplay, it raises a genuinely useful management question.
Not simply: How much AI are we using? but: What are we spending our intelligence on?
Training Builds It. Inference Spends It.
Inference is one of those words that has migrated out of technical conversations without becoming particularly well understood inside business ones. The simple distinction is this:
Training acquires capability. Inference exercises it.
Training is the long, expensive process through which a model learns patterns from enormous amounts of data. Inference happens when the trained model is put to work.
Ask it to interrogate a customer segment, stress-test a strategy, draft a hundred product names, classify an image, model a competitor’s response, or explain why a campaign underperformed, and the model is performing inference.
Training built the capability. Inference spends it against a problem.
That framing is a metaphor, not an accounting standard. Infrastructure teams may use Return on Inference to ask how much value a company receives from its compute spend. I am interested in the organizational question: what value came back from the intelligence we chose to deploy? No single, interchangeable unit of intelligence is being withdrawn from a corporate account. But the metaphor matters because the marginal cost of asking the model to think again, to produce another analysis, alternative, simulation, objection or possibility, is falling extraordinarily quickly.
For most of business history, that was never true.
If you wanted to understand a market properly, you commissioned research and waited six weeks. If you wanted to model a decision, you pulled analysts away from something else. If you wanted to explore what might happen to your category in five years, you convened the kind of workshop that ends with sticky notes covering a wall and a facilitator asking everyone to “park that thought.”
Sophisticated thinking was expensive because people had to do the thinking, and people are finite. That is why organizations rationed it. Not because three scenarios were correct and eleven would have been reckless. Three were what the budget, timetable, and available human attention could support.
That constraint is beginning to dissolve.
Two Rooms
Look closely at how AI is landing inside organizations, and it tends to separate, reliably, into two rooms.
In the first room, AI has been absorbed into the existing shape of the work. Meetings get summarised. Decks get built in a quarter of the time. Reports that once took an analyst two days now take ninety minutes and read better. Customer-service teams handle more tickets. Marketing teams produce more variations. Administrative work quietly disappears.
Someone, usually the CFO, usually with a genuine and defensible sense of achievement, puts a number on the productivity gain and takes it upstairs.
The number is real. Nobody should sneer at it. A four-hour task completed in one hour returns three hours to the organization. Multiply that across a few thousand people, and you have a business case any board can understand.
In the second room, something else is happening. The same tools are being used to interrogate customer behavior continuously rather than quarterly. Teams are generating hundreds of products, pricing or positioning hypotheses and killing most of them cheaply before they reach a meeting. Competitor responses are modeled before a strategic move, not explained after it. Someone goes looking for the anomaly buried in the data that nobody previously had a reason to search for, because searching for it used to cost more than it was worth finding.
One room uses intelligence to make existing work cheaper. The other uses intelligence to make previously uneconomic thinking possible. Both may use the same model. Both may report similar AI spend. Both may appear in the same board pack under the same transformation program, with similarly impressive arrows pointing upwards. But they are producing different returns.
One returns efficiency. The other returns possibility. That is not a difference of degree. It is a difference of kind.
The Question Isn’t the Number
What would it mean to examine a thousand plausible versions of your market instead of three?
For almost the whole of recent business history, the question would have been absurd. It was the sort of provocation somebody used in a workshop to expand everyone’s thinking before returning to the three scenarios there was time to develop.
Put the same request to a well-instructed model today, and the absurdity relocates. A thousand outputs are achievable. The real question is whether you have a thousand things worth asking. And, more uncomfortably, whether you would recognize the useful ones if they arrived.
This relocation is the central AI story.
The constraint moves from production to selection. From: Can we make enough? to: Can we distinguish what matters?
Productivity measures the difference between what the work used to cost and what it costs now. It tells us almost nothing about the value of questions the organization could not previously afford to ask.
Inference Allocation
A CEO’s responsibility is not merely to make sure the company has capital. It is to decide where capital should be deployed. Against which opportunities? At what level of risk? Over what period? What deserves another round? What should stop being funded? Where might a disproportionate return exist precisely because other organizations cannot see it yet?
Capital allocation became one of the defining disciplines of modern management.
AI creates the possibility of a similar discipline around intelligence. Call it inference allocation.
Where should the organization deploy its capacity to analyze, simulate, question and imagine, and against which problems? How much belongs in making existing systems more efficient, and how much in questioning whether those systems should exist at all? Which assumptions deserve to be tested continuously, which decisions deserve more intelligence rather than faster execution, and which parts of the business should receive exploratory capacity even when the return cannot yet be placed into a spreadsheet? And afterward: which of the possibilities it generates are consequential enough to deserve scarce human judgment?
