Your Agents Did Not Blow the Budget. Your Architecture Did.

· 4 min read ai-finopsagentsenterprise-ai

Navy and amber graphic with the headline Your Architecture Spent the Budget

KPMG’s Q3 2026 AI Pulse Survey landed this week with two numbers that belong on the same slide. Multi-agent systems are now in use at 25 percent of large enterprises, up from 6 percent across all of 2025. And roughly 93 percent of participants blew past their AI budgets. Scale is arriving four times faster than the controls to pay for it. KPMG’s Q3 2026 AI Pulse Survey

McKinsey’s estimate, quoted in the same coverage, explains part of the gap: agent-based workflows are five to 30 times more computationally intensive than chatbot queries. That is a wide band, and the exact number does not matter. What matters is that architectural choices are now the primary cost lever as inference spend scales. The model you picked is not the budget line. The chain of loops, retries, and tool calls around it is.

The fence goes up after the money is spent

A few weeks ago I wrote that every agent needs a budget, enforced in code at the execution layer. I still believe that. The budget is the fence. But a fence does not change what the animal eats. Per-task budgets stop the bleeding. They do not fix the burn rate, and 93 percent overrun tells me most shops are designing the burn first and noticing it later.

The overspend is decided at design time, in choices that look innocent on a whiteboard:

  • One more agent in the chain. Every hop multiplies token burn linearly, because each agent re-reads the context it was handed and writes a full reply.
  • The planner running on the flagship model. Planning is the cheapest work in the loop, and teams run it on the most expensive model out of habit.
  • Unbounded retries and tool calls. The agent decides its own retry policy, and the default policy of every agent I have met is “try again.”
  • Full context passed down the chain instead of a summary. Context bloat is where the money quietly goes.

None of these are billing problems. Nobody can negotiate their way out of them. They are design problems, and they compound: multi-agent adoption quadrupled in a year, which means the compounding is quadrupling too.

Cost-shape the chain before you budget it

The fix is to shape cost at architecture time, when the choices are still cheap to make. A 30x cost multiplier sounds abstract until you price one task: a support triage that costs two cents as a chat transcript costs sixty as an agent loop with three retries and two tool calls. Multiply by ten thousand tickets a month and the 93 percent stops looking like a rounding error.

  1. Split the roles across model tiers. The planner thinks, the investigator gathers, the executor acts. Only the executor needs the heavy model, and even that is per task, not per everything. This is the planner/investigator/executor separation I have used in agentic IT operations design: it stops the failure mode where the model acts without proving anything, and it stops the budget failure mode where the model reasons expensively about cheap work.
  2. Budget tokens in zones, not in one lump. Session memory gets a three-zone budget with eviction rules per zone: keep what the task needs, summarize what it merely referenced, drop the rest. Whole-context handoffs between agents are the silent killer of per-task cost.
  3. Cap the loop, cap the tools, name the kill criteria. Maximum iterations, maximum tool calls per task, and a stall rule: if the agent cannot reach a coherent result within the configured cost budget, it stops and escalates instead of burning more. That rule belongs in the agent contract, in writing, the same way I put owner and error-budget discipline in SRE agent contracts.
  4. Meter per agent, per chain. Cost per completed task for every agent in the chain, including retries and tool calls, measured in production. The budget post covers the method. The point here is where to aim it: the chain, not the model. Audit the chain the way you would audit a data pipeline that suddenly costs four times what it did last year.

Gartner projects more than 40 percent of agentic AI projects will be cancelled by the end of 2027, naming unclear value and governance failures as the causes. I expect a large share of those “governance failures” will turn out to be cost failures with a better-sounding name. Nobody cancels a pilot that pays. They cancel the one whose invoice from the API provider looks like a typo, and they call it governance.

Put the budget in the harness, yes. But before that, shape the architecture so the budget is not fighting a losing battle from the first token.