What an AI agent actually costs to build and run for a service business
By Mariam Raouf. Published 2026-07-22. 8 min read.
Every quote you get for an AI agent is one number, and that number hides two different things.
One is calculable to the cent, today, from published pricing pages. The other has no public benchmark anywhere, and anybody giving you an industry average for it made it up. So this article splits AI agent development cost in half: the running cost with the arithmetic done, and an honest account of what moves the build price.
The running number is usually the surprise. It is smaller than people fear on model costs and larger than people expect on platform metering.
Running cost, part one: the model
This is the bill from whoever makes the model. It is per token, published, and the part everyone worries about unnecessarily. Pricing per million tokens, September 2026:
| Model | Input | Output |
|---|---|---|
| Claude Opus 5 | $5 | $25 |
| Claude Sonnet 5 | $2 | $10 |
| Claude Haiku 4.5 | $1 | $5 |
| GPT-5.6 Sol | $5.00 | $30.00 |
| GPT-5.6 Terra | $2.00 | $12.00 |
| GPT-5.6 Luna | $0.20 | $1.20 |
| Gemini 3.5 Flash | $1.50 | $9.00 |
| Gemini 3.5 Flash Lite | $0.30 | $2.50 |
| Gemini 2.5 Pro | $1.25 up to 200k context | $10.00 |
Sources: Anthropic pricing, OpenAI pricing, Google Gemini pricing. Checked September 2026.
Two levers cut that materially. Anthropic's Batch API is 50% off input and output, and prompt caching reads cost 0.1 times the input rate, which matters for an agent resending the same instructions on every call. OpenAI publishes cached input at roughly a tenth of standard. Google's batch discount is also 50%, and its priority tier costs an 80% premium.
The dated cost cliff you should budget for now
Google's Gemini 3.8 and 3.7 Flash prices are promotional through 31 December 2026 and double on 1 January 2027, going to $1.50 input and $7.50 output. If you are building a business case on Flash pricing for next year, use the post promotional numbers. This published cliff will quietly break a lot of spreadsheets.
Running cost, part two: the platform, which is where it actually hurts
Agent platforms do not bill per token. They bill per action, and their unit of counting is where the surprise lives.
Zapier Agents offers a free tier at 400 activities a month and Pro at $33.33 a month billed annually, $400 a year, for up to 1,500 activities. The definition is the important bit: activities include actions the agent takes, plus browsing the web and looking up attached knowledge. Both burn an activity.
So an agent that researches before it acts might spend three or four activities on one useful outcome. At 1,500 a month, that is closer to 400 real outcomes than 1,500.
Microsoft Copilot Studio is the clearest metering arithmetic in the category. It charges $0.01 per Copilot Credit on pay as you go, billed in arrears:
| Action | Credits | Cost |
|---|---|---|
| Classic answer | 1 | $0.01 |
| Generative answer | 2 | $0.02 |
| Agent action | 5 | $0.05 |
| Tenant graph grounding | 10 | $0.10 |
| Content processing | 8 per page | $0.08 |
| Generative voice | 35 a minute | $0.35 |
Source: Microsoft Copilot Studio message and credit management.
Work one real interaction: a generative answer plus one agent action is 7 credits, so 7 cents. Ten thousand interactions a month is about $700 a month, before a dollar of model API cost. Agents get disabled at 125% of prepaid capacity, so budget deliberately rather than discovering the ceiling on a Monday. Compare that $700 against the model table and you can see where an agent's running cost really comes from.
The cost nobody puts in the quote: agents loop
This is what makes agent economics different from ordinary automation, and it comes from the people who make the models. Anthropic's Building Effective Agents defines an agent as a system where the model "dynamically direct[s] their own processes and tool usage", and describes them plainly as "LLMs using tools based on environmental feedback in a loop."
In a loop. That is the cost.
A deterministic workflow makes one model call per run because a developer wrote the path. An agent makes as many calls as it decides it needs: read the request, pick a tool, look at the result, decide whether that was enough, try again. Nobody can tell you the number in advance, because deciding the number is the agent's job.
Anthropic states the tradeoff directly: "Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense." And more bluntly: "The autonomous nature of agents means higher costs, and the potential for compounding errors."