These are allocation questions. Yet most enterprise AI strategies remain acquisition strategies — which models to use, how many licenses to buy, which vendors are approved, how to train employees, what belongs inside the governance framework.
Necessary questions. But temporary ones. Access will not remain the advantage for long. It rarely does with a general-purpose technology. The advantage will come from what an organization decides to do with the access.
The Cheapest Intelligence in History
There is a darker possibility here.
The cheapest intelligence in history could still produce some of the most intellectually passive organizations we have ever seen.
A company can possess enterprise licenses, multiple models, thousands of active users, automated reports, AI-generated ideas, continuous analysis, and a board-approved transformation program — and remain fundamentally incurious. If every model is asked only to accelerate the current workflow, the organization may have abundant inference and no imagination about where to deploy it.
Exploration Isn’t Automatically Virtuous
Exploration sounds like the more sophisticated use of intelligence until you ask how anyone is meant to govern it. Productivity is wonderfully measurable. You can put the hours saved into a spreadsheet and defend them in a budget review. Exploration is not.
What is the value of hypothesis 783 when the first 782 went nowhere?
What is the return on discovering a customer behavior nobody thought to look for, measured against the analyst time required to work through nine hundred plausible observations?
How do you distinguish a useful anomaly from a beautifully articulated hallucination?
An organization that generates infinite cheap thinking without a serious discipline for judging it is not operating at the frontier. It is drowning politely, with excellent production values, in its own inference.
“We are exploring too” can become another kind of corporate theatre: a slide showing a thousand AI-generated ideas that nobody had the judgment, nerve or time to cull.
Cheap thinking without curation is not insight. It is noise with a research methodology attached.
Judgment Was Never the Cheap Part
The constraint has not disappeared. It has moved.
For most of business history, scarcity sat near the point of production: who has the time, who has the analysts, who has the budget to investigate this, who can make the thing. Increasingly, it sits at the point of allocation and selection: which questions are worth funding with inference, which outputs deserve human attention, which patterns are real, which possibility changes the decision, who is prepared to act, who remains accountable.
A prompt tells a system what to do. A question decides what is worth thinking about. Those are not the same capability.
An AI can produce an excellent answer to an irrelevant question. It can optimize a process that should not exist. It can generate a hundred well-argued ideas for a product nobody needs and do it beautifully, on request.
Left unsupervised, it can make an organization spectacularly efficient at doing the wrong thing. Which is considerably worse than being inefficient at it, because at least inefficiency slows you down long enough to notice. Back to that fiction thing.
If inference becomes abundant, curiosity and judgment do not become obsolete. They become the allocation system.
What are we not seeing? What are we treating as a settled fact that is really an assumption nobody has tested since the last downturn? What would we investigate if the investigation were nearly free? What are we spending intelligence on simply because the answer is easy to measure?
Those are not merely prompting skills. They are organizational capabilities closer to editorial judgment, strategy, and leadership than technical fluency. They do not appear prominently on the roadmap of any AI vendor because no vendor is particularly incentivized to tell a customer that the next bottleneck may be their own curiosity.
Return Without a Ratio
Here is what I think Return on Inference does not resolve.
It does not tell us how to price a hypothesis, provide a clean P&L line for curiosity, explain how a CFO should report “questions we finally had the capacity to ask” beside the productivity numbers, or tell us how to compare the inference used to generate marketing variants with the inference used to model a strategic risk.
Somebody may eventually create a credible measurement system. I have not seen one yet.
Return on Investment took decades to become the apparently clean, single-line number organizations now use, sometimes with more confidence than wisdom.
This second meaning, Return on Inference, will not arrive fully formed. But the absence of one perfect ratio does not remove the management responsibility. Return on Inference could include better decisions, faster decisions, more experiments, earlier rejection of weak ideas, improved customer experiences, fewer errors, new products, hidden patterns, increased output — and questions the organization previously could not afford to ask. The return depends on what the intelligence was deployed to do.
Which brings us back to allocation.
The Questions Worth Funding
Return on Investment isn’t going anywhere. Capital matters. It always will. But capital may no longer be the only resource a leadership team is judged on how well it allocates.
The last few years were a race to acquire AI capability: Licences. Models. Governance committees. Training programs. The slide deck nobody finished. That race is nearly over. Winning it won’t distinguish anyone.
The next race isn’t about access. It’s about inference allocation.
Where is intelligence actually being deployed, and against which problems? Is it making what already exists cheaper, or discovering what doesn’t exist yet? Who chose the questions worth funding, who judged what came back, and who’s on the hook when inference turns into action?
Return on Inference isn’t a clean number. It may never be one. But the accountability question has already arrived, whether or not anyone’s named it yet.
The next advantage will not come from having access to intelligence. Access alone will not distinguish anyone for long.
The next advantage will come from knowing what to spend it on.
