Where the Money Actually Lands
Vishal Sachar
Co-Founder & CEO of CLRT
There are two maps of enterprise AI. One is drawn in keynotes and analyst decks, and on it the technology is everywhere: legal teams reviewing contracts in minutes, finance closing the books with agents, HR running itself. The other map is drawn in budgets, and it looks nothing like the first. In late 2024, Menlo Ventures surveyed 600 IT decision-makers at US enterprises, venture research, so read it directionally, and put the year's spend on generative AI at $13.8 billion, six times the $2.3 billion of 2023. The total made the headlines. The composition did not, and the composition is the story, because a budget records what organisations actually did after the keynote ended.
Break the spend down by the function that carries the budget and the concentration is immediate. IT holds 22 percent, the largest single share of the $13.8 billion. Product and engineering hold another 19 percent, data science 8 more. That is nearly half of every dollar, 49 percent, that never leaves the technical functions. The go-to-market side takes roughly a quarter: support at 9 percent, sales at 8, marketing at 7. And the departments the transformation story was actually about split what remains. HR and finance carry 7 percent each, design 6, and legal, the most document-bound function in the company, carries 3. The functions drowning in exactly the repetitive, text-heavy work this technology is best at received the scraps. The map is precise, and it is upside down relative to the story that raised the money.
The explanation is not that CIOs know something the keynotes do not. It is procurement gravity. Budget lands where there is an owner, a line item, and a familiar buying motion, and in most organisations only the technical functions have all three for software. IT can classify an AI tool as infrastructure and buy it the way it has bought everything for twenty years. Engineering can trial a coding assistant on Tuesday and expense it on Friday. Legal has no equivalent muscle: no evaluation habit, no vendor shortlist, no budget line called software that thinks. So the money followed the path of least resistance, and the spend map became a map of purchasing convenience. There is nothing wrong with copilots for engineers. The error is reading that map as a verdict on where AI creates value, when it mostly records who already knew how to buy.
Look inside the stack and the same survey shows where the dollars sit vertically. The application layer took $4.6 billion, the largest tracked slice, against $3.5 billion for foundation models, evidence that enterprises buy outcomes wrapped in software rather than raw capability. And the pace is the other half of the story. By Menlo's mid 2025 update, enterprise spend on model APIs alone had reached $8.4 billion, up from $3.5 billion roughly eight months earlier. The money is accelerating faster than its composition is changing, which matters, because whatever pattern was set early is now compounding. Each renewal cycle deepens the groove: the funded functions generate usage data, case studies, and internal champions, and next year's budget is written by the people who spent this year's.
Now set the budget map against a different axis: what a wrong decision costs in each function's core workflow. Support answers at volume, and a bad answer usually costs an apology. Legal reviews contracts, where one missed clause can cost more than the entire AI budget. Finance closes the books and signs numbers regulators read. HR decides matters that end up in tribunals. These are the functions where the work is repetitive, text-heavy, deadline-driven, and verifiable, the precise profile that rewards well-engineered automation, and where the stakes make trustworthy automation genuinely valuable rather than merely convenient. They are also the functions carrying 3 and 7 percent of the money. The gap between where the budget landed and where the leverage sits is not a rounding error. It is the largest unpriced opportunity in enterprise AI, and it is sitting in plain sight.
Why has nobody closed it? Because closing it is not a purchasing decision. Pointing AI at a legal review or a financial close is not a copilot with a different prompt. It requires decomposing the workflow into the steps a system can own and the judgments a person must keep, building verification that defines wrong before the first run, and governance that makes a general counsel or a CFO willing to sign. None of that ships in a product, which is exactly why no vendor motion exists for it and why the budget has not followed. The companies that treat the spend map as the truth will keep deepening the groove they are already in. The ones that read it as a record of convenience, and go looking for the corners it missed, are buying leverage at 2024 prices.
The spend map is a map of purchasing convenience, and the market keeps reading it as a map of value.
A deeper dive
The composition is self-reinforcing in a way that should worry anyone waiting for it to correct itself. Vendors build products for the budgets that exist, so the tooling for engineers compounds while the tooling for a general counsel stays thin, which keeps legal's spend low, which tells the next vendor there is no market there. Benchmarking culture completes the loop: boards ask how their AI spend compares with peers, and because every peer is subject to the same gravity, the comparison certifies the distortion instead of correcting it. Meanwhile the functions that never received budget never generate the baseline data that would prove the case for one. Nobody measures the error rate of a process no system touches, so the cost of the manual status quo stays invisible while the cost of any proposed automation is audited line by line. A map like that does not redraw itself. Someone inside the organisation has to decide the map is wrong, and that decision is a judgment call made against the grain of every vendor pitch and every peer benchmark available.
There is also a reason the underfunded corner stays hard, and it is the honest half of the argument. The workflows with the highest stakes are precisely the ones where a plausible-looking wrong answer is most expensive, so pointing AI at them without verification is not an opportunity, it is a liability. That is why the gap persists, and why it is worth something. Closing it requires the unglamorous work the funded functions never needed: decomposing a contract review or a close process into steps whose outputs can be checked, deciding which decisions must stay with a person, and building the audit trail that lets a regulated function adopt automation without betting its licence on a model's confidence. The organisations that do this first will not look impressive on a spend benchmark, because the benchmark measures conformity. They will simply have automation where their competitors have subscriptions, in the functions where an hour saved or an error caught is worth the most.
Work with CLRT
Reading that gap for a specific business is the work CLRT does. Our diagnostic starts where the budget maps stop: which of your workflows carry real leverage, what a wrong output costs in each, and what it would take to make automation trustworthy where the stakes are high. If your AI spend looks like everyone else's, that is not evidence of good judgment; it is evidence of the same gravity. Start with the diagnostic at ascent.clrtstudio.com, or bring us the function everyone agrees is next and nobody has funded.
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.
