Budget season has a smell. It is the smell of February's optimistic spreadsheets being reopened in October by the person who has to sign them. On October 2, 2026, Glenn Hopper published a piece in his Deep Finance Dispatch newsletter called "Your AI Pilots Have Until Budget Review." One line from it has stayed with me since: "A blended number got these projects funded once. It won't get them renewed."
Most advice on measuring ROI on AI initiatives assumes the hard part is the measuring: picking metrics, building dashboards, attributing revenue. I think the hard part is the decision a measurement is supposed to force. A number that never makes you cancel anything is ornament. The return on an AI portfolio depends on how fast you bury the projects that lose money, more than on how good your best ones are. Three releases from the past six weeks make that case better than I could make it alone.
The average is a hiding place
On September 24, 2026, Gartner released a survey of 160 senior finance leaders, run from January through April. CFO Dive reported the two headline figures side by side: 55% of CFOs reported positive overall returns from their 2025 AI initiatives, and 57% said returns from individual use cases were unclear.
Put those numbers together. More than half of finance chiefs believe their AI spending pays off in aggregate. More than half cannot say which parts of it pay off. Both beliefs live in the same building, often in the same head, and that pairing is how a portfolio quietly fills with dead weight.
A blended return works like a group photo. Everyone looks fine at that resolution. Say a company runs ten AI initiatives in a year. Say two of them, perhaps invoice extraction and a code assistant, return four times their cost. Say five roughly break even and three lose money every month they run. Add it up and the portfolio is comfortably positive, the board hears "AI is working," and the three losers get renewed because nobody ever looked at them alone.
Hopper's point is that finance teams already know this trap from every other part of the business. His line: "Finance would send back any department report that showed only net income." No controller would accept a regional P&L reported as a single number. AI spending gets that treatment all the time, because it arrived as a strategic bet and strategic bets are judged by mood.
The blended number does more harm than plain uselessness. It protects the weak projects. The two winners carry the three losers into next year's budget, and the bill for the losers keeps growing while everyone talks about the winners.
Where the 81 percent comes from
Three weeks before the CFO survey, on September 1, 2026, Gartner released a larger study of 1,303 respondents at organizations with at least $50 million in annual revenue. The Gartner release found that only 22% of organizations had scaled AI across multiple business units or adopted an AI-first approach. It also found that about 11% did not know what their function had spent on AI in 2025.
The figure I keep returning to is the split between high and low performers. Gartner defines high performers as organizations that track the ROI of AI initiatives and regularly reallocate or discontinue the ones that underperform. Those organizations reported positive returns on 81% of their AI initiatives. Low performers did not know the rate of return for 29% of theirs.
Read the definition again before you read the result. A high performer is partly defined as a company that kills its losers. A company that kills its losers will, by construction, report a high share of winners among the projects still running. Some of that 81% is pure arithmetic. I don't think that weakens the finding. It is the finding. The best-performing AI portfolios in Gartner's sample got there by subtraction. They placed more bets than they kept, and they had a mechanism for ending bets that went wrong.
The 29% figure says just as much from the other side. A project with an unknown return cannot be cancelled, because nobody can win the argument for cancelling it. The sponsor says it needs another quarter. The vendor says adoption is ramping. Finance has no line item to point at. So the project survives on what I would call default immortality, and every month it survives it takes budget, engineering attention and executive patience that a better project could have used.
Popular is a different list from profitable
The same Gartner release has a quieter finding that should worry anyone who picked their AI roadmap from a vendor's slide deck. Among IT functions, the most-pursued use cases were cybersecurity threat detection and response (54%), IT service desk automation (54%), and automated code generation and refactoring (44%). The use cases most often producing positive returns were a different list: intelligent IT asset and cost optimization (40%), synthetic data generation (28%), and automated code generation and refactoring (23%).
Service desk automation tied for most popular and does not appear on the returns list. Asset and cost optimization, which nobody demos at a keynote, leads it. Companies choose AI projects by what looks good in a pilot and judge them by what shows up in the ledger, and those two lists overlap far less than the demos imply.
Measuring ROI on AI initiatives starts with a death date
If the return comes from subtraction, the measurement system exists to support one decision: keep this or stop it. Everything else is reporting. Here is how I would design ROI measurement for an AI portfolio with that decision at the center.
Write the obituary first
Clinical trials pre-register their endpoints for a simple reason. Once the data arrives, every researcher can find a story in it. AI projects need the same discipline. Before a project starts, write down what result by what date ends it. Write it while the team is optimistic and has nothing yet to defend.
