"How much does an AI agent cost" is really the question "how much judgment are you asking software to exercise, across how many of your systems". An agent that drafts replies for a human to approve is a different project from one that updates your CRM, files the invoice and only escalates the weird cases, even if both get described as "automating our inbox". That is why we scope these individually rather than publishing tiers.
What actually drives the price
- How many workflows, across how many systems. One workflow touching one tool is the floor. Every additional system the agent has to read from or write to adds integration and testing surface.
- How much judgment is involved. Rule-shaped work ("if the invoice matches the PO, file it") is cheap to automate reliably. Work that needs interpretation needs guardrails, and guardrails are engineering.
- What happens when the agent is not sure. A well-built agent has a designed escalation path to a human, not a guess. Designing that handoff, and the review tooling around it, is part of the build, not an extra.
- Whether it serves one team or many customers. An internal automation and a multi-tenant product with billing are different projects, even with the same core logic.
Reference points from our published tiers
Agent work is custom-scoped, but it runs on the same fixed-scope, fixed-price delivery model as everything else we build, so our published tiers for adjacent work are honest reference points for where projects tend to land.
| Adjacent work | Scale of project | Timeline | Published starting price |
|---|---|---|---|
| AI chatbot build | One knowledge source to live actions against your systems | 2–5 weeks | $3,000–$9,000 |
| Web application | Single workflow to complex permissions and reporting | 2–5 weeks | $3,000–$9,000 |
A thirty-minute scoping call produces a written scope, timeline and fixed price before any work starts. The number doesn't move once you've agreed to it, and scope added mid-build is priced honestly rather than absorbed into surprise invoices.
The running cost
Model usage scales with how many tasks the agent handles, not with how many people are on your team. For most workflow automations the model bill is a modest monthly line, because a well-scoped agent run is a short prompt and a short response, not an open-ended conversation. Usage is billed to your own provider accounts at provider rates. We don't resell model access at a markup, which also means you can audit the running cost yourself.
What gets left out of most automation quotes
Our build process has four stages, and the cheapest quotes usually price only the third one.
- Mapping the workflow. Sitting with the people who do the work now and writing down what actually happens, including the exceptions nobody mentions in the kickoff meeting.
- Designing the guardrails. Deciding what the agent may do on its own, what needs approval, and what gets escalated. Skipping this is how automations quietly do damage.
- Building and integrating. The part everyone quotes.
- Monitoring and adjusting. A period after launch where the agent's decisions get reviewed and the rules get tightened against reality. The automation that works in month three is rarely identical to the one that launched.
Want a fixed number instead of a range? Tell us which workflow eats the most hours and we'll scope it on a thirty-minute call.
Book Free Consultation