Token prices keep falling, yet businesses keep getting surprised by AI bills. The reason is simple: an agent does not make one call. It makes many, in loops, often re-sending everything it has read so far. Add retries, tool calls and a weekend nobody is watching, and a pilot budgeted at a coffee a day can arrive as an invoice nobody approved. Here is how the bill really grows, and how archCTO keeps it boring.
Why agent bills grow faster than you expect
Most AI providers bill for every token sent and received. In a chat, that is one exchange. In an agent run, each step usually re-sends the instructions plus everything gathered so far. Ten steps do not cost ten times one call; the history is billed again and again, so cost grows more like a triangle than a line.
The usual suspects behind surprise bills
Industry reports of runaway agent costs share a few patterns. None of them need exotic technology to fix; they need limits that are designed in, not added after the invoice.
- Loops: two agents passing work back and forth, or an agent retrying an error forever
- Context bloat: whole documents or inboxes re-sent on every step
- Cache misses: a changing timestamp in the instructions stops cheaper cached reads
- Over-powered models: the most expensive model used for simple sorting
- Tool storms: hundreds of API or search calls per run with no cap
- No alerts: monthly billing tells you after the money is spent
The bill you do not see: everything around the model
Even a well-behaved agent has costs beyond tokens: hosting and orchestration, logging and storage, test runs, and above all the time people spend reviewing drafts. For small firms, these fixed costs are often larger than the model bill. A business case that ignores them will look great in month one and disappointing in month six.
How archCTO keeps costs predictable
Cost is designed in the Engineer phase and reviewed every month in Track. Before we build, we count calls per run and estimate cost per run; after go-live, cost per run sits on the same report as time saved.
| Control | What it does |
|---|---|
| Caps per run | Maximum steps, tool calls, time and spend; when hit, the work goes to a person |
| Right model per step | Small, cheap models sort and extract; stronger models only where judgement pays |
| Cached instructions | Stable rules and playbooks reused at a lower rate instead of re-sent |
| Summarised history | Agents carry a short summary, not the whole thread |
| Query, do not stuff | Structured data is queried; only documents are searched and quoted |
| Budgets and alerts | Monthly limits at the provider plus alerts on unusual spikes |
| Use what you pay for | Start inside existing AI subscriptions before adding pay-per-use APIs |
Questions to ask any AI provider before you sign
Whoever builds your agents, ask these five questions. If the answers are vague, the bill will not be.
- What is the estimated cost per run, and how was it calculated?
- What caps stop a loop or retry storm?
- Who gets an alert when spend spikes, and how fast?
- Which costs are outside the model bill?
- Whose account is billed, and can we see usage ourselves?
A document-processing agent with no limits
An agent summarised every attachment in a shared inbox and retried failed downloads. Over a long weekend it re-read the same large files hundreds of times; nobody saw it until the invoice.
Redesigned with a per-run spend cap, a three-retry limit, cached instructions, a cheaper model for classification and a daily usage alert. The same work now costs a small, predictable amount each month.
AI agents are cheap to run when cost is a design requirement and expensive when it is a surprise. Ask for cost per run, caps and alerts before you ask for features.
Questions we hear
How much does it cost to run an AI agent for a small firm?
It depends on volume and design. Small workspace-based agents can run within existing subscriptions plus light usage; custom cloud agents add hosting. archCTO always estimates monthly run cost before building and reports it after.
Are cheaper models good enough?
Often, for sorting, extraction and routing. We test each step on your own past cases and only pay for a stronger model where it measurably improves the result.
Sources and further reading
- Gartner (June 2025): over 40% of agentic AI projects predicted to be cancelled by end of 2027
- Anthropic documentation: prompt caching
- OpenAI API pricing (check current rates before budgeting)
Figures from third parties are quoted as reported by their publishers. archCTO diagrams are illustrative.