Ninety Percent Expected It. Eleven Percent Shipped It.
Vishal Sachar
Co-Founder & CEO of CLRT
Every January produces a number the rest of the year must live with. This year it was BCG's: roughly 90 percent of CEOs said AI agents would deliver measurable ROI in 2026, and they committed more than 30 percent of their AI budgets to the belief. By midsummer the delivery figures had arrived from other desks. Deloitte counted 11 percent of organisations with agents actually in production. The distance between those two numbers is the story of the year in enterprise AI, and almost everyone is reading it wrong.
Read the four entries together, because they come from four independent research desks and they do not disagree; they describe different floors of the same building. BCG's AI Radar, published in January, found that roughly 90 percent of CEOs believe agents will produce measurable ROI this year, with more than 30 percent of 2026 AI budgets committed to agentic AI. Deloitte's State of AI in the Enterprise, a survey of 3,235 leaders across 24 countries, mapped where organisations actually stand: 30 percent exploring agents, 38 percent piloting, 14 percent with solutions ready to deploy, and 11 percent in production. McKinsey found 62 percent of organisations at least experimenting with agents, yet in no single business function do more than 10 percent report scaling them. And Gartner forecasts that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, in a vendor landscape where, of the thousands claiming to sell agents, it counts only around 130 as the real thing.
Then, on 2 July, the ledger acquired its most expensive line. Meta has reorganised thousands of people into AI groups, one of them named Agent Transformation, against roughly $145 billion of expected AI infrastructure spend this year, which makes it the most heavily resourced agent bet on earth. TechCrunch reported that Mark Zuckerberg told staff that agent development had not "accelerated in the way" leadership expected, and that the benefits of the restructure had not "come to fruition yet". Strip the corporate softening and the admission is stark: near-unlimited capital, a full reorganisation and the largest infrastructure budget in the industry did not, by themselves, produce working agents on schedule. If money and models were the constraint, Meta would not have this problem.
The comfortable explanation, that the models are simply not ready, no longer survives contact with the capability data. Stanford's AI Index tracked agent success on real computer tasks, the OSWorld benchmark, rising from 12 percent to roughly 66 percent, within six points of human performance. Capability crossed the usefulness threshold this year. What did not cross is everything around it. The gap between ninety and eleven is a direction gap, knowing which specific workflow in a specific business an agent should take, and a verification gap, engineering the checks, the governance and the escalation paths that make an agent's output trustworthy enough to run without a human hovering. Deloitte's own numbers make the diagnosis: 74 percent of organisations plan to deploy agents within two years, while only 21 percent have a mature model for governing them. The market bought capability and skipped judgment, and it is judgment that ships.
Here is the contrarian reading, and I hold it with some conviction: the reckoning now under way is good news, provided you are a buyer rather than a vendor of hype. Gartner's cancellation forecast is not a prophecy of winter; it is a clearing mechanism. Every cancelled project returns budget to an organisation that now knows one more place agents do not belong. Every exposed case of agent washing narrows the field toward the small number of vendors doing real work. Expectations reset from ninety toward something honest, and the price of discipline, of diagnosis before deployment and verification before trust, stops looking expensive and starts looking like the only spend that survived. The second half of 2026 is a buyer's market for judgment.
The gap between ninety and eleven is not a capability gap. It is a judgment gap, and judgment does not ship in a model release.
A deeper dive
It is worth being precise about why production is where agent programmes die, because the funnel is routinely misread as a maturity queue, a set of stages every organisation will pass through given time. It is not a queue; it is a filter, and the filtering criterion is not technical sophistication. A pilot is a demonstration that an agent can do a task. Production is a commitment that the business will run on the agent doing it, which imposes a different class of requirement: a definition of wrong precise enough to check automatically, a maker-and-checker separation so the system that did the work is never the only judge of it, escalation paths for the cases the agent cannot decide, cost ceilings so a loop cannot bill you into a hole, and a named owner who answers when it slips. None of that arrives with the model. All of it is engineering, and all of it depends on an upstream act of judgment most programmes skipped: choosing the workflow in the first place. A large share of the pilots stalled between 38 percent and 11 percent were pointed at the wrong work, tasks chosen because they demonstrated well rather than because a wrong output would be cheap to catch and a right one would move a number. Wrong-workflow pilots do not fail loudly; they succeed in the demo and then quietly refuse to clear the reliability bar production imposes, which is exactly the pattern Deloitte's funnel records.
Read Gartner's forecast through that lens and it stops being pessimism and becomes a diagnosis. The three cancellation drivers it names, escalating costs, unclear business value and inadequate risk controls, are not model failures; each is a direction or verification failure, visible before a single token was spent to anyone who asked the diagnostic questions first. The same lens explains the Meta admission better than most coverage did. A company can reassign thousands of people and spend $145 billion on infrastructure, and neither act substitutes for the two inputs that do not scale with capital: knowing precisely which work the agents should take, and building the verification that makes their output trustworthy. Note also who is telling you what. The expectation numbers come from CEOs surveyed by firms with transformation practices to sell; the delivery numbers are the ones I would weight. For a buyer, the practical consequence is leverage. In the second half of 2026 you can demand what the first half let vendors refuse: a named workflow before a licence, a definition of wrong before a deployment, evidence of governance before scale, and the figure the agent is supposed to move.
Key terms
- Agent washing
- Gartner's term for rebranding existing software, chatbots and RPA as agentic AI without substantial agentic capability. Of the thousands of vendors claiming to sell agents in mid-2025, Gartner counted only about 130 as real.
Work with CLRT
The eleven percent are not better funded than the rest; Meta has just demonstrated that funding is not the variable. They pointed agents at the right workflow and engineered the verification that let the business trust the output, which are the two things CLRT builds for its clients and the two things no model release will ever include. If you want to know which side of the ledger your organisation actually sits on, and what crossing would be worth, CLRT Ascent will tell you: a score, the workflow where your leverage actually is, and the opportunity in dirhams. This is a buyer's market for judgment. Buy well.
Vishal Sachar
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