Gartner has handed CFOs a usable starting clock. Per CFO Dive, straightforward finance use cases like data extraction, accounts payable and receivable automation, and report creation typically reach expected value in nine to 10 months. Marco Steecker, a senior director analyst in Gartner's finance practice, framed the cut-off plainly: "if you haven't gotten value by this point in time," and you have diagnosed the root causes, the initiative is likely consuming money with no clear path to value. Gartner's report puts the aim as knowing "where to invest, when to cut underperforming initiatives."
Use different clocks for different work. Forecasting, scenario planning and insight generation take longer to build, and Gartner says so. The risk is that "this one takes longer" becomes the excuse every project reaches for. The defence is to set the longer clock at the start, in writing, and to hold it.
Give every project its own meter
You cannot kill what you cannot see, and most companies cannot see their AI costs. On August 11, 2026, CFO Dive covered the 2026 State of AI Cost Governance Report from Mavvrik and Benchmarkit. Only 11% of organizations could forecast AI spending within plus or minus 10%, down from 15% in 2025. Token costs were a top source of surprise spending for 43%. And 98% of engineering organizations use AI coding assistants, while only 42% include developer AI tool spending in their AI cost reporting.
That last gap is the one I find most telling. The most widely used AI tool in the company is often missing from the AI cost report. If a cost line is invisible, the return on it can only be guessed, and guessed returns are always positive.
A per-project meter has to hold three kinds of cost. The first is the vendor bill, including tokens, which now move with usage rather than seats. The second is integration and maintenance engineering. The third is the one almost everybody leaves out, the human supervision hours spent checking, correcting and re-running what the system produces. Hopper's kill review asks for exactly this all-in line, internal supervision hours included. An agent that saves forty hours of clerical work and needs thirty hours of senior review has a very different return from the one in the pitch.
Give the number to someone with nothing at stake
The third design rule is about people. Whoever sponsored a project should not own its results number. Hopper names the missing independent owner as one of the three reasons pilots stall, alongside missing project codes and missing end dates. It is an old governance idea, and AI projects keep skipping it because they were treated as innovation, and innovation was supposed to be exempt from audit.
Productivity is the softest number on the page
The default AI business case is productivity, and productivity is where measurement goes to die. Gartner's September 1 release found productivity was a target outcome for 75% of functional leaders and took about 30% of functional AI spending, nearly twice the next objective. The September 24 CFO survey had 73% of finance leaders naming productivity as a 2025 goal.
Hours saved become money only when someone decides what to do with them. You can stop hiring into a role, close a backlog that was losing revenue, or move people onto work that sells. Without one of those decisions, saved hours spread into longer lunches, better slide formatting and a slightly calmer team. Those are pleasant results that never appear in a cash flow statement.
Buyers are starting to notice. Futurum's survey of 830 IT decision-makers, published February 17, 2026, found productivity's share as the primary ROI metric fell from 23.8% to 18.0%, while direct financial impact rose to 21.7%. Keith Kirkpatrick, Futurum's research director, put it bluntly: "Sales teams leading with 'save 4 hours per week' are entering a losing conversation."
My rule for clients is simple. A productivity claim must name the decision that turns the hours into money, and the date that decision gets made. If nobody can name it, score the project as a cost with an option attached and treat it that way. Options can be worth holding. They should be priced as options, with an expiry date.
The reallocation gap
Measurement without reallocation is bookkeeping. The IBM Institute for Business Value's 2026 CFO study, released September 30 and based on 1,500 CFOs across 33 geographies and 26 industries, shows how wide the gap between watching and acting still is.
Forty-eight percent of CFOs say finance monitors AI-driven value and shifts capital in response. Only 8% say finance leads enterprise-wide AI value goals with automated investment triggers. Just 6% let AI propose or carry out reallocations within set limits. So most finance teams can see AI value, a small minority have wired that sight to the budget, and only a sliver have let the machine move money.
The authority is arriving faster than the machinery. In the same study, 56% of CFOs report more authority over portfolio management and capital reallocation. Fortune's coverage of the study quotes IBM CFO Jim Kavanaugh describing the shift from "guardians of stability" to "agents of transformation," and says he sees sequencing as the main obstacle. Companies buy the AI tools before they fix the data architecture and workflows underneath.
Two CFOs quoted in IBM's release describe the operating model I think most companies need. Hisako Takada of Advantest: "Capital allocation should operate dynamically and be reviewed continuously as opportunities and conditions change." Lindsey Crossland of KPMG UK talks about the need to "keep an element of funds ready to allocate to new initiatives as they take flight."
Crossland's line holds the whole argument in small. The reserve for new bets has to come from somewhere. In a flat or shrinking budget, it comes from the bets you end. A company that cannot kill projects has to fund new ones with new money, so its AI spend climbs every year while its learning rate stays flat. Gartner's September 1 release found 85% of functional leaders planned to increase AI spending in 2026, after putting an average of 12% of functional budgets into AI in 2025. I would like to know how many of them plan to cut anything.
