The AI ROI Map: Where AI Actually Pays Across 127 Business Operations (and Where It Doesn't)

Ask a CFO how much their company spent on AI last year and the number comes back to the decimal. Ask what it returned and the room goes quiet.

That gap is the defining problem of enterprise AI ROI right now. In April 2026 Gartner surveyed 782 infrastructure and operations leaders and found that only 28% of their AI projects delivered the return they were promised, with one in five failing outright (Gartner, via The Register, 2026). The spending has not slowed to match. Almost every enterprise now runs a discipline just to track its AI and token spend, up from roughly a third two years ago, and the skill those teams most want to add is measuring the value they get back (FinOps Foundation, State of FinOps 2026).

Of course numbers like that get read as proof that AI does not work. In my view they point at something narrower and far more fixable. The money went in before anyone answered the question that decides the return: where does AI actually pay, and where does it quietly not? This is a map of that answer, across 127 business operations. Call it the AI ROI Map.

Why the AI ROI numbers you are handed cannot be trusted

Return on investment did not change because the investment is artificial intelligence. Gain minus cost, over cost, a number a finance team can check. What changed is that every part of that sum got harder to pin down, and that most of the numbers in circulation come from someone with a reason to inflate them. Measuring AI ROI honestly comes down to three things, the return, the cost, and whether it was realized, and AI strains all three.

The benefit side is soft. Most "AI ROI" claims are not booked financial return, they are proxies: hours saved, tickets deflected, a productivity lift from a pilot. Useful signals, of course, but a proxy presented as a return is how a program convinces itself it is winning before the P&L agrees. The cost side keeps moving, because consumption pricing runs the meter every time the model does, which is why finance teams are suddenly tracking tokens the way they once learned to track cloud. And the return rarely realizes, because almost nobody captured the baseline before the spend. One 2026 survey put the share of CEOs seeing no return at all from AI so far at 56% (Forbes, 2026).

Underneath all of it sits the simplest problem. Most AI ROI figures a buyer sees are supplied by the party selling the AI, and a vendor's job is to make the case for spending, in every operation, all the time. The claims are rarely audited, and almost never graded for how much evidence stands behind them. Anchoring an AI investment on a vendor's own number is like pricing a house on the seller's word.

What we built, and why you can trust it

We built the AI ROI Map to answer the question with numbers that do not come from anyone selling AI. It is an assessment of 127 business operations across 11 business areas, triangulated from 245+ quantified data points across more than 120 independent research organizations: consulting studies, analyst reports, academic work, and vendor benchmarks with the vendor bias discounted where it could be. No single source sets a figure. Every number in the tables that follow comes from that triangulation, which is the whole point of the exercise, an AI decision anchored on evidence rather than on the pitch. The full source list and grading rules live in the methodology.

Two things get measured, and they are worth keeping apart. The AI value of an operation is the share of its work that AI can realistically take on given today's technology and evidence. That is the percentage in the tables, drawn from the research. The AI mix is which of the six AI levers does the most work in that operation, a directional read of where the leading lever sits, honest about being a pattern rather than a split measured to the decimal.

Then the part most maps skip. Every AI value estimate carries a confidence grade, Strong through None, set by how much independent published evidence actually stands behind it. We grade our own numbers, and the grades are humbling. Of 127 operations, only six carry Strong or Solid evidence today.

Where the proof is thin, the map says thin. That honesty is the point. A number you can plan against tells you how sure it is. A pitch deck never does.

One thing to hold in mind while reading any of it. These are figures on a moving curve, not verdicts carved in stone. AI value climbs as adoption deepens and the technology matures, so an operation sitting at 40% today can sit higher a year from now. Confidence moves the same way: a grade of Thin reflects the evidence that exists right now, and thin becomes Solid as the studies land. The map is a snapshot of a continuum, re-read as both the adoption and the evidence move. Read every number as where things stand today, not where they settle.

Most AI budgets fund the wrong category

Start with the single most expensive misread in enterprise AI. Generative AI takes an estimated 60 to 70% of AI budget. Assessed across the 127 operations, it accounts for an estimated 18% of the aggregate AI value, and it leads in only 15 of 127 operations. The quiet workhorse is machine learning: 32% of the value, leading in 65 of 127.

The pattern has little to do with technology and everything to do with attention. Companies fund the category that is loudest in the market, not the one that does the most work in their operations. Kahneman, in Thinking Fast and Slow, called this the availability bias, the tendency to weigh what comes easily to mind over what the evidence supports. Enterprise AI budgeting is availability bias with a purchase order attached, the demo that impressed the board becoming the line item while the quiet lever that moves the P&L stays under-funded. The same reflex shows up one level down in the token bills, in the habit of running every task on the most powerful and most expensive model whether it needs it or not. Paying for the loudest option instead of the one the work calls for is the whole mistake, at the category level and the model level both.

The distance between the budget column and the value column is where most enterprise AI spend quietly disappears. Not because generative AI has no value, but because it gets funded as if it were the whole answer when it is one lever of six.

Where AI ROI is highest

The highest-return operations share a shape. High volume, rule-bound or document-driven work, where a model runs the task at scale and a person checks the exceptions. Average AI value across all 127 operations is 49.4%. These clear 55 to 85%.

Read the leading-lever column, because it carries the lesson. The best AI ROI rarely comes from the loudest category. Reconciliation and matching pay off on machine learning, contract work on generative AI, defect detection on computer vision. The right lever is set by the nature of the work, not by the vendor cycle.

Where AI ROI is lowest

The lowest-return operations share the opposite shape. Judgment-heavy, lower-volume, or accountability-bound work where a person stays in the loop by design. These settle around 25 to 45%. Real value still, but the place an AI-first budget is most likely to be wasted.

No vendor will publish this half of the map, because a vendor's job is to sell AI into every operation. That is precisely why it is worth publishing. The ceiling tells you where to move first. The floor tells you where the money is lost, and a real AI ROI decision needs both. Putting a generative-AI budget into enterprise risk assessment is one of the faster routes to a stalled project and a bad ROI story, and it is exactly the kind of spend the map is built to stop.

The floor is a read on today, though, not a life sentence. Some of these operations will climb as the technology matures and the evidence catches up. The discipline is to fund them when that happens, not before, which is the difference between timing an investment and hoping for one.

How to read the map for your own operations

This is the aggregate picture, and there are three ways to go deeper without waiting on us. Browse the full operations map to see AI ROI by operation across all 127 and sort the ceiling from the floor. Open a single operation profile for the detailed read on one process, its mix, its evidence, and the tools that serve it. Or run your own numbers through the AI Value Assessment, which applies the same 127-operation model to your volume, your systems, and your risk tolerance, and returns your version of the map: AI ROI measurement at the operation level, where AI pays for you, in what order, on which lever, and what evidence stands behind each call. All of it is re-read as adoption and evidence move, so it is worth checking back as the curve shifts.

The discipline is easy to state and hard to practice. Start at the ceiling, not at the loudest category. Fund the levers the work rewards, not the ones that demo well. And treat any AI ROI number that will not show you its evidence the way you would treat any other number a salesperson made up on the spot.

The companies that will look smart about AI in three years are not the ones that bought the most of it. They are the ones that knew, operation by operation, where it paid. So before the next AI line item goes in, the question worth asking is a plain one. On this operation, what comes back, on which lever, and what evidence stands behind the call?

Proprietary data: the BizBlocz AI Value set, 127 subprocesses across 11 business areas, 245+ quantified data points from 120+ independent research organizations, each graded for evidence confidence. External research cited inline. Related reading: The AI Solution Mix, how the six categories work together.