The CFO conversation no one saw coming: Your enterprise rolled out AI coding tools to 500 engineers in January, projected a full-year budget, and by April the money was gone. Not over budget — completely gone. This is not a hypothetical. It is happening right now at companies across the Fortune 500, and the root cause is a single word that was not in last year's AI budget model: agentic.
The shift from AI-as-assistant to AI-as-agent changes the economics of every AI deployment your organization has made. And the numbers coming out of recent evaluations by Microsoft, IBM, and GitHub suggest that most enterprises have wildly underestimated what this transition costs.
The Multiplier Nobody Told You About
When engineers use a standard AI coding assistant — autocomplete, a one-shot code generation, a quick question — the interaction is bounded. The model receives a prompt, generates a response, the context window closes. Costs are predictable.
agentic AI works differently. A single developer instruction — "refactor this payment processing module and write tests" — triggers a planning phase, file reads across the repository, dependency analysis, code generation, test generation, error analysis, and iteration. Every step adds to the model's context window. Every subsequent inference call processes the entire accumulated history.
A Microsoft internal evaluation published this month found that upgrading from one generation of AI model to the next multiplied token consumption by a factor of 10 to 12 in complex agentic coding scenarios. One test run consumed 69 million tokens in a single task — the kind of number that, at enterprise scale, turns a sensible annual budget into a four-month runway.
That 69 million token figure is an extreme outlier. But it illustrates something important: agentic workflows have a long tail of cost that standard request-based billing models were never designed to handle, and that IT teams have no native visibility into.
The GitHub Copilot Shock
On June 1, 2026, GitHub moved its 4.7 million paid Copilot subscribers from premium request units to token-based billing. The base subscription price for Copilot Business ($19/user/month) and Copilot Enterprise ($39/user/month) stayed the same. But the way those seats actually get consumed changed fundamentally.
Core features — code completions, Next Edit suggestions — remain unlimited for paid plans. But Copilot Chat, Copilot CLI, agentic coding sessions, and code review now draw from a monthly GitHub AI Credits pool. Overages are charged at per-token rates. Enterprise customers running active agentic sessions are seeing individual sessions consume $30 to $40 in credits per developer, per session.
Do the math for a 200-person engineering org with developers running three to four agentic sessions per week. The $39/user/month seat fee is no longer the right number to plug into your annual cost model.
This is not a GitHub-specific phenomenon. It is the industry moving toward honest accounting for what agentic AI actually costs to run. The seat fee model assumed bounded, request-based usage. Agentic usage is unbounded by design.
"Tokenmaxx" and the Uber Warning
The cultural dimension of this problem deserves attention from business leaders, not just engineering finance teams.
When organizations push AI adoption through leaderboards, utilization metrics, or mandates to "use AI for everything," they create incentive structures that drive toward maximum token consumption. One major tech company reportedly pushed internal teams to maximize token usage — treating token spend as a proxy for AI engagement.
The result at multiple enterprises has been consumption curves that look like the early days of cloud cost overruns: teams moving fast, building value, but burning through budget at a rate that no one in procurement or finance had modeled for. One large company reportedly consumed its entire 2026 AI coding tools budget in four months.
Gartner analyst Will Sommer put the structural issue clearly: teams should not mistake falling token prices for falling enterprise AI costs. Per-token rates have dropped significantly over the past 18 months. But agentic models require orders of magnitude more tokens per task than the standard models those rate decreases were benchmarked against. Consumption growth is outpacing unit-cost declines across the board.
Goldman Sachs has forecast a 24-fold increase in enterprise token consumption by 2030 as agentic AI scales. If your current cost model does not account for compounding consumption growth, your AI budget will be structurally wrong for every year in that forecast window.
Why Agentic Architectures Drive Exponential Cost
Understanding the technical mechanism helps leaders make better architectural and vendor decisions.
The transformer architecture underlying every major large language model attends to every prior token when generating the next output. Computational cost scales roughly with the square of the context sequence length. A 10,000-token context is not twice as expensive as a 5,000-token context — it is four times as expensive. An 80,000-token context (a realistic multi-step agentic session) is not eight times as expensive as a 10,000-token context. It is 64 times.
This is why multi-step agentic workflows feel economical in testing (short contexts, controlled scope) and become expensive in production (long contexts, open-ended tasks, user-driven scope creep). The cost dynamics are non-linear, and most enterprise cost models are linear.
