What Does an AI Agent Actually Cost Per Month?

Ask five vendors what an AI agent costs and you will get five different shapes: a seat licence, a consumption estimate, a build fee, a retainer. None answers the question that decides whether a programme survives a finance review: what will it cost every month once live? Agents are not priced like software seats. A seat costs the same whether it is used or not. An agent behaves more like a process: it consumes models when it works, infrastructure when it waits and human attention when it fails. The monthly bill moves with volume, complexity and how much of the stack you run yourself. Three pricing shapes dominate. Platform licences charge per user or workspace — easy to budget, with a ceiling on flexibility. Consumption pricing meters the work itself — tokens, minutes, documents — scaling with value but needing guardrails. Managed delivery converts engineering effort into a service fee — predictable until you need to leave. The one-off build side is covered in what an AI agent costs to build; this article covers the number that repeats.

The monthly number is rarely one line. It is a stack of lines, each with a different growth behaviour:

• Model usage. Every conversation, lookup and tool call consumes tokens. The line grows with volume unless you engineer it down through caching and routing.
• Platform and orchestration. The runtime that routes tasks, manages memory, connects tools and enforces policy — usually a licence per seat or usage.
• Infrastructure and hosting. Containers, queues, databases and vector storage. Cloud cost discipline applies directly: published benchmarks put cloud waste at around 29% of spend — an agent stack without capacity planning adds to it.
• Knowledge maintenance. Retrieval indexes must be refreshed as documents, prices and policies change — which is why RAG implementation services are continuous work, not a one-off project.
• Channels. Voice deployments add telephony and speech minutes — a real variable cost for teams building AI voice agents.
• Evaluation and observability. Scoring runs, red-teaming, logging and spend dashboards. Small relative to the stack, and the first thing cut — usually a mistake.
• Human review. High-stakes decisions still need a person in the loop, part-time or full-time.

Four levers explain most of the variance between two companies running similar agents:

• Volume. The gap between a pilot's few hundred interactions a month and production's tens of thousands is the biggest difference between the demo bill and the real one.
• Steps per task. A single lookup is cheap; a ten-step workflow that reads six systems, reasons over the results and writes back is not.
• Autonomy. Agents that confirm every action spend human attention instead of rework; fully autonomous agents invert that trade.
• Knowledge freshness. Fast-moving content needs frequent re-indexing — a line that never appears in pilot estimates.

None of these is fixed. Scope tightly, bound the tools each agent may touch, and assign someone to keep the bill down.

Pilots are priced to start, not to stay. Vendors subsidise proofs of concept because volume is low and scope is narrow, and approvals come easier when the number is small. Production removes every subsidy: monitoring, redundancy, security review, integrations with real systems of record, and traffic that does not stop at 5pm. This is also where most programmes stall. More than 50% of GenAI pilots are abandoned before they reach production — rarely because the model underperformed, usually because the surrounding work had no owner. An abandoned pilot is a fully sunk cost: build fee, pilot subscription and internal hours, all gone.

• Steady volume: a licence is usually cheapest and easiest to forecast.
• Spiky or unproven volume: consumption pricing keeps you honest, but set spend limits before the first surprise invoice.
• Scarce engineering attention: managed delivery buys patterns and maintenance you would otherwise learn by failing — check who owns the data and the exit.

The wrong answer is a licence sized for a pilot, a meter with no cap, or a contract with no exit. Ask for production pricing, and ask what the bill looks like at three times current volume.

• Count interactions, not users. Estimate steady-state interactions per month, then stress-test at triple that figure.
• Split fixed from variable. Platform licences and base infrastructure are fixed; tokens, minutes and review hours scale with volume.
• Price the governance lines. Evaluation, monitoring and human review are cost lines, not overhead.
• Add integration upkeep. Agents read and write to systems of record, and those systems move: with SAP ECC support ending in December 2027 and roughly 60% of enterprises already migrated, integration often rides along an active platform programme.
• Set a review cadence. Monthly early on, then quarterly, with optimisation targets attached.

Predictable does not mean cheap; it means planned. One owner for the run cost. One dashboard, reviewed monthly. Optimisation targets you actually pursue: cache frequent answers, route simple requests to smaller models, retire workflows nobody uses. The organisations that keep agent bills under control run them the way FinOps runs cloud spend — engineering discipline applied to a meter that never stops. If you want a cost model built against your volumes and systems, talk to our team. For the wider view of what production agents involve, start with our AI agents practice; for customer-facing deployments, see AI agents for customer support.