In August 2025, MIT published an uncomfortable snapshot of enterprise AI adoption: only 5% of generative AI pilot projects reach measurable financial impact. The research, from MIT's Media Lab through Project NANDA, is built on interviews with business leaders, a broader survey of employees, and an analysis of hundreds of real AI deployments inside companies — not a survey of expectations, but a look at what actually happened once the contract was signed.
The report doesn't blame the quality of the AI models involved. It points to the gap between what companies budgeted and what the project actually required once it left the pilot stage. And that gap follows a pattern: the license fee is the easiest number to calculate — and the smallest part of the total cost.
For whoever signs off on the technology budget, that distinction is what separates a project that works from one that quietly joins the pile of forgotten pilots. The right question was never "how much does the tool cost" — it's "how much does it cost to make this tool work inside the company, and keep it working."
What The Original Budget Missed
A survey from IDC, commissioned by DataRobot and published in December 2025, put a number on that gap. It covered 318 technology leaders at companies with more than a thousand employees in the United States, the United Kingdom, and Ireland. Among companies using generative AI, 96% paid more than they had planned. Among those using AI agents, the figure was 92%. And 71% admitted having little to no control over where those extra costs were coming from.
The same research points to where the money leaks out: usage running higher than expected, fixing wrong answers the AI produced, and technology teams spending their time stitching different tools together so they actually work with each other. At companies with heavier AI use, close to half of the technology team's time goes into that upkeep — not into what the company actually set out to automate.
A separate study from the technology-management firm Flexera showed the other side of the same problem: only 31% of companies say they have real visibility into what they spend on AI, and just 36% see their entire technology spend clearly. Without that number, budgeting the next project turns into a guess — and 59% of companies reported that AI waste grew over the past year.
Both studies point to the same pattern: the license has a fixed price from day one. Everything else — preparing the content the AI will draw on, connecting it to the systems that already exist, fixing what goes wrong, and tracking how much each team actually uses — only shows up once the project is running. And that remainder is usually what decides whether the project survives its second year.
More Licenses Won't Fix This
The most common reaction to a blown AI budget is to buy more: more licenses, a higher-tier plan, a bigger contract with the same vendor. That solves the wrong problem — MIT's research found that more than half of enterprise AI budgets go to sales and marketing tools, exactly where measured returns were lowest. The biggest returns showed up in back-office work: replacing outsourced services, cutting spend on outside agencies, automating repetitive daily tasks.
The same study found a difference worth noting: buying a specialized vendor solution succeeded about 67% of the time, roughly triple the rate of projects built in-house from scratch. What separates success from failure isn't how much was spent — it's where that spend goes, and how much rebuilding it avoids that a specialized vendor already solved.
Upgrading a plan doesn't fix the visibility problem either. If nobody knows which team is consuming what, a bigger contract only produces a bigger number at the end of the month — without answering the question that actually matters to whoever approves the budget: why the spend grew, and where to cut without blocking whoever depends on the tool.
Seventy-one percent of companies admit having little to no control over where their extra AI costs come from.
What has to be in place
Both studies point in the same direction: the problem isn't the amount spent, it's the missing structure around that spend. A well-designed AI environment fixes this with a few concrete mechanisms, not a pricier contract.
Transparent cost per conversation, not a flat, idle subscription. Every interaction carries a visible cost, and the company pays for what actually gets used — not a fixed number of licenses, many of them sitting idle.
A shared usage budget for the whole team, with spend visible by department. Instead of one license per person, a common balance any team can draw from, with each team's spend tracked in real time — the same visibility 31% of companies say they lack today.
Automatic routing to the cheapest AI model for each task. A routine question doesn't need the most expensive model available; saving the more capable model for what actually requires it is what prevents the runaway usage Flexera measured growing year over year.
Knowledge and connections reused, not rebuilt for every new project. Connecting AI to a system another team already uses shouldn't mean redoing that integration from zero — that's where much of the cost of standing up a new tool lives.
Human approval before a sensitive spend or action, with an audit trail. Every relevant decision gets logged and traceable, which cuts down on the kind of late, expensive correction that shows up when nobody reviewed it beforehand.
This is how Skyller was designed: a usage budget shared across the team, with spend visible by department, and automatic routing to the cheapest model for every routine task.
Where The Budget Shows Up Again
The most immediate gain from a budget built this way is predictability: whoever approves spend knows, month by month, what each team is actually using — and can act before the number turns into a surprise at quarter-end, the same surprise that caught 49% of companies off guard enough to delay or scale back AI projects, according to KPMG data cited by Flexera.
There's a second, quieter gain: reusing what has already been connected and approved costs far less than rebuilding it. A connection to an internal system, an approval flow already set up, a body of knowledge already organized — all of that, done once by one team, becomes available to the next one, instead of turning into a new project with its own integration cost built in.
The result is a budget that grows in proportion to actual use, not to the size of the next contract. On a platform like Skyller, that consumption shows up by department, month by month, for whoever decides where to invest next.
The Questions The Budget Skips
Before signing off on the next AI project, it's worth bringing these questions to whoever approves the budget:
- How much will it cost to prepare and maintain the knowledge the AI will use? Internal documents, policies, and processes are rarely ready for this — organizing and keeping them current is ongoing work, not a one-time step.
- Who is going to connect this AI to the systems the company already runs, and how long will it take? A proposal that doesn't mention that timeline is usually underestimating the whole project.
- Cost scales with usage — is anyone tracking that consumption month by month, by department? Without that visibility, the budget only shows up finished, at month's end, with no explanation.
- If the pilot doesn't move beyond its first team, how much has already been spent, and what's left afterward? An isolated pilot with no reuse plan tends to become a sunk cost, not a lesson learned.
