How Much Does an AI Agent Cost a Business

The listed price isn't the price you pay

When asked how much an AI agent costs, the first answer that comes up is almost always a monthly subscription. That's incomplete. An autonomous AI agent in production adds up several cost line items that never appear on the same invoice: model inference, the platform subscription, integration time, human oversight, and ongoing maintenance. EY, in an analysis of enterprise token costs, sums up the problem well: these total costs stay structurally invisible, because they're scattered across several budgets that never get consolidated in one place (source: EY, "Agentic AI enterprise token costs," https://www.ey.com/en_us/insights/ai/agentic-ai-token-costs, accessed 2026-09-04).

Line item 1: LLM inference, the price of tokens

This is the most visible line item, and the most misunderstood. Every bit of reasoning an agent does (planning, deciding, calling a tool) consumes tokens, billed by model providers per million tokens, with a different price for input (what you send the model) and output (what it produces). Here's the public pricing on display as of September 4, 2026.

On the Anthropic API (source: Claude, pricing page, https://claude.com/pricing, accessed 2026-09-04): Claude Haiku 4.5 is billed at 1 dollar per million input tokens and 5 dollars for output. Claude Sonnet 5 is 2 dollars for input and 10 dollars for output. Claude Opus 5, the most capable model, is 5 dollars for input and 25 dollars for output.

On the OpenAI API (source: OpenAI, pricing documentation, https://developers.openai.com/api/docs/pricing, accessed 2026-09-04): GPT-5 is 1.25 dollars for input and 10 dollars for output. GPT-5.4 is 2.50 dollars for input and 15 dollars for output. GPT-5.5 is 5 dollars for input and 30 dollars for output.

These figures give you an order of magnitude, not a final budget: an agent that plans, calls several tools, and reviews its own results can burn through tens of thousands of tokens for a single task. EY illustrates this drift with an example it presents itself as an illustration, not a verified market average: a customer conversation that cost 0.04 dollars in 2023 rose to roughly 1.20 dollars in 2026, a 30-fold increase driven by the growing complexity of agent orchestrations (same source as above). This is exactly what makes the BYOK option appealing to businesses that want to stay in control of this line item: it lets you connect your own API key from an inference provider directly, the BYOK option, with billing that then follows the chosen provider's pricing rather than a credit allowance.

One point often overlooked in this calculation: the price per million tokens says nothing about the actual volume an agent consumes. A support agent that answers a simple question in a single pass uses very little. An agent that plans a task, calls three tools, reviews its results, and revises its own draft can multiply that volume by ten or twenty on a single request. It's this mechanic, more than the provider's listed rate, that explains why the inference bill for an agent in production often comes as an unpleasant surprise to those who only budgeted from the catalog price.

Line item 2: the platform and slots

On top of inference comes the cost of the platform itself: the environment that hosts the agent and manages its permissions, its tool connections, its history, and its observability. This is not the same thing as token pricing, it's the price of the infrastructure that makes the agent usable and manageable day to day, rather than a bare script calling a model API. On Atako, this line item is billed per active agent, with a monthly allowance of inference credits included. The exact tier breakdown is on the pricing page. This is a common model among agent platforms: you pay for capacity, how many agents can run at once, not for the number of employees using them.

Line item 3: integrating with existing tools

An agent that only talks to itself has zero value. It needs to read a CRM, write into a ticketing tool, post in a chat channel. Every connection requires configuration: setting up API or OAuth access, precisely defining what the agent is allowed to do, not just what it can reach, then testing that it works on real cases. This work is routinely underestimated when a project is first scoped, even though it directly determines whether the agent is actually useful or stays a gadget that performs well in a demo but never touches the business's real systems.

This line item varies enormously depending on how many tools need connecting and how complex they are. Wiring an agent to a single, well-documented system with simple API-key authentication takes a few hours. Multiplying connections, each with systems that require their own setup (subdomain, region, object-specific permissions), pushes this line item up much faster than the tool count alone would suggest. This is one of the arguments for a platform that has already industrialized a broad catalog of ready-made connectors, rather than rebuilding every integration in-house.

There's also a hidden recurring cost inside this line item: every connector needs to be re-tested whenever the underlying tool changes its API, its permission model, or its object schema, which happens more often than teams expect on fast-moving SaaS products. A connector that worked perfectly at launch can silently start failing months later if nobody is watching for it.

Line item 4: human oversight

An autonomous agent isn't an agent with no oversight. The most sensitive actions, sending a cold email to a customer, publishing public content, editing a billing record, deserve a human checkpoint before they execute, at least in the early stages. This oversight time carries a real cost, often in the form of employee hours spent reviewing, approving, or correcting the agent's work. It normally decreases as trust builds, but it never fully disappears: governance and guardrails are part of the total cost, not an optional add-on. This is also one of the factors Gartner cites to explain why more than 40% of agentic AI projects are expected to be abandoned by the end of 2027: underestimated governance and risk-control costs from the outset (source: Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027, accessed 2026-09-04).

Line item 5: ongoing maintenance

A deployed agent is never finished. The instructions it's given need adjusting whenever a process changes. Connectors break when a third-party provider changes its API. Activity logs need watching to catch drift before it turns into an incident. This maintenance line item is the one project owners forget most systematically when budgeting, because it doesn't show up at launch: it's paid month after month, in the form of engineer or admin time keeping an eye on what's running.

Build it yourself or go through a platform

Faced with this list, two paths exist. Build in-house: assemble the model calls, permission management, observability, and connectors yourself, with a dedicated technical team. Or go through a platform that has already industrialized these building blocks, and charges a per-agent subscription rather than a custom engineering project. The choice mostly depends on how many agents you're targeting and how rare your use cases are: the more standard the needs (support, HR, monitoring), the less custom-building is justified. This comparison is covered in more depth in a dedicated article. Either way, before deciding, it's worth having a realistic estimate of the five line items above rather than a single subscription price splashed across a marketing page.