AI Research
Vol. 1·Monday, March 23, 2026
Companies Are Spending the Most on AI Where It Works the Least
Noah Ogbi9 min read
Tips, corrections, or questions? support@omniscient.media
Tips, corrections, or questions? support@omniscient.media
Consequential AI, explained and evaluated, every weekday.
The Omniscient Bulletin: 5 to 7 items a day with the take, not the recap.
The numbers are hard to square. Global AI spending is on track to hit $2.52 trillion in 2026, a 44% increase over last year, according to Gartner.[1] Meanwhile, a July 2025 report from MIT's NANDA project found that 95% of task-specific, integrated enterprise AI deployments deliver zero measurable P&L impact - this, against a backdrop of $30 to $40 billion already deployed into enterprise GenAI.[2] You do not need to be an economist to sense something has gone badly wrong with the allocation.
The MIT report offers a specific explanation for why. When researchers asked executives to distribute a hypothetical $100 across business functions, sales and marketing captured approximately 70% of GenAI budgets - a figure the report itself notes may be closer to 50% depending on the mix of companies surveyed.[2] Back-office automation - document processing, accounts payable, compliance monitoring, internal workflow orchestration - consistently outperformed every other category on actual ROI, yet it attracted a small fraction of total spend. The money is going where the cameras are pointed, not where the returns are.
Back-office automation is genuinely unglamorous. It rebuilds margins quietly, through reduced BPO spend, faster contract processing, and fewer manual exceptions in finance workflows. These gains are real, but they do not make for compelling board presentations or LinkedIn announcements. Sales and marketing AI, by contrast, is loud: AI-generated outreach, smart lead scoring, personalized campaign content. Metrics appear on dashboards within days. The causal chain from tool to revenue looks neat, even if the underlying returns are thin.
This is the structural incentive problem the MIT data exposes. Executives fund what they can attribute, and attribution models overwhelmingly favor visible, top-line functions. The result is a kind of collective misallocation - every individual budget decision looks rational, but in aggregate they point enterprise AI investment toward its weakest applications.
Gartner has placed AI squarely in the "trough of disillusionment," with recovery still ahead.[1] The trough arrives when early pilots fail to scale, when proof-of-concept energy collides with operational reality. Gartner's framing implies the hangover is not yet finished.
The critical question is which departments absorb the most pain when disillusionment peaks. If 70% of GenAI budgets are flowing into sales and marketing, that is where the failed pilots are concentrated. Those are also the functions where executives are most exposed - where overpromised AI tools produce no visible lift despite visible spend. The trough, in other words, will not strike AI budgets evenly. It will hit hardest in exactly the departments that received the most investment.
There is a revealing subplot buried in the MIT data. While official enterprise AI initiatives stalled, a parallel economy of personal AI use flourished. Workers from over 90% of surveyed companies reported regular use of personal AI tools - ChatGPT, Claude, and similar - for work tasks, even as only 40% of those same companies had purchased an official LLM subscription.[2] Employees were, in effect, solving the ROI problem themselves, routing around the corporate procurement cycle entirely.
This matters for understanding where AI value actually lives. The productivity gains showing up in individual workflows - faster drafting, quicker research, better first drafts of code - are real. They are just not appearing in P&L reports because they are not being captured by sanctioned tools. The enterprise AI problem is not purely a technology problem. It is a measurement and incentive problem dressed up as one.
None of this appears to be slowing the marketing industry's appetite. The IAB's 2026 Outlook Study, based on responses from more than 200 brands and agency buyers, found that five of the six top areas of increased advertiser focus are directly tied to AI.[3] Two-thirds of buyers are now focused specifically on agentic AI for ad buying and campaign execution. U.S. ad spend overall is projected to grow 9.5% in 2026, with AI described by IAB as having moved from experimentation to core infrastructure.
This is precisely the dynamic the MIT data should be warning against. The marketing sector is accelerating its AI investment at the moment the evidence most clearly suggests it is the wrong place to concentrate it. Agentic AI for media buying may prove different from chatbots for outbound email - the technology is meaningfully more sophisticated. But the underlying incentive structure has not changed. Marketers still measure what is easy to measure, fund what produces visible outputs, and declare success in a language that boards understand.
The MIT report's finding that external vendor partnerships outperform internal builds by a factor of two is instructive here.[2] The organizations crossing what the report calls the "GenAI Divide" are not the ones with the largest AI budgets or the most ambitious internal programs - they are the ones that targeted specific, process-level problems in operational functions and demanded measurable business outcomes rather than software benchmarks from their vendors. Mid-market companies reached full implementation in an average of 90 days; enterprise-scale organizations took nine months or more.
The implication is not that enterprises should abandon AI in marketing. It is that the current allocation - roughly 70 cents of every AI dollar pointed at top-line functions, with a handful of cents going to the back office - is badly inverted relative to where measurable returns actually concentrate. Correcting that imbalance will require CFOs willing to fund unglamorous infrastructure and executives willing to present boring wins to their boards. That is a cultural problem, not a technical one. And culture changes more slowly than hype cycles.
After publication, we returned to the MIT NANDA primary report to verify the figures above against the original document. Three corrections and one substantive addition follow.
The piece uses 70% as the operative number for sales and marketing's share of GenAI budgets throughout. The primary report presents 70% as the raw survey result from executive interviews, but includes a research note stating that the general functional allocation is "closer to 50%" and that "sub-category and use-case breakdowns should be treated as directional at best." The 70% figure appears in an exhibit; the 50% figure is the report's own headline characterization of the same data. The piece should have led with the 50% figure and treated 70% as the upper bound of survey responses, not the primary claim. The argument holds either way, but the more defensible number is 50% - at that level, the misallocation case is harder to dismiss as a sampling artifact, because it is the report's own conservative estimate, not the ceiling of a single survey cohort.
The 95% failure rate cited in the opening paragraph refers specifically to custom or vendor-sold, task-specific enterprise AI tools reaching production scale with measurable P&L impact - not to all enterprise AI initiatives broadly. The report defines "successfully implemented" as tools where "users or executives have remarked as causing a marked and sustained productivity and/or P&L impact." The piece's parenthetical "(task-specific, integrated)" is accurate, but the framing could imply a wider population than the data covers. General-purpose LLM tools like ChatGPT show a very different pilot-to-implementation curve. The distinction matters: the 95% failure rate is a warning about custom enterprise deployments, not a verdict on AI tools overall.
The original post cited an AI Governance Library summary page for the MIT NANDA data. The footnote has been updated to link directly to the primary PDF hosted by the report's authors. The underlying data is unchanged; the citation is now more precise.
The original piece treats misallocation as the root cause of enterprise AI's underperformance. A closer reading of the primary report suggests misallocation is better understood as a symptom. The report identifies a "learning gap" as the core structural failure: most enterprise GenAI tools do not retain feedback, adapt to context, or improve over time. This is why back-office automation consistently outperforms sales and marketing AI even when budgets are comparable - operational workflows are more structured, which makes them more amenable to tools that can integrate tightly and iterate. The report is explicit: "The core barrier to scaling is not infrastructure, regulation, or talent. It is learning." Back-office processes happen to be the terrain where learning-capable systems can demonstrate value fastest. The misallocation of budget toward sales and marketing makes the problem worse, but it is not the origin of it. That distinction changes the prescription: fixing the budget split is necessary but not sufficient. The underlying requirement is tools that learn, not just tools that automate.