The pilot cost about forty dollars a month and everyone was delighted. Then it went to two hundred users and the bill arrived looking like a car payment. Nobody had done anything wrong. They had just never counted the retries.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, with cost and unclear value near the top of the list. Here is where the money actually goes.
The Four Costs Nobody Budgets For
- Retries. A failed run still costs full price. If one in three attempts fails, your real unit cost is 50% higher than your spreadsheet says.
- Context growth. Long sessions resend an ever larger conversation. Cost rises with the square of the conversation, not with the answer.
- Tool call chatter. Every list, every lookup, every clarification is another round trip.
- Idle supervision. A coordinating agent that adds a model call to every step is a tax you pay on all traffic.
The Metric That Actually Matters
Stop tracking cost per call. Track cost per successful outcome, with failures included in the numerator.
An agent that costs four cents a call but succeeds half the time costs eight cents per result plus the human cleanup. That last part rarely appears in anyone dashboard, and it is often the largest number.
Five Cuts That Do Not Hurt Quality
- Route by difficulty. Most agent steps are classification and extraction, not reasoning. One report found infrastructure costs falling from around $3,000 to $127 a month for standard enterprise workloads after moving high-volume tasks off frontier APIs to purpose-built smaller models.
- Cache the catalogue. The current MCP spec adds
ttlMsandcacheScopehints with deterministic ordering, so tool lists stop being re-fetched and upstream prompt caches stay stable. - Compact deliberately. Decide what must survive a summary, and drop the rest instead of resending it forever.
- Cap the loops. Hard limits on hops and retries turn an unbounded bill into a bounded one.
- Fail fast and loudly. An agent that gives up after two bad attempts is cheaper than one that tries eleven times and still escalates.
The Latency Cost Nobody Prices
Latency is a cost with a different currency. A support agent that takes ninety seconds to answer is one your customers route around, so you pay for the tokens and keep the human queue too.
Set a latency budget before you design the workflow. Then choose your orchestration pattern to fit it, rather than discovering afterwards that a debate pattern triples both cost and wait.
Conclusion
Agent economics are not mysterious, they are just uncounted. Measure cost per successful outcome, route cheap work to cheap models, cache what does not change, and put hard caps on loops. Do that and the bill becomes predictable, which is the only condition under which a finance team lets a project reach year two.
Frequently Asked Questions
Is a cheaper model always worth it?
No. A smaller model that fails more often can cost more per successful outcome. Route by task difficulty, then measure, rather than switching wholesale.
How do I attribute cost to teams?
You have to build it. MCP has no defined multi-tenancy or cost attribution model yet, so tag calls at your own gateway using the method and name headers.
What is a reasonable target?
Cost per resolution below the fully loaded cost of the human path, with retries and escalations counted honestly. If you cannot calculate that, you are not ready to scale.