So multiply your per interaction model cost by a loop factor, and remember the platform meters every one of those tool calls too. Looping plus per action billing is why agent bills come in above the estimate. Our AI automation cost breakdown covers the same metering problem for non agent workflows, where the numbers are friendlier.
What AI agent development cost actually depends on
Now the honest part. There is no credible published benchmark for what a US small business pays to have an AI agent built. No federal survey, no representative study, nothing. Any article quoting "$15,000 to $50,000 for a custom AI agent" is reading one agency's rate card and calling it a market.
I am not going to add another invented range. Here is what genuinely moves the price.
How many systems have to talk to each other. One system with a good API is a small job. Four, one of them a legacy tool with a CSV export and no API, is a different job.
Whether your process is written down. This moves the number more than the technology does. A builder working from a documented process prices a known task. One working from "we sort of just know how it goes" prices discovery, and discovery is where estimates go wrong.
How much scaffolding the agent needs. Anthropic's own shop experiment went from losing money to profitable mainly by adding a CRM, costed inventory data, price comparison and an oversight agent. That scaffolding is the build. The deployment guide for support agents shows it in practice.
What the reliability bar is. An agent drafting internal summaries needs far less testing than one touching customer money. Evaluation and guardrail work scales with consequence, not complexity.
Who maintains it. APIs change, prices change, and somebody owns the thing after launch. If that is not in the arrangement it becomes awkward in month four.
For scale, Gartner puts enterprise generative AI deployment approaches at $5M to $20M. That is nowhere near your number, but it does explain why quotes in this space spread so wide.
The test for whether you need an agent at all
Run this before you price anything. It takes ten minutes and saves a lot of money.
Write the process down step by step. Then ask: can I write the sequence in advance?
If yes, you want a workflow, not an agent. It costs one model call instead of ten, it can meter as one run instead of many actions, it fails predictably, and you can debug it. Anthropic's own guidance is to find "the simplest solution possible, and only increasing complexity when needed", and that workflows "offer predictability and consistency" while agents suit cases where flexibility is genuinely needed at scale.
If the sequence genuinely varies every time, based on things you cannot enumerate, an agent earns its cost. That is a real category, and a smaller one than the marketing suggests. Most of what gets sold as an agent is a workflow with better branding.
How to sanity check an AI agent development cost quote
Calculate the running cost first, because you can, and because it tells you within an hour whether the idea clears its own bar. Steps times runs times loop factor, against your platform's unit. If the answer is $700 a month to handle work costing $400 in staff time, you have your answer and you did not spend anything to get it.
Then treat the build as what it is: mostly integration and mostly your own process documentation. The businesses that get good prices arrive with the process written down, and that is something you can do yourself over a weekend.
At Calpir we build AI agents and the ordinary automation underneath them, and our packages page sets out how we scope work. If you want a straight answer on whether a process justifies an agent, ask, and I will tell you when it does not.
Frequently asked questions
How much does it cost to run an AI agent per month?
It depends on volume and your platform's unit. Zapier Agents Pro is $33.33 a month billed annually for up to 1,500 activities. Copilot Studio arithmetic gives about 7 cents per useful interaction, so 10,000 interactions is roughly $700 before model API costs. Model costs are the smallest line unless you run a premium model at volume.
Why is AI agent development cost so different between quotes?
Mostly because of discovery, not building. A quote against a written process with clean APIs prices a known job. A quote against a vague description prices risk. If you want comparable numbers, write the process down first and send the identical document to every builder you ask.
Is it cheaper to build an agent or buy an agent platform?
Buy first, almost always. Platform subscriptions start around $33 a month and tell you within a fortnight whether the task suits an agent. Build when you hit a real limit: a system with no connector, a compliance requirement, or metering costs above a developer's time.
Do model prices go up?
They mostly fall, but not always, and one increase is already published. Google's Gemini 3.8 and 3.7 Flash rates are promotional through 31 December 2026 and double on 1 January 2027. Any budget running into next year should use the higher figure rather than today's promotional one.
What is the cheapest way to test whether an agent is worth it?
Free tiers and a narrow task. Zapier Agents has a free tier at 400 activities a month, enough to learn how quickly research and knowledge lookups consume quota. Run one real process through it for two weeks and count the activities a single useful outcome takes.