Proscendia

What does it cost to build an AI agent?

Prices reviewed August 2026 · USD · excludes tax

A working pilot that does one real job on your own data costs $2,400 to $4,500 and takes one to two weeks. A production build of a single agent across two or three systems is typically $8,000 to $22,500 over three to six weeks. Agents spanning several systems with approval steps and compliance requirements run $22,500 to $52,000. Running costs after launch are usually $50 to $450 a month in model and hosting spend. Most agencies will not publish these numbers; there is no good reason not to.

Build cost by scope

ScopePriceTimeWhat you get
Discovery callFree30 minAn honest read on whether an agent is the right tool, and a rough figure. No charge, no deck.
Pilot$2,400–$4,5001–2 weeksOne real workflow, working end to end on your own data, so you can watch it before committing to more.
Production build$8,000–$22,5003–6 weeksA hardened agent across two or three systems: permissions, audit trail, error handling, monitoring and handover.
Multi-system build$22,500–$52,0006–12 weeksSeveral systems, approval steps, role-based access and the evidence trail a regulated process needs.
Support retainer$800–$2,500MonthlyMonitoring, model updates, and changes as the work changes. Optional — you own the code either way.

What it costs to run, after launch

This is the number almost nobody publishes, and it is the one that decides whether an agent is worth keeping. An agent is not a one-off purchase: it consumes model tokens every time it works, and it needs somewhere to run.

ItemTypical monthlyWhat drives it
Model tokens$25–$320How often the agent runs and how much it reads each time. A daily agent over long documents costs more than an hourly one over short messages.
Hosting$12–$75Usually a small container. Agents that idle most of the day cost very little on scale-to-zero hosting.
Third-party APIs$0–$150Anything the agent calls that bills separately — search, enrichment, transcription, e-signature.
Typical total$50–$450Most single-workflow agents sit at the lower end. We estimate this before the build, not after.

What moves the price

  • How many systems it touchesOne system in, one system out is the cheap case. Every additional system adds authentication, error handling and a new way for things to go wrong.
  • What a mistake would costAn agent drafting internal summaries needs far less scaffolding than one emailing customers or moving money. Approval steps, rollback and audit are where the engineering time goes.
  • How messy the input isClean structured data is quick. PDFs, scans, forwarded email chains and free text take longer to handle reliably — and reliability is the whole job.
  • Whether the process is written downIf the rules live only in one person’s head, part of the work is extracting them. That is normal, and worth doing regardless of whether you build the agent.
  • VolumeVolume barely changes the build cost and directly changes the running cost. High volume usually improves the business case rather than worsening it.

When it is not worth building

If the process is fixed and predictable, a Zapier or n8n automation will do it for a fraction of these numbers and your own team can maintain it. If the task happens a handful of times a month, the maths rarely works — an agent replacing twenty minutes a week saves roughly seventeen hours a year, which will not repay a $12,000 build. We will say so on the call rather than quote for it.

Compare custom agents against Zapier and n8n if you are not sure which side of that line you are on.

Common questions about cost

Why do most agencies not publish prices?

Because ranges look risky and they would rather qualify you on a call first. The trade is that you cannot tell whether they are in your budget without spending half an hour finding out. Publishing ranges costs a few unqualified enquiries and saves everyone else time.

Is it cheaper than hiring someone?

For a single workflow, usually yes. A production build at $8,000 to $22,500 is a fraction of a junior hire’s first year, and it does not take holiday. For continuous, judgement-heavy work that changes weekly, a person is still the right answer.

What is the cheapest way to start?

The pilot, at $2,400 to $4,500. It exists so you can see one real workflow running on your data before deciding whether to spend more, and if it does not convince you, you stop there.

Do you charge for the discovery call?

No. Thirty minutes, no deck, and if an agent is the wrong answer for your problem you will be told on that call.

Are these fixed prices or estimates?

The ranges are indicative; the quote you get after the call is fixed for the agreed scope. Changing the scope changes the price, which is agreed in writing before work restarts rather than appearing on an invoice.

Who pays the model and hosting costs?

You do, directly to the provider, on your own accounts. That keeps the running cost transparent and means you are not paying a margin on tokens — and if you leave, nothing needs migrating.

What if the agent needs changes later?

Small changes are covered by a retainer if you want one, at $800 to $2,500 a month. Otherwise changes are quoted as work. You own the code, so you can also have your own developers do it.

Every number here is a range because scope genuinely varies. Thirty minutes on a call turns it into one fixed figure, and there is no charge for that conversation.