Published ranges: simple assistant $10,000 to $30,000; agent that takes actions in your systems $30,000 to $150,000; multi-agent or engine $100,000 to $500,000 and up. Mid-size operating businesses mostly land in the second band.
$500 to $20,000 a month: model usage, hosting, monitoring, and the person or firm operating it.
Number of systems connected, whether it acts or only drafts, evaluation and testing, data quality, who operates it.
A per-seat desk is quoted at a fixed price from a two-week assessment. Engines carry a monthly fee because Keystone operates them.
The ranges, by kind of agent
| Kind | What it does | Published 2026 build range |
|---|---|---|
| Assistant on your documents | Answers questions from your files and manuals; no actions | $10,000 to $30,000 |
| Task agent with integrations | Reads email, files, and one or two business systems; drafts outputs; a person approves | $30,000 to $150,000 |
| Agent that acts | Sends, submits, updates systems on its own within limits | $75,000 to $250,000, plus the governance to make it safe |
| Multi-agent system or engine | Several agents in sequence on a continuous supply of work | $100,000 to $500,000 and up |
Those bands are what US development firms publish and what search engines quote. They are wide because they cover everything from a one-person shop to an enterprise vendor with a sales team. Most mid-size operating businesses that need one seat's admin work handled land in the second band, at the lower end when the work is well-defined.
What moves the number
- Integrations. Each system the agent reads (email, project platform, accounting, CRM) is a connection to build and keep working. Read-only connections cost a fraction of write access.
- Draft or act. An agent that drafts and waits for a person is far cheaper and far safer than one that acts. The approval gate removes an entire class of failure modes and the engineering to guard against them.
- How much is deterministic. Most of a good agent is not AI. Collecting, matching, formatting, and routing are plain software, written once and verified. Agents only where judgment is required. Firms that make everything an LLM call pay for it in testing and in reliability.
- Evaluation. A golden set of real examples and a way to score the agent against it. Skipping this is how projects reach a demo and never reach production.
- Data quality. If the source documents are inconsistent, the extraction layer costs more.
- Who operates it. An agent needs someone watching the queue, the errors, and the model changes. In house, a vendor, or the builder.
What a per-seat desk costs
Keystone builds agents as desks: one seat, read-only connections into the systems that seat already uses, a memory that compounds with every job, one play per task, and a gate on everything that leaves. A desk cannot send, submit, or sign.
The price is fixed per desk and quoted from a two-week assessment that counts the hours the seat loses today, so the number is set against a return you can check. The second desk costs less than the first because the connections and the memory carry over. Engines, which run continuously on their own supply of work, carry a monthly fee because Keystone operates them. There is no hourly rate. Ask for the current figures.
Running costs
Model usage for a desk handling one seat's daily queue is usually tens to a few hundred dollars a month at 2026 prices. Hosting on a small dedicated server is similar. The larger running cost is operation: someone reviewing errors, updating plays as the business changes, and re-testing when models change. Published ranges for all of that together are $500 to $20,000 a month, with a single-seat desk at the bottom of that range.
When you should not build one
- A vendor already sells the job as a product. Takeoffs, scheduling, phones, CRM.
- The work is not repetitive. Agents pay back on the queue that arrives every day, not on the project that happens twice a year.
- Nobody in the company owns the seat. A desk supports a person; without one it has no gate.
- You cannot describe what a correct output looks like. If the person in the seat cannot grade the draft, the agent cannot be evaluated.