Nobody has ever asked for the ROI on a stapler. Nobody built a business case for the office Wi-Fi, or for email, or for the lights in the parking garage. These things are overhead. You buy them, you keep the cost sane, and you stop thinking about them. On November 4, 2026, diginomica wrote up McKinsey's latest Global Survey on the state of AI, and buried in it was a sentence from McKinsey Senior Fellow Michael Chui that should change how every executive committee handles measuring ROI on AI initiatives: "Spend on horizontal AI-enabled tools, such as chatbots, is increasingly managed as a necessary cost of doing business."
Read that slowly. The biggest consultancy in the business is reporting that companies have quietly stopped asking a large slice of their AI to pay for itself. Some of it became a stapler.
I think this is the most useful thing to happen to AI return on investment all year. It also shows why most AI ROI reporting is broken. Companies have been running one measurement system across two very different kinds of spending. Half their AI is a utility. The other half is a wager. They put both on the same spreadsheet, ask both for a payback period, and then wonder why the numbers come out as fog.
What the autumn surveys actually say
A pile of surveys landed in the last few weeks, and they rhyme.
McKinsey's survey, as reported by diginomica's Katy Ring, found that 80% of respondents say AI has improved their productivity, and half say it helps them make better decisions. Then the floor drops out. Only 37% attribute any EBIT impact to AI, roughly where the number sat last year. The "high performers", companies that attribute at least 5% of EBIT to AI with a significant impact, are still about 6% of the sample. McKinsey Senior Partner Dan Tinkoff named it plainly: "A striking finding this year is the gap between individual gains and enterprise impact."
Deloitte published its State of AI in the Enterprise report on October 29, 2026, built on 3,500 director-to-C-suite leaders across 25 countries, with fieldwork in May and June. Of those companies, 81% report faster decision-making and execution. Only 28% report revenue gains from AI, though that is up from 20% in Deloitte's January report. Deloitte's US head of AI, Jim Rowan, put the gap in one line: "The next challenge is safely turning speed into meaningful change."
On November 12, 2026, CFO Dive's Alexei Alexis pulled together two more data sets. A Genpact survey of 2,005 enterprise executives found that, on average, 49% of AI initiatives funded over the previous two years delivered their expected value, and only 15% of organizations had fully tied AI investments to specific profit-and-loss outcomes. A Capgemini Research Institute survey of 500 senior executives at companies with more than $1 billion in revenue found that about 60% said AI spending over the prior 12 months ran well above expectations or budget. Among the overspenders, the average overrun was 24%, and 21% blew past budget by more than 30%. Two-thirds of the overspenders reported consequences such as lower free cash flow.
Marketing teams are getting squeezed from the same side. Validity's Winter 2026-2027 Marketer Survey, released November 11, found 61% of marketers facing C-suite pressure to demonstrate measurable AI gains. Half said their 2026 AI investments met or beat ROI expectations. A quarter of respondents, 24%, could not yet say whether their investments had paid off at all.
Put those side by side and you get a strange picture. Individuals feel faster. Executives feel faster. Costs run over. And the P&L mostly shrugs. The usual reading is that AI is overhyped, or that adoption is immature, or that companies need better change management. I read it differently. The measurement is asking the wrong question of most of the spend.
Measuring ROI on AI initiatives starts with sorting
Every dollar of AI spending belongs in one of two ledgers, and the first job of an AI ROI framework is deciding which.
The first ledger is overhead. This is the chat assistant on every laptop, the meeting summarizer, the coding copilot that every engineer now expects, the writing help baked into the office suite. Its value is real and spread thin across thousands of small moments. It shows up as Deloitte's 81% feeling faster and McKinsey's 80% feeling more productive. Trying to trace it to EBIT is like trying to trace the revenue contribution of the fact that your staff can read.
