AI budgets fail at a startling rate: 79% of organizations experienced AI-related cost overruns in the past 12 months (Sapio Research for DoiT, February 2026, 500 finance leaders at 1,000+ employee companies, US/UK), and only 15% of companies forecast AI costs within 10% of actual — while 80% miss by more than 25% (Benchmarkit with Mavvrik, 2025, 372 enterprises). The cause is structural, not sloppiness: teams budget AI the way they budget SaaS — per seat, flat rate — while AI agents bill by consumption. The fix is a consumption-based forecast plus a short list of engineering guardrails, both below.
Those are enterprise surveys, and that should worry a smaller business more, not less: those companies have FinOps teams and dedicated tooling and still miss. An SME running AI agents on a usage-based bill has the same mechanics with none of the slack.
Last verified: August 10, 2026.
What is token shock?
Token shock is the moment a usage-based AI bill decouples from the budget that approved it. Traditional SaaS pricing is a flat, per-seat subscription: 10 seats cost the same in a quiet month and a busy one. AI agents consume tokens — units of text processed — and every conversation, retry, tool call, and document fed into context burns them. Success makes the bill grow: the pilot that delighted customers at 200 conversations a month costs something entirely different at 20,000.
The DoiT survey has a finding that makes the point sharply: organizations with the most mature FinOps practices reported the highest overrun rates — 89% among leading-edge teams, with a mean overspend around 31%. Maturity doesn't prevent the surprise; it just detects it. Prevention has to happen in how the system is designed and bought, which is good news if you haven't built yet.
Why AI budgets miss
Five mechanical drivers, all invisible in a seat-based budget:
- Agentic loops multiply consumption. A "single" customer question can trigger multiple model calls — planning, tool calls, retries, summarization. The user sees one answer; the meter sees six requests.
- Context grows over time. Longer conversations, bigger knowledge bases, and richer prompts all increase input tokens per request — costs drift upward with no code change and no decision anyone remembers making.
- Model choice is a 10–100x price spread. Frontier models and small models differ by orders of magnitude per token. A system that silently defaults everything to the biggest model pays the frontier rate for work a small model handles fine.
- Overage tiers hide in SaaS AI tools. Many "flat-rate" AI products carry conversation-volume or credit ceilings with steep overage pricing — the flat rate is flat until the month it isn't.
- Nobody owns the number. In the DoiT survey, accountability for AI spend splits almost evenly between technology and finance leadership — a seam where overruns live, and where a 37% share of respondents also report finance and engineering can't even agree on what AI success means.
How to forecast AI agent costs: the consumption model
Build the forecast from the unit of work up, not from last month's invoice down. Five steps:
- Define the unit of work. One resolved support conversation, one qualified lead, one processed invoice. Everything gets priced per unit.
- Measure tokens per unit — including the invisible calls. Run 20–50 real examples through the system and take the full trace: input + output + tool calls + retries. This number is the heart of the model; guessing it is how the 80% miss.
- Apply the per-model price. Tokens per unit × price per token, at the model mix you actually route to (see guardrails below).
- Model volume in scenarios, not a single number. Price the expected case, then 3x, then 10x — growth is the goal, so the budget should already know what success costs.
- Add an honest buffer and a review date. Retries, context drift, and prompt changes push real consumption above the trace. Start conservative, then re-measure monthly against actuals until forecast and invoice converge.
An illustrative example (round numbers, not a quote): a support agent that averages 15,000 tokens per resolved conversation at a blended $3 per million tokens costs about $0.045 per conversation — roughly $45 at 1,000 conversations/month, $450 at 10,000. The arithmetic is trivial; what makes it a forecast is that the 15,000 came from measured traces and the volume came from scenarios. Our AI agent cost calculator runs this model interactively, platform fees included.
The guardrails that keep the forecast true
| Guardrail | What it does |
|---|---|
| Hard budget caps + alerts | Spending alerts at 50/80/100% of the monthly budget, and a hard cap that degrades gracefully instead of billing infinitely. Every major AI provider supports this; most deployments never turn it on. |
| Model routing | Small, cheap models by default; escalate to frontier models only when the task needs it. This is the single biggest cost lever in agent design. |
| Caching and batching | Prompt caching for repeated context and batch processing for non-urgent work — provider-level discounts that exist precisely to reward predictable usage. |
| Per-agent cost attribution | Every agent logs tokens per task, so the invoice decomposes into units of work. "AI spend went up" becomes "the invoicing agent's context doubled in March." |
| Monthly finance–engineering review | One owner, one meeting: forecast vs. actual per agent, unit costs trending, caps adjusted. This closes the accountability seam the surveys keep finding. |
What to ask any AI vendor before you sign
Token shock most often arrives through a tool, not an API. Five questions that surface it in the sales call:
- Is pricing per seat, per usage, or hybrid — and what exactly counts as usage (conversations, credits, tokens, minutes)?
- What happens at the volume ceiling: overage billing, throttling, or a forced tier upgrade — and at what rates?
- Can we see per-task consumption logs, or is usage a monthly black box?
- Can we set hard spending caps ourselves?
- If we leave, what do we keep? (If the answer is "nothing," the subscription is the ceiling on your leverage.)
That last question is why we build the way we do: PxlPeak builds AI agents with cost observability designed in — per-task logging, model routing, caps — and the client owns the system, so the economics are inspectable instead of bundled into someone else's margin. If you're weighing a subscription tool against building your own, our pricing comparison and is-it-worth-it breakdown cover the tradeoff, or talk to us about your volume math.
Is AI automation still worth it at usage-based prices?
Usually yes — when the unit economics are measured, not assumed. The per-unit numbers are typically small (cents per conversation); the overruns come from unmeasured multiplication, not from AI being expensive. Price the unit of work, compare it to the loaded cost of the human-hours it replaces or augments, and the ROI question becomes ordinary arithmetic instead of a leap of faith.
Sources
- DoiT — Why 79% of Enterprises Overspent on AI (Sapio Research, February 2026, n=500 finance leaders, 1,000+ employee orgs, US/UK, ±4.4pp)
- DoiT — The AI Spend Reality Check: full survey (2026)
- Mavvrik — 2025 State of AI Cost Management (Benchmarkit with Mavvrik, n=372 enterprises)