IBM's July 9 update to its Bob agentic development platform addresses this directly with a subagent isolation architecture. Each subagent receives only the inputs it needs for its specific subtask — a codebase section, a dependency graph, a test suite — and performs its work in an isolated context window. Only the result returns to the main workflow. The accumulated intermediate context is discarded rather than carried forward.
IBM's product page claims approximately 40% reduction in AI compute spend from task-aware model routing combined with subagent isolation. That figure is IBM-stated and has not been independently audited. But the architectural principle it reflects — minimize shared context across parallel workstreams — is sound, and it is the same principle that makes microservices cheaper to run at scale than monoliths.
The Visibility Gap: Why Finance Can't See the Problem
One of the consistent failure modes I see in conversations with enterprise AI leaders is the visibility gap between what AI tools are being used for and what that usage is actually costing at a department or team level.
Most enterprise AI tools have historically given procurement teams a single number: seats. The seat count model is familiar, predictable, and easy to audit. Finance can model it. Legal can negotiate it. IT can track it.
Token-based consumption billing is a fundamentally different model. It requires infrastructure teams to instrument usage, create cost attribution by project and team, set consumption alerts, and build dashboards that connect token spend to business outcomes. Almost none of this infrastructure exists in organizations that adopted AI tools under the seat fee model.
IBM's Bobalytics dashboard — part of the July Bob update — attempts to address this with three visibility layers: an administrator view covering seat usage, consumption metrics, governance controls, and activity logs; a manager view showing team-level adoption patterns and where usage is high but outcomes are unclear; and a cost attribution layer that traces spend spikes to specific projects, workflows, and model invocations.
The IBM VP of Bob Michael Kwok framed the goal clearly: help enterprises understand not only how much AI is being used, but whether it is creating meaningful value. That is the right frame. The question is not "are we using AI" — it is "are we getting $1.40 in value for every $1.00 we spend on tokens."
What Enterprise Leaders Should Do Right Now
The organizations navigating this well share a few common patterns.
Instrument before you scale. Before expanding agentic AI rollouts, get baseline consumption data at the team and project level. Token spend per developer per week, broken down by workflow type, is the minimum you need to model forward costs accurately. Without this baseline, every future budget conversation is guesswork.
Separate the seat fee from the usage cost in your contracts. When negotiating or renewing AI tooling agreements, understand what is included in the seat fee, what triggers overage billing, what the overage rates are, and what controls exist on consumption. Many 2025 contracts were written before agentic features existed. They need to be renegotiated with usage-based billing explicitly modeled.
Match model capability to task complexity. Not every agentic subtask requires a frontier model. Running IBM's Granite small language models for straightforward code completions and routing only complex reasoning tasks to Claude or GPT-4 class models can reduce total token costs by 30 to 50 percent without meaningful quality degradation on the simpler workloads. The IBM Bob architecture makes this routing automatic. Other platforms require teams to implement it manually in their agentic pipeline design.
Build a token FinOps practice. Cloud FinOps became a discipline when cloud cost overruns became a CFO problem. The same transition is happening now with AI token spend. The teams that will manage this well are the ones that treat token budgeting with the same rigor they apply to cloud infrastructure spend: cost allocation by team, project-level budgets, alert thresholds, and regular review cycles.
Set consumption expectations culturally, not just technically. If your organization is driving AI adoption through utilization metrics, you are creating the conditions for a budget crisis. Measure outcomes — features shipped, bugs caught, documentation completed — not token consumption. Adoption leaderboards that reward usage without tying it to outcomes are expensive and counterproductive.
The CFO's New AI Question
For the past two years, the CFO's primary AI question was "what is the ROI on this investment." That question has not gone away. But it now has a companion question that matters equally: "what is the fully-loaded cost of the AI we are already running."
The seat fee on the invoice is no longer the cost. The token consumption behind that seat fee — multiplied by the complexity of the agentic workflows your teams are running — is the cost. And for most enterprise organizations, that number is not currently visible, not currently modeled, and not currently tied to the outcomes that would justify it.
The good news: this is a solvable problem. The tooling for token FinOps is maturing quickly. The architectural patterns that reduce unnecessary token consumption are well understood. The governance frameworks for tying AI spend to business outcomes exist and are being implemented at leading enterprises.
The window for getting ahead of this is open. But the organizations that wait for a budget crisis to start building the visibility infrastructure will spend the next 12 months reacting. The ones that instrument now will spend those same 12 months scaling.
How is your organization tracking AI token spend? Are you seeing consumption patterns that differ from your original cost models? I am hearing consistent themes across enterprise conversations — reply or connect on LinkedIn or X if you want to compare notes.
