The honest short answer is that an AI agent I build for a business starts at $25,000, after a Proof Sprint that starts at $10,000 and is credited in full toward the build. Running it usually costs far less than people expect. The agent on my own site costs me about half a cent each time someone uses it, emails included.
The rest of this post explains where those numbers come from, what drives them up, and when you should not build an agent at all.
Why the prices you find online are all over the place
Search this question and the first page is mostly agency guides. Their ranges run from about $5,000 to well past $400,000 (ProductCrafters and Appinventiv, both checked November 2026). Those ranges are wide because "AI agent" covers everything from a chatbot that answers FAQs to a system that reads your email, decides what to do, and updates five other tools.
A range that wide cannot help you budget. What helps is knowing which parts of the work cost money, so you can tell which end of the range your project sits on.
What you are actually paying for
Connecting it to the tools you already run
An agent that only talks is cheap. One that acts has to read from and write to your inbox, CRM, accounting system or database, and every one of those connections needs permissions, error handling, and a plan for when the other system is down.
At Nordic Global I built features for an internal AI resume platform for healthcare IT consultants. Who could see which profiles came from Microsoft Entra roles plus Workday data, and I rolled that access control out in three modes. Off, then shadow, where it only recorded who would have lost access, then enforce. That staged rollout is real work, and it is a big part of why connecting to an HR system costs more than plugging in an API key.
Your real data
Demos run on clean examples. Your data has scanned PDFs, missing fields, three date formats and the one customer who writes in all capitals. Getting an agent to handle that is most of the work, and it is the part a demo hides.
Proving it gets the job right
Before anyone pays for a full build, I run the agent on a sample of real work and score how often it gets the job right and where it fails. That scorecard is the main thing a Proof Sprint delivers. Without one, nobody knows whether the agent is ready, including the person who built it.
Permissions and a record of what it did
If the agent touches company documents, it has to show each person only what they are cleared to see, and keep a record of every answer. On that same platform I built the audit reporting, with before and after views of every change and exports for review. The small details add up. A hidden record has to look exactly like a missing one, so nobody can probe for what exists, and an export has to be safe to open in a spreadsheet.
I do the same kind of work at Fox, on a newsroom production platform shared by many shows. Much of the job is making sure each show only ever touches its own data, and that the AI features can only be used the way they were meant to. Hundreds of tests check that isolation on every change. None of it shows up in a demo, and all of it is in the price of a build you can trust.
What it costs to run each month
Running cost depends on how many times the agent runs and how much text it reads and writes each time. Here is a real example, the free AI project planner on this site. It reads a description of a job, drafts a Proof Sprint plan, and emails it.
| Item | Per visitor |
|---|---|
| Model, about 385 tokens in and 330 out | $0.0041 |
| Email, the plan plus three follow ups and my alert | $0.0013 |
| Total for 1,000 visitors | $5.32 |
Measured in November 2026 on Azure OpenAI at list prices. Hosting and the database ride on plans I already pay for.
Most business agents cost more than this because they run more often and read more, a long email thread or a contract instead of a paragraph, so the model bill grows with use. The bigger ongoing cost is a person keeping it working when your systems change.
Running Answerhall, my own product, taught me where the money really goes. Most of it is spent at the start, building the pipeline, and then on refactoring it to cut what each answer costs wherever I can. Once it was built, hosting was the small part. In October 2026 the app, its admin, build and Shopify services, the image registry and storage came to about $25 on Azure. The model and document reading bills sit on an account I share with my other projects, so I won't pretend to split out Answerhall's share.
The costs most quotes leave out
- Upkeep when a connected tool changes its API or your process changes.
- Model updates. A new model version can change answers, so the scorecard gets rerun.
- Someone owning the edge cases the agent hands back to a person.
I include 30 days of support after launch. After that, a retainer starts at $3,000 a month for teams that want me on call.
How I price an agent
- We talk for about thirty minutes about the job, the data and the timeline.
- A Proof Sprint, about three weeks, from $10,000. You get a working version on your data and a scorecard.
- A fixed price and date for the full build, from $25,000, with the sprint fee credited in full.
- An optional retainer after the 30 days of included support.
After our call, you get a scope and a price within 48 hours. You own the code, the data and the scorecard.
When you should not build one
Some jobs do not need AI. When I gave my planner a sample job, a support team answering order status questions all day, it said plain automation in the store would do it cheaper, and it was right. If your job is a lookup with clear rules, build the lookup. Save the agent for work that needs reading and judgment.
If you are not sure which one you have, describe the job in the planner or tell me about it.