The speed of no
IBM reports that organizations led by what it calls AI-first CFOs approve funding for new AI initiatives 15% faster than peers. The study is correlational, and IBM frames it as a finding about the speed of yes. I read it as a finding about the speed of no. You can only say yes faster if you have the capacity, the budget and the attention to absorb another project. That capacity comes from somewhere. In a well-run portfolio, it comes from the funeral last quarter.
What a funeral tells the board
The last step of Hopper's four-step review strikes me as the most radical, and also the most obvious. Cancel at least one project, and document why. His reasoning is that a documented cancellation proves finance can tell its projects apart, which makes boards more willing to fund the rest.
Directors have sat through years of AI updates in which every project was "promising" or "scaling." A list like that carries no information, because a measurement system that rates everything as promising cannot tell projects apart. A board that hears "we stopped the support agent in July because resolution rates stalled at the level we set as our floor, and we moved its budget to cost optimization" learns two things at once. It learns about the support agent. It also learns that the measurement system has teeth, which raises its trust in every number that follows.
This is why I now ask clients one question before any other about their AI program. How many AI initiatives did you end last year, on purpose, against criteria written in advance? Zero is a common answer. Zero tells me the portfolio's reported return is a story rather than a number, however good the dashboards look.
Gartner's IT Symposium/Xpo opens in Orlando on October 19, and the opening keynote, given by Daryl Plummer and Kristin Moyer, is titled "Reaching Escape Velocity - Shaping an Effective Enterprise AI Economy." I like the metaphor more than its authors may intend. Rockets reach escape velocity by dropping stages. Every spent stage they carry is mass they must push, and enough of it keeps any rocket on the ground. Most enterprise AI portfolios I see in 2026 are still carrying every stage they ever lit.
A kill rate is a metric
Treat the kill rate as a first-class number on the AI dashboard, next to spend and return. Suppose a company launches twelve AI initiatives in a year and ends none of them. Either its selection process is perfect or its measurement cannot tell success from failure, and nobody's selection process is perfect. Suppose instead it ends four, and the money from those four funds the second phase of the two that worked. That company will report a lower count of active AI projects and a higher return per dollar, and it will have learned more in a year than its rival learned in three.
There is a cultural cost, and pretending otherwise would be a lie. Teams resist ending projects because ending feels like failing, and because careers in most companies are built on what was launched. The fix is to reward the decision rather than the outcome. A product lead who writes clean cut criteria, hits the date, reads the data honestly and recommends ending her own project has done the most valuable thing an AI program can produce. She has turned a bet into knowledge. Promote her.
Build the machinery before you buy more tools
I can predict what a lot of companies will do with September's numbers. They will buy an ROI dashboard. A vendor will sell them one, it will show tidy charts of blended returns, and next October the same 57% will still be unable to say which use cases pay.
The tools are the cheap part. The expensive part is the set of commitments that make a measurement bite. Every AI initiative gets its own cost line, supervision hours included. Every one gets its own clock, set in writing before launch. Results belong to an owner who does not sponsor the project. Spend moves on a calendar, with a reserve for new bets funded by the ones you end. A board sees the kill rate next to the return. None of that comes in a box. It has to be designed into how the company allocates money, and it touches finance, IT, procurement and the incentive plan all at once. That is why most companies have not done it, and why the 22% who have scaled AI in Gartner's data look so different from everyone else.
This is the work Agor AI Advisory does. We design the decision system around your AI portfolio: the cut criteria, the cost meters, the reallocation cadence and the board reporting that turns a pile of pilots into a portfolio that learns. We have no product of our own to protect, so we can tell you which of your projects to bury. Budgets lock in the next few weeks. Every month an unmeasured project survives, it takes money from a project that might have worked. Make the cuts deliberately before the calendar makes them for you.
Sources
- Glenn Hopper, "Your AI Pilots Have Until Budget Review," Deep Finance Dispatch, October 2, 2026
- CFO Dive, "CFOs need realistic AI time-to-value expectations: Gartner," September 24, 2026
- Gartner, "Gartner Survey Finds Only 22 Percent of Organizations Have Successfully Scaled AI Across Multiple Business Units," September 1, 2026
- IBM Newsroom, "IBM Study: As AI Scales Enterprise-Wide, CFOs Play an Expanded Role in Transformation," September 30, 2026
- Fortune, "AI is giving CFOs a bigger slice of the C-suite, new IBM data shows," September 30, 2026
- CFO Dive, "1 in 4 companies delay or cancel an AI project over cost," August 11, 2026
- Futurum Group, "Enterprise AI ROI Shifts as Agentic Priorities Surge," February 17, 2026
- Gartner, "Gartner Announces Keynote Speakers at Gartner IT Symposium/Xpo, October 19-22, in Orlando," September 17, 2026