The LLM Token Trap: Why AI Tooling Costs Escalate

Why This Just Became Procurement's Problem

For decades, procurement has priced software the same way: a seat, a term, a predictable line item on next year's budget. AI has broken that model. When your organisation calls a large language model — inside a coding assistant, a source-to-pay platform, or a customer-facing chatbot — you're no longer paying for a licence. You're paying per token, every single time the model reads a word in or writes a word out. There is no ceiling unless someone builds one into the contract and the operations behind it.

In 2026, two of the most sophisticated technology buyers on the planet found that out the hard way, within months of each other. If Uber and Microsoft can get caught by a pricing model nobody capped, procurement teams with far less engineering muscle behind them are exposed to the same risk — and it's arriving faster than most category plans account for.

What's a token?

The metered unit LLM vendors bill on instead of a seat – roughly three-quarters of an English word, counted on the way into the model as a prompt and on the way out as a response. Every prompt, every generated reply, every automated re-check runs the meter.

Two Warnings From the Frontier

Uber: a year's budget gone in 4 months

Claude Code adoption spread fast across Uber's engineering organisation — from 32% of engineers using it in February 2026 to 84% by March, with 95% on it monthly by spring. Average monthly spend ran $150–$250 per engineer; power users hit $500–$2,000 a month, and Uber's own CTO logged a single two-hour session that cost $1,200. By April, the company had burned through its entire 2026 AI tooling budget, set against a $3.4B 2025 R&D base. Internal leaderboards ranking engineers by usage made it worse — they rewarded exactly the behaviour that was blowing the budget.

Microsoft: licences pulled after 6 months

Microsoft actively encouraged broad employee adoption of Claude Code, then reversed course within six months — cancelling most direct licenses and steering engineers toward GitHub Copilot instead. The trigger wasn't tool quality. It was consumption at scale against a metered bill with no predictable ceiling, echoing a warning that's become common across the AI infrastructure world: for teams running agentic tools, the cost of compute can now exceed the cost of the people using it.

Why cheaper tokens won't save your budget

Frontier model prices are expected to fall roughly 90% by 2030 — but agentic workflows already consume 5–20x more tokens than a simple autocomplete-style call. Consumption is growing faster than the unit price is falling, which is why Gartner's guidance to enterprise buyers is blunt: cheaper tokens will not translate into cheaper enterprise AI.

The 4-Stage Pricing Cycle Every AI Vendor Is Running

It isn't new, procurement has seen this movie before, just with different vendor logos on screen. Cloud storage ran the same playbook a decade ago: cheap entry pricing, then invoices carrying fees nobody saw in the sales deck. A recent Backblaze survey of IT decision-makers found 95% had been hit with cloud charges they didn't see coming. That wasn't an accident — it was the plan. AI vendors are running the sequel, with much bigger numbers attached.

● Stage 1 -> Subsidised adoption. Flat, seat-like pricing designed to get a fast internal “yes” and a clean, predictable line in the budget.

● Stage 2 -> Embed & depend. Once the tool is stitched into daily workflows, removing it becomes organisationally painful. Adoption quietly becomes dependency.

● Stage 3 -> Absorb the real cost. Running frontier models is expensive, and vendors absorb the loss short-term — they're buying future pricing power, not customer goodwill.

● Stage 4-> Shift to usage-based billing. Once switching costs are high enough that leaving isn't realistic, the bill changes shape. Anthropic's move to usage-based billing after July 2026 and Microsoft's $0.01-per-credit Copilot Cowork pricing are the leading edge of this shift across enterprise software broadly — not just chatbots.

Beyond Coding Tools: Where This Lands Next

It's tempting to file Uber and Microsoft under “that's a dev-tools problem.” It isn't. The same metering sits underneath every AI feature vendors are racing to bolt onto source-to-pay suites, intake and orchestration platforms, and contract lifecycle tools. A single multi-step agent reading a contract, checking it against policy, and routing it through approvals can burn hundreds of thousands of tokens in one run — and today, that cost sits on the vendor's API bill, not yours. It won't stay there.

Gartner projects AI agents will pull $234B away from traditional SaaS spend by 2030 as agent-based consumption displaces fixed-scope contracts. Left unmanaged, that means procurement could end up paying for a legacy seat license and an AI usage meter for the same workflow — unless contracts are renegotiated deliberately, before renewal forces the conversation.

The Procurement Response Framework

Nobody has shipped mature token-budgeting or governance tooling for procure tech yet — but it's coming. Until it does, the leverage sits in three places: what you put in the contract, what you build internally, and what you ask before you sign.

Contract Clauses to Negotiate

ClauseWhy It Matters

Usage caps & hard ceilings -> Without a negotiated ceiling or overage alert, a single automation script can become a six-figure surprise.

Cost transparency & reporting -> Require itemized token/usage reporting, not a lump-sum invoice — you can't manage what you can't see.

Pricing model change notice -> Require 60–90 days' written notice before any shift from seat-based to usage-based billing, with a right to renegotiate or exit.

Model change control -> A vendor updating the underlying model can silently change token consumption per task. Require notice and impact disclosure before material model swaps.

Rate-lock / escalation caps -> Where usage pricing is unavoidable, negotiate a ceiling on per-token rate increases for the contract term.

Audit rights -> The right to audit usage reporting and billing methodology — not just the vendor's word for it.

Exit & data portability -> Defined deletion timelines and portability terms cap your exposure if you need to walk away from a runaway bill.

Internal Governance Controls to Stand Up

● Baseline before you budget. Pull 60–90 days of actual usage data before setting next year's AI line item — don't budget off the vendor's demo numbers.

● Departmental quotas. Allocate token budgets by function based on actual workload intensity, not headcount alone.

● Tiered alerts, not just hard stops. Flag usage at 50%, 75%, and 90% of allocation so teams can self-correct before an automatic cutoff disrupts work.

● Chargeback, not overhead. Route AI spend to the department ledger that generates it. Cost visibility changes behavior faster than a policy memo does.

● Model routing. Push routine tasks to smaller, cheaper models and reserve frontier-model tokens for work that actually needs them.

● One written policy. A short, standardized AI usage policy — approved tools, expense tracking, consequences for overages — beats an ad hoc “just be reasonable.”

Vendor Due-Diligence Questions

1. Is this priced per seat, per token/credit, or a blend — and under what conditions does that change?

2. What does a “typical” user cost per month once agentic or automation features are in daily use, not just chat?

3. Can we negotiate a hard usage cap or an automatic pause instead of open-ended overage billing?

4. How much advance notice do we get before a pricing-model change takes effect on our contract?

5. Will you provide itemized, near-real-time usage reporting broken out by team or workflow?

6. When the underlying model itself changes on your side, does that change our cost per task — and do we get advance notice?

7. Can lower-value tasks be routed to a cheaper model tier within the same platform?

8. What's the overage rate, and is it capped or open-ended?

9. If we exceed budget mid-term, is service throttled, paused, or does billing simply continue?

10. Do you have a roadmap for token-budgeting or governance features — or is that entirely on us to build?

30–60–90 Day Action Plan

Days 0–30 -> Pull usage data on every AI-enabled tool already in the stack. Identify which are seat-based, usage-based, or blended today.

Days 31–60 -> Add the clause checklist to your standard AI vendor addendum. Flag any active contract renewing in the next two quarters for renegotiation.

Days 61–90 -> Stand up departmental budget owners, tiered alerts, and a one-page internal AI usage policy. Brief finance on the shift from fixed to metered AI cost lines.