What drives the cost of an AI agent project?
The factors that actually determine the build effort and running cost of an AI agent: integrations, data readiness, actions and permissions, evaluation, interfaces and volume, and how to keep it economical.
Cost questions deserve straight answers. We do not publish a price list, because two agents with the same name can differ by an order of magnitude in effort. What we can do is explain exactly what drives the number, so you can judge any estimate you receive.
What drives build effort
- SystemsIntegrations
Each system the agent reads from or writes to adds work, especially systems without a modern API or with complex permissions.
- DataData readiness
Clean, accessible data is fast to work with. Scattered, inconsistent or undocumented data needs preparation before an agent can rely on it.
- ActionsActions and permissions
Read-only agents are simplest. Every action that sends, books or updates a record needs confirmation flows, validation and audit.
- TestsEvaluation
High-stakes workflows need larger test sets, more careful rollout and more review before each change.
- InterfaceWhere people use it
An agent inside an existing tool is lighter than one that needs its own product surface, accounts and design.
What drives running cost
| Driver | What it depends on | How to manage it |
|---|---|---|
| Volume | Conversations, documents or tasks per day | Automate the high-volume path first |
| Model choice | How capable a model each task needs | Route routine work to lighter configurations |
| Context size | Length of documents and histories sent | Retrieve only what each step needs |
| Monitoring and improvement | Time to review failures and refine | Weekly review of logs and escalations |
Two agents with the same name
- Answers questions from one approved help centre
- Read-only, no actions
- Hands off to the existing support inbox
- Lives inside the current website chat
- Reads orders, payments and delivery systems
- Issues refunds and changes bookings with approval
- Routes cases to specialist teams with context
- Works across chat, email and voice
Both are called support agents. The second has several integrations, external actions with confirmations, a larger evaluation set and three channels. That is why a quote without a scope is not meaningful.
How to keep it economical
- Start with one workflow
Choose one high-volume workflow with a clear outcome, rather than a general-purpose assistant.
- Use the smallest permission set that delivers value
Read-only first. Add actions only when the evidence shows they are needed and safe.
- Measure cost per completed outcome
Track the cost of each enquiry qualified or document processed, not cost per request.
- Widen on evidence
Extend scope when the numbers show the current scope works, not on a schedule.
Frequently asked questions
How much does an AI agent cost?
It depends mainly on the number of systems it must connect to, how ready your data is, how many actions it can take and how carefully those must be controlled, and the volume it handles. A narrowly scoped first agent on one workflow is the most economical way to start.
What are the running costs of an AI agent?
Model usage driven by volume and context size, hosting and storage, monitoring, and the time spent reviewing failures and improving the system.
How do you measure the return on an AI agent?
Measure cost per completed outcome, such as an enquiry qualified or an invoice processed, against the cost of the same outcome before. Usage numbers alone do not show value.
Get the next note by email
One email a month at most. Practical, from real work. Unsubscribe in one click.