The second ledger is bets. A bet is an AI system aimed at a specific line in the P&L, with a specific owner, where the business would look measurably different if it worked. An agent that clears supplier invoices end to end. A claims triage system that changes how many adjusters you need next year. A pricing model that moves gross margin. These are the projects that should produce the EBIT impact McKinsey's 37% are reporting, and they are the only place where a classic ROI calculation makes sense.
Most companies never make this split. They run every AI line item through the same intake form, with the same request for "expected annual savings", and the result is predictable. The overhead tools get justified with invented hours-saved numbers that nobody can ever collect. The bets get diluted, because they share a budget and a governance process with fifty copilot licenses, and nobody gives them the operational redesign they need to hit a real number.
Chui's sentence about necessary costs describes companies sorting by accident. I want you to do it on purpose.
The overhead ledger
Overhead AI gets three questions, and none of them is "what is the return?"
What does it cost per unit of use? Deloitte found that 71% of companies track total AI inference and API spend, 51% track token volume, and only 44% track per-unit inference cost. Costi Perricos, Deloitte's Global GenAI Business leader, said most firms know their spend but "far fewer understand the unit cost underneath it." If you treat a tool as a utility, you have to manage it like one, and every utility manager knows the cost per kilowatt-hour. For a chat assistant the unit might be cost per active user per month. For a coding agent it might be cost per merged pull request.
What is the ceiling? Utilities have budgets with hard caps and alerts. The Capgemini data is a picture of what happens when AI overhead has no ceiling: costs drift up with usage, nobody owns the total, and the overrun shows up in free cash flow before it shows up in a meeting. According to CFO Dive's summary, six in 10 respondents said their organizations lack a comprehensive view of AI expenses, and half said AI spending was approved without adequate visibility into full costs.
What would it cost to not have it? This is the only value question worth asking of overhead, and it is a recruiting and retention question more than a finance one. Suppose your engineers can get a strong coding agent at every competitor and cannot get one at your company. Your hiring funnel will tell you the cost long before your EBIT does.
Overhead AI fails when it gets expensive and nobody notices. It does not fail by missing a return target, because it never had one. Freeing these tools from fake ROI cases is a relief to everyone, including the finance team that has been pretending to believe the hours-saved numbers.
The bet ledger
Bets get the opposite treatment. Every bet names the P&L line it is supposed to move, the person who owns that line, the date by which it should move, and the number it should move to. That is a short list, and the Genpact result shows how rarely it gets written down. If only 15% of organizations have fully tied AI investments to specific profit-and-loss outcomes, then most AI bets in the market today are bets without a stated payout.
The bet ledger also needs a kill rule. Capgemini found that only 5% of overspending organizations paused or stopped an affected initiative, while 53% tightened controls or cut scope. Trimming scope is what you do when nobody agreed in advance on what failure looks like. A bet with a named target and a named date can lose cleanly. A bet without one just shrinks slowly until it becomes a line item nobody wants to defend or cancel.
Here is the test I use with clients. Say the project works exactly as promised. Which number in next year's budget changes, and who signs off on the change? If nobody can answer, it belongs on the overhead ledger, with a unit cost and a cap, or it should not be funded.
Hours saved are a receipt you cannot cash
The single most popular AI ROI metric is time saved. It is the one vendors put on slides and the one employees can report without effort. In Validity's survey, among marketers reporting some return, 35% pointed to time saved. Cost savings came first at 41%, and only 26% measured by comparing revenue to AI cost.
Look at what the executives above those marketers say they care about. Asked what the C-suite prioritizes, Validity's respondents put marketing ROI and return on ad spend at 25%, revenue and pipeline contribution at 19%, customer acquisition cost at 16%, and efficiency gains at 6%. The people reporting results are speaking in hours. The people reading the reports are listening for dollars. Cynthia Price, Validity's SVP of marketing, said it gently: "Saving time matters, but it is only part of the equation."
