4 days– That's how long it took one company to consume 106% of its monthly artificial intelligence budget. The company had expanded its use of autonomous agents and cloud coding tools in May. By June 8, the budget was gone, leaving 22 days of the month unfunded. This is not some isolated AI cost crisis. In fact, many large-sized companies have already blown through their annual AI budget within 3-4 months. While others are watching their budgets quadruple almost overnight.
The problem is not simply that AI is expensive. It is that most organizations are still budgeting for predictable software seats while consuming AI through variable, autonomous AI workloads that can scale faster than anyone anticipated.
Want the full picture? Read more about the enterprise AI budget crisis, the failures of old cost-thinking, how to calculate AI costs and match governance to its actual stage of AI adoption.
Why traditional pricing models fails for AI– and how to calculate AI cost the right way
To understand why 106% of a month's AI budget disappeared in 4 days, you have to first understand how most finance teams still calculate AI cost:
1. Cost model has shifted to recursive loops over prompts
The chatbot economics of 2023-2025 ran on a linear model where one prompt equals roughly one predictable cost unit:
Human asks → Model answers→ Session ends
In long-term, however, a single "task" passed to an AI agent can trigger dozens of underlying AI model invocations before a human ever sees a result:
Plan → gather context → call tools → edit/review → test → retry → escalate → verify
A prompt-response exchange is too small a unit to price against anymore. The real cost unit is the loop and unlike a prompt, a loop has variables:
- Number of retries
- Turns taken
- Context drag across the chat
- Model called, and how many copies of itself run at once.
Pricing AI spend off "number of prompts" was never going to catch this, because prompts aren't where the multiplication happens. Which means the old mental model — one exchange, one cost — needs to be replaced with a formula that actually accounts for every place cost gets multiplied. The total cost now will be:
Agent Cost ≈ Tasks × Attempts × Agent Turns × Context Size × Model Price × Parallelism
| Term | Impact on enterprise spend |
|---|---|
| Tasks | The raw count of distinct user objectives or workflows initiated. This is the single lever corporate finance anticipates and plans for when scaling AI access. |
| Attempts | The number of full passes an agent makes to self-correct failures, fix broken code, or re-parse inputs. Every retry executes an entire hidden loop that never registers as a "new message" in traditional prompt tracking. |
| Agent turns | The total sequence of tool calls, environment checks, API queries, and intermediate file edits required to move a task from start to "done." |
| Context size | The volume of tokens loaded per AI model invocation (e.g., entire repositories, system logs, tool returns, and prior turn histories). |
| ModelPrice | The base rate of the target model family. |
| Parallelism | The number of concurrent sub-agents running simultaneously across developer worktrees or background pipelines. |
We ran our client's June numbers through this formula, lever by lever and found that 72.1% of total spend was concentrated in Opus-class AI models compared to Sonnet 4.6 at 15.8% and all other models combined at 12.1%. Unconstrained, premium model defaults multiply every loop parameter upstream.
2. Headcount-based pricing breaks when 1% of users drive the spend
Traditional financial modeling assumes software spending scales neatly with headcount: Seat count × Average usage. That method works when usage is roughly uniform — when one licensed user consumes roughly the same amount as another.
It completely breaks down for AI, because AI usage isn't uniform, but concentrated with a Gini coefficient of 0.75 for spend and 0.72 for tokens consumed. Which means the top 1% of users generate 46 times more AI-written code and merge 15 times more pull requests than the median developer.
A seat-based forecast has no way to represent that. It assumes user #50 and user #500 cost roughly the same, when in reality a small tail of high-intensity users can be running 10, 20, even 50 times the loop-volume of everyone else on the same license.
Anchoring enterprise governance to AI spending
Governance only works if it's matched to where an organization actually sits in its adoption journey, and split cleanly between who's accountable and what's controlling. Most engineering organizations follow a predictable, six-phase adoption lifecycle:
| Phase | What happens |
|---|---|
| 1. Push | Leadership encourages employees to "use AI everywhere." Adoption and experimentation are the primary goals. |
| 2. Power-user | A small group begins generating a disproportionate share of usage and value. |
| 3. Agentic | Work shifts from isolated chats to long-running loops, background agents, retries, and parallel execution. |
| 4. Invoice | Usage-based charges reveal the real cost of autonomous consumption. |
| 5. Governance | Controls are formalized around budgets, model access, workflows, and business accountability. |
| 6. ROI | Spend is judged against accepted code, completed workflows, time saved, revenue, or other verified outcomes. |
Our client followed the arc diligently. May was Push. Without anyone naming it, they were already sliding through Power-user and Agentic by early June. June 8 was the Invoice-phase shock where the consumption became visible, hard enough that it forced Governance overnight.
The goal is to build Phase 5 capabilities while agentic adoption is still expanding, before the Phase 4 invoice forces blunt controls into place. Preventing another June cannot be delegated entirely to finance, engineering managers, or the platform team. Effective governance requires two connected ownership layers for cost optimization:
- An accountability layer that determines whether spending is justified.
- A control layer that keeps consumption within those decisions.
AI tooling can detect an expensive loop, but it cannot decide whether that loop is strategically valuable. A manager can approve a high-value workflow, but cannot reliably control model routing, background retries, or parallel execution manually.
Layer 1: Managers own accountability
Managers are the boots on the ground who can explain why a line item moved in the first place. Say an engineer spins up an agentic CLI to refactor a legacy codebase overnight. The agent gets trapped in a self-correction loop, retrying 400 times against a premium flagship model. Under this framework, p99 usage alert brings the manager in before that single loop consumes the team's monthly allocation.
Rather than treating generative AI consumption as an unmanageable capital expenditure, managers evaluate performance against business outcomes:
- Determine whether higher consumption of AI models is producing more accepted code, faster delivery, fewer defects, or measurable customer value.
- Separate productive power users from abandoned experiments, pet projects, and workflows with no accountable owner.
- Identify where one project or account is effectively subsidizing another's AI consumption.
- Sponsor exceptions when premium models, additional retries, or parallel agents have a genuine business case.
- Stop or redesign workflows whose costs continue rising without a corresponding improvement in outcomes.
Layer 2: Platform/FinOps own control
Platform teams must deploy automated technical guardrails targeting the primary multipliers: ModelPrice and Parallelism. The platform can:
- Set cost-effective mid-tier models (e.g., Sonnet-class) as the global default. Escalation to flagship models (e.g., Opus-class) requires explicit approval or triggered thresholds.
- Trace every model invocation back to the task, team, and workflow that triggered it including sub-agent spawns and background threads
- Surface retry counts, turn counts, and context size per task, not just per month
- Flag concentration in real time. If 70%+ of spend is running against a single premium model tier, that needs to be visible before it's a line item, not after
From financial accountability to competitive advantage
Attribution is the heart of accountability, and accountability is the heart of ROI. Miss just one step in that chain, and the entire AI investment house of cards crumbles. This is why reaching the Governance Phase by design rather than reacting after an invoice shock is the fastest route to real ROI.
But governance can't stop at spending caps and approval gates. Moving from uncontrolled token burn to predictable delivery means connecting every model call back to the project, workflow, team, and outcome it supports. Something dashboards and caps alone don't do.
EPAM's AI/Run was built to exactly solve this adoption-ROI dilemma. It gives teams a unified view of model usage, cost, performance, and business attribution to calculate AI cost in real-time. Leaders can identify expensive workflows, route tasks to cost-effective models, enforce guardrails, and manage agentic AI cost based on outcomes rather than raw token volume.
