How much does an AI agent cost?
An AI agent's cost includes development and the ongoing cost of a working process. A model tariff covers only part of that total: integrations, retries, maintenance and human review also matter. Before comparing proposals, define the result you are paying for.
5 min read
Separate setup and running costs
A setup estimate includes process analysis, data preparation, integrations, implementation, testing and operational handover. Data migration and access configuration may require additional work. Ask which tasks the proposal covers and who owns its dependencies.
Running costs depend on the actual workload. A single user request may trigger multiple model calls, document retrieval and an external service call. Resources spent on an unsuccessful action still count.
| Cost item | Estimation input | Question to resolve |
|---|---|---|
| Development | First release scope and acceptance criteria | Are integrations and failure testing included? |
| Data preparation | Sources, quality and access | Who cleans and maintains the data? |
| Model and tools | Typical task usage, retries and retrieval | Which failed attempts still incur charges? |
| Infrastructure | Load, storage and hosting requirements | What do production and test environments require? |
| Human review | Handover rate and review time | Who handles exceptions, and within what response time? |
| Maintenance | Monitoring, updates and integration support | Where is the boundary between a defect fix and new work? |
Calculate cost per accepted result
For a chosen period, use total processing cost / results meeting the agreed acceptance criteria. The numerator includes failed attempts and human review. Count a result in the denominator only when it meets the rule defined beforehand.
Report automatically completed results separately from those completed with staff assistance. This shows how much work is actually automated. If no result is accepted, a unit cost cannot be calculated; investigate the failures.
Illustrative example, not an IOANN or provider price: 1,000 tasks arrive in a month, processing costs total 10,000 currency units and 800 results pass acceptance. Cost per incoming task is 10; cost per accepted result is 12.5. This example excludes development. Include it separately for a complete estimate.
If you spread development costs across months, state the chosen period and its rationale. Compare several workload and usage-duration scenarios. The estimate supports a decision; it does not guarantee a return.
What changes the estimate
Multiple systems, data correction and high-consequence actions add complexity. Access to the CRM, document quality and write validation can demand more work than the model call itself.
A short demonstration request is a weak basis for a cost estimate. Test long inputs, multiple steps, retries after failures and human handover. In our own product, Inner Voice, we assessed model alternatives using evaluation scenarios defined in advance. This illustrates a testing method; it is not a published client agent cost estimate.
- Which systems must be read or updated?
- Which data is ready, and which needs preparation?
- What counts as a completed action, and who accepts it?
- What workload is expected, and how variable is it?
- Which actions require staff approval?
Compare with the current process
Measure staff time, rework and failure costs on comparable tasks. Use the same definition of a result for the existing process and the pilot. Otherwise, a cheap model reply may be compared with a person's fully completed job.
Hours freed and cash savings are different measures. State what work the team can do with the available time. Treat any sales-growth forecast separately: it requires funnel data and does not follow automatically from faster processing.
An existing product, a conventional integration and custom development should face the same checks on your examples. Compare workflow coverage, human review and full cost, including the transition.
What to send for a development estimate
Describe one process: inputs, systems, expected result and current volume. Include anonymised examples of routine and difficult cases. If baseline measurements are unavailable, say so; the first step can be defining how to measure them.
This brief supports a discussion of first-release scope and the basis of an estimate. IOANN prices work after reviewing the task. A proposal should make its scope, acceptance checks and ongoing costs clear.