Time saved belongs in the overhead ledger as a satisfaction signal. It cannot carry a bet, because a saved hour does nothing to the P&L until someone decides what happens to it. Suppose a finance team of twenty each saves four hours a week with an AI assistant. That is eighty hours a week of capacity, roughly two people's worth. If the company keeps all twenty people, gives them the same mandate, and lets the hours dissolve into slightly longer lunches and slightly better slide decks, the AI ROI is zero on paper and arguably zero in fact. If the company redeploys the capacity to collections work it had been ignoring, the return becomes a number on the cash flow statement.
That decision is a management act. The AI cannot make it for you, and no survey tool can measure it after the fact unless someone decided it in advance. I suspect this explains most of the gap between McKinsey's 80% and its 37%. The individual gain is real. The harvest never happened, because nobody was assigned to harvest.
Deloitte's numbers point the same way from a different angle. Sixty percent of companies say they have redesigned jobs around AI capabilities to at least a moderate degree, but only 15% report a significant shift in decision-making authority. Jobs got new tools. Power and accountability mostly stayed where they were. Capacity freed inside an unchanged structure leaks out of it.
The cost side of the fraction moves
ROI is a fraction, and most AI business cases treat the bottom half as fixed. It stopped being fixed when the bill went from per seat to per token.
A seat license costs the same whether the user asks one question or one thousand. Inference does not. When an agent loops, retries, or pulls a longer context, the cost moves with the work. McKinsey found that about 20% of respondents say AI operating costs, including token costs, have constrained their use. Capgemini found that only 56% of executives were fully confident they could forecast AI costs as adoption grows, and only a quarter of CFOs were fully confident they could estimate indirect costs such as human oversight and governance.
That last point matters more than it looks. Every serious AI bet carries a human tax: reviewers checking outputs, engineers maintaining evaluations, compliance staff reading logs. Deloitte found that 40% of companies allow AI systems to execute decisions or take actions without human approval, at least to a moderate degree, and 23% say accountability is unclear when an AI system produces a negative outcome. The cheaper you make oversight, the more risk you carry. The more oversight you add, the more the denominator grows. A business case that ignores this trade is a guess with a spreadsheet around it.
Volker Darius, CEO of Capgemini Invent, told CFO Dive that organizations should "stop treating AI costs as a technology or procurement issue" and that they "need to measure cost, quality and business value together." I agree, and I would add a sequencing rule. Measure cost first, per unit, on everything. You can know your cost per invoice processed or per support ticket closed within weeks of launch. Value takes longer to show. If you cannot see the unit cost, the value number you produce later will be wrong anyway, because you will not know what you paid for it.
The CFO is moving into this seat whether IT likes it or not. Capgemini found that CFOs take part in approving major AI investments at 52% of organizations, up from 42% a year earlier. That is the right direction. Finance knows how to run two ledgers. It has been separating operating overhead from capital projects for a century.
The build-or-buy line is a bet in disguise
One more figure from the McKinsey write-up deserves its own section. Nearly a third of respondents say their organizations decided against buying one or more software products or features because agentic coding tools could build them internally. McKinsey Senior Partner Lieven Van der Veken said: "Leaders are asking what their organizations need to build AI tools themselves."
This is a return on investment that almost nobody is booking. When a team declines a software contract because a coding agent can produce the internal equivalent, the company avoids a cost. That avoided cost is real money, and it rarely appears in any AI ROI report, because the AI budget pays for the coding agent while the savings land in a different cost center that never saw the vendor quote.
It also carries a cost that almost nobody books. Software you build is software you maintain. Suppose the internal tool replaces a $60,000 annual license and takes one engineer a fifth of their time to keep alive. Depending on salary, that is either a fine trade or a wash, and you only find out which if someone writes both numbers down at the moment of the decision.
Every "we'll build it" call is a bet in the sense I defined above. It should get a named owner, a target, and a review date. Otherwise the company ends up with dozens of small internal tools, each justified once in a hallway conversation, none of them measured, all of them consuming engineering time that used to go somewhere else.
