The Short Answer
Enterprise AI budgets break because per-seat licenses bill on headcount while agents multiply token spend invisibly, so the fix is structural rather than tactical: run the platform on infrastructure where you own all the code and the data, pay by usage against hard caps you set, and route each request to the cheapest capable model.
The FinOps Foundation's 2026 survey found 73% of enterprises exceeded their original AI cost projections. Dashboards did not prevent it, because a dashboard reports spend after it happens.
What holds a budget is enforcement at the infrastructure layer: tenant ceilings that stop rather than alert, per-agent allocations, per-user limits, and cost-aware model routing.
How many enterprises actually go over budget on AI?
Roughly three in four. The FinOps Foundation's State of FinOps 2026 drew on 1,192 practitioners responsible for more than $83 billion in annual cloud spend, and 73% of them reported AI costs that exceeded original projections.
The same report shows how fast this became everyone's problem: 98% of respondents now actively manage AI spend, up from 31% two years earlier. AI cost management is now the single most-wanted skill on finance-operations teams, cited by 58%.
A second, independent survey lands in the same place. A WitnessAI study of 300 business executives published in July 2026 found 68% had AI initiatives run over budget in the past year, and 33% said overruns happened "mostly or always."
The return side is thinner still: only 9% of those executives said more than three-quarters of their AI initiatives produced measurable financial return. Overspending on projects that mostly do not pay back is a procurement failure, not an engineering one.
Why does per-seat AI pricing break at enterprise scale?
Per-seat pricing assumes usage is roughly constant per person. That assumption held for traditional SaaS — a CRM license generates similar load whether someone logs in once a day or fifty times, and the vendor absorbs the variance.
Agentic AI breaks the assumption in both directions at once. Headcount sets the bill, but behavior sets the cost, and the two curves are unrelated. You pay for every employee while a handful of automated workflows consume most of the tokens.
The result is a line item that scales with hiring instead of with value. Here is what the same 2,000-person deployment costs under each shape:
| Platform | List price | 2,000 users / year | Scales with |
|---|---|---|---|
| ChatGPT Enterprise | ~$60/user/mo | $1,440,000 | Headcount |
| Glean | ~$40/user/mo | $960,000 | Headcount |
| Microsoft Copilot | ~$30/user/mo | $720,000 | Headcount |
| ibl.ai (self-hosted) | Flat license + compute | Tracks actual usage | Work performed |
The gap is not a discount. It is a different billing shape: the first three charge for people who may never open the tool, while the fourth charges for work the system actually performed.
That is why hiring 500 people should not raise an AI bill by $360,000 before anyone uses anything — and under per-seat pricing, it does.
Where does the money actually go in an agentic deployment?
Not where the pilot suggested. A pilot measures single-turn chat; production measures agents, and agents behave nothing like a chat box.
Three multipliers stack. First, task depth: summarizing one document costs X, comparing twenty costs roughly twenty times X, and an agent iterating toward a goal may call the model hundreds of times for one user action.
Second, model choice. The spread between a lightweight model and a frontier reasoning model runs one to two orders of magnitude per token. Without routing logic, every trivial lookup pays frontier prices.
Third, delegation. Agentic architectures let agents call other agents, so a single click can cascade into ten model invocations across three services — none of which appeared in the pilot's per-user average.
None of these are visible in a per-seat contract. The subscription is the number procurement approved; the token consumption underneath it is the number that actually moves, and it surfaces at the quarterly review.
What does infrastructure-level AI cost control look like?
It looks like enforcement, not reporting. Most platforms offer cost visibility — usage dashboards, spend reports, per-team analytics. Those tell you what happened. They do not stop anything.
Four controls make the difference, and all four have to live below the application, where an agent cannot route around them:
- Tenant-level spend ceilings. The organization sets a hard cap across all users and agents. When it is reached the system stops. Alerting is not a control; stopping is.
- Per-agent budgets. Each agent gets its own allocation — a support agent at $500/month, a research agent at $2,000 — and pauses when exhausted rather than quietly continuing.
- Per-user limits. No single person, and no single runaway automation, can consume a disproportionate share of the pool.
- Cost-aware model routing. Simple queries go to efficient models, hard reasoning goes to capable ones, decided by task complexity rather than by which model a vendor prefers to sell.
The distinction that matters: a dashboard is an after-the-fact narrative, while a cap is a policy the infrastructure enforces in real time. Only the second one survives contact with an agent in a retry loop.
What changes when you own the stack instead of renting it?
Control over the cost controls themselves. On a managed platform, the ceilings, routing rules, and budget policies are product features — you get the ones the vendor built, priced the way the vendor prices them, and you cannot inspect the meter.
ibl.ai is the agentic AI platform where you own all the code and the data.
You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.
Ownership is what makes the budget legible. You see the token meter directly, set caps in your own configuration, and move a workload to a cheaper model — or to your own fine-tune — without renegotiating a contract.
It also removes the variable that broke the forecast. When cost tracks usage rather than headcount, the finance question becomes "what work did we run" instead of "how many people did we hire," which is a question the business can actually answer in advance.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
How do you keep next year's AI budget from breaking?
Model the production shape, not the pilot shape. Before signing anything, run the arithmetic that the FinOps 2026 respondents ran too late:
- Project cost at full headcount, not pilot headcount. Multiply the per-seat price by the population procurement will eventually be asked to cover, then ask whether that number is defensible on its own.
- Estimate tokens per completed task, not per chat turn. Include retries, agent-to-agent calls, and refinement passes. This is the number that diverges most from the pilot.
- Require enforcement, not visibility. Ask the vendor what happens when a cap is hit. If the answer is "we notify you," there is no cap.
- Price the exit. If migrating off the platform means rebuilding every integration, the renewal is not a negotiation.
The pattern underneath all four is the same one the deployment-yield analysis found on the delivery side: AI spending fails on procurement architecture long before it fails on model capability.
For the segment-level math, see the cost comparison for enterprise ownership versus rental, or start with the ibl.ai pricing model and the Agentic OS platform.