A single unflattering number keeps surfacing in this year's AI reports: somewhere between five and seven percent. That is the share of corporate AI projects that can point to a real return on the money spent. An MIT study found just five percent of generative AI projects delivering identifiable returns. BCG put the share of companies consistently generating substantial value at the same five percent. KPMG's spring survey of more than 2,000 senior leaders across 20 countries was starker still: only seven percent could say they had established a return, 42 percent admitted they had only partial visibility into what they were even spending, and nearly a quarter said investors were already pressing them to prove it. Gartner expects 40 percent of AI projects to be scrapped by 2027 for unclear value.
The comfortable reading is that this is a discipline problem. Companies bought too fast, tracked too little, and now the bill has arrived. Ajay Pundhir, an AI leader writing for the Forbes Technology Council, argues that reading is wrong. The real issue, he says, is a category error: finance is applying the right tool to the wrong thing. We are measuring AI like software when it has started to behave like labor.
The distinction matters. Software has been evaluated the same way for thirty years. You buy a license, count the seats, amortize the cost, and treat the tool as a multiplier of human effort: a person uses it to work faster. An agent breaks that assumption. It does not sit on the desk waiting to be used; it reads the contract, drafts the reply, runs the reconciliation and closes the ticket. When the thing you bought produces output rather than merely enabling it, the seat stops being the unit that matters. The unit of work does. That is why vendors are drifting from seat-based licensing toward billing based on consumption, and why payback math built for software keeps misleading buyers.
Pundhir's trap is worth spelling out. Picture a support agent with genuinely cheap tokens. Per interaction it costs a fraction of a human, so the cost-per-seat comparison looks like an easy win and leadership celebrates. Then you measure the actual work and find a 40 percent rework rate: four in ten "resolved" tickets reopen, get re-escalated, or need a human to redo them. Counting the cleanup and the lost trust, that cheap agent is net-negative, and nothing in the traditional model surfaces it. Firms end up mispricing in both directions at once, overpaying for agents that quietly destroy value while killing agents that look expensive on tokens but produce clean, low-rework work. His fix is three questions: what work unit does the agent finish, what is the fully loaded cost per unit once you include human supervision, and what is the rework rate.
The companies that do show returns tend to prove the point in reverse. In a survey of eight firms with measurable AI value, Bernard Marr found a common thread: they started with a business problem, not with AI. Walmart used it to build or fix 850 million product-catalog data points that would otherwise have required ten times the staff. UPS lifted the share of small packages clearing customs in a single day from 21 to 97 percent. IKEA, rather than cutting staff after an assistant resolved 47 percent of queries, retrained 8,500 agents as remote interior-design consultants and booked 1.4 billion dollars in new sales. Newsweek's reporting echoes the lesson from the other side: leaders who chase every new "shiny spoon" lose momentum, while those who define the destination first use AI as a tool rather than mistaking it for the strategy.
Even the vendors now sell discipline. At its first press briefing in Seoul this month, the enterprise AI firm Cohere spent more time on its agent platform than on its own models, and argued that companies should choose models suited to a specific objective rather than reaching for the biggest one available. It also prices by annual users rather than tokens, to make costs predictable. Coming from a model maker, that is a quiet admission that competing on raw performance is no longer enough.
The seven percent, Pundhir argues, will not climb just because the models improve; they already finish work that once took people hours. The bottleneck is the ledger, not the intelligence. Until finance measures agents the way it has always measured labor, by throughput, by quality, and by what comes back clean, the proof of return will stay stuck in single digits. The encouraging part, for buyers, is that this is a problem on their side of the table.