What the high performers seem to share
McKinsey's high performers are a small group, still about 6%, and the diginomica write-up does not give a full recipe for what they do. I will not invent one. What the surveys do show is a pattern among the numbers that moved.
Deloitte's revenue-gain figure climbed from 20% to 28% in under a year. McKinsey found that 40% of large enterprises (over $1 billion in revenue) report scaling AI agents, up from 27% last year, while smaller organizations held flat at 22%. The growth is concentrated where companies have the scale to redesign whole processes rather than sprinkle tools across them. Agents that own a workflow end to end are bets almost by definition. They replace a measurable process, so they produce measurable results, good or bad.
Contrast that with the horizontal tools Chui described. They spread across every function and own no workflow, so they cannot move a single line. Both kinds of AI are worth having. Only one kind can tell you what it earned.
My reading is that the companies seeing EBIT impact are the ones who stopped asking their copilots to explain the P&L and started asking a small number of specific systems to change it. They concentrate measurement where measurement can work. Everywhere else, they manage cost.
How to run the split this quarter
None of this requires a new platform. It requires a meeting, a spreadsheet, and a willingness to reclassify things.
Start with an inventory of every AI line item, including the ones buried in software renewals and the ones that individual teams expensed on a corporate card. Capgemini's finding that six in 10 organizations lack a comprehensive view of AI expenses suggests your list will be longer than you expect.
Sort each item. If it is horizontal, spread across many people, and aimed at no specific P&L line, it goes on the overhead ledger. Give it a unit cost metric, a monthly cap, and an owner who watches both. Retire its ROI case. You will be surprised how much management time this frees.
For everything left, write the bet card. Name the line it moves, the owner of that line, the target, the date, and the kill rule. Include the human oversight cost in the denominator. If the card cannot be completed, the project either moves to overhead or stops.
Then review the bets on a fixed cadence, with finance in the room. Say a bet misses its first checkpoint by a wide margin. The conversation should be short, because the rule was agreed in advance. That short conversation is what the 5% of Capgemini respondents who actually paused failing initiatives were able to have, and what the other 95% could not.
Finally, assign the harvest. For every bet that frees capacity, someone has to decide, in writing, where that capacity goes. Redeployment, slower hiring, new work you had been turning away. Without that decision the gain stays individual and never becomes enterprise, and next year's survey will show you the same gap Dan Tinkoff called striking.
Why this needs architecture
The tempting response to a season of disappointing AI ROI surveys is to buy something: an AI cost dashboard, a value-tracking module, another vendor promising to quantify productivity. Those tools can help. They cannot do the sorting, and the sorting is the whole job. Deciding which of your AI systems are overhead and which are bets is a decision about your operating model. It determines who owns what, which processes get redesigned, where freed capacity goes, and which budgets the CFO controls. A dashboard reports on that structure. It cannot build it.
Companies that skip this step will spend 2027 the way most spent 2026: buying more AI, feeling faster, overspending by double digits, and writing ROI reports nobody believes. The companies that do the work will know, line by line, what they are paying for and what it is earning, and they will be the ones who can keep investing when the board starts asking harder questions. At Agor AI Advisory, we sit with leadership teams to draw that line between utility and wager, build the bet cards, wire unit-cost visibility into the overhead, and design the harvest so that individual speed turns into enterprise return. The surveys this autumn told you where the gap is. The gap closes only when someone redesigns the operation to close it, and every quarter you wait, your stapler budget grows and your bets stay unmeasured.
Sources
- diginomica, "We're on the road to ROI: McKinsey's verdict on AI in 2026", November 4, 2026
- Deloitte press release, "Enterprise AI Enters an ROI Reckoning", October 29, 2026
- CFO Dive, "CFOs confront cash-flow risks as AI costs rise", November 12, 2026
- Validity via PR Newswire, "61% of Marketers Face C-Suite Pressure to Prove ROI", November 11, 2026