The software fee is only one line of the cost. An AI customer support agent also costs model usage per conversation, integration with your systems, rewriting the knowledge base, building and running tests, and the time of the people who review its drafts and take its handovers. The useful figure is cost per contact resolved without recontact.
What are the cost lines?
Platform or software. Usually a subscription, sometimes priced per conversation or per resolution. Read the definitions carefully: "resolution" may mean the conversation ended, not that the problem was solved.
Model usage. Each conversation uses model capacity, charged by volume of text processed. Long conversations, large retrieved documents and voice cost more than short text exchanges. This line scales with traffic.
Integration. Connecting the agent to your order system, CRM, billing and ticketing. A one-off build cost plus maintenance whenever those systems change.
Knowledge work. Rewriting help content so an agent can use it, and keeping it current. Often underestimated, and ongoing.
Testing and certification. Building a case set, replaying it, reviewing failures, and repeating after every change.
People in the loop. Staff who approve drafts, take escalations and sample conversations for quality. This is not a failure cost; it is part of running the agent safely.
Monitoring and records. Storing conversation logs and verdicts for the period your policy or regulation requires.
What drives the total up or down?
| Driver | Effect on cost |
|---|---|
| Contact mix | Simple, repetitive questions are cheap to automate; complex ones need more people in the loop |
| What the agent may do | Actions need integration, limits and more testing than answers |
| Channels | Voice costs more per conversation than chat or email |
| Knowledge quality | Poor content means more rewriting and more escalations |
| Escalation rate | Each handover is a person's time; see the escalation matrix for customer support agents |
| Review effort | Drafts that people heavily edit cost nearly as much as writing from scratch |
The review effort line is worth measuring directly. If staff change most of the agent's drafts, the agent is not yet saving their time, and the measure that tells you so is described in how to measure how much reviewers change an AI agent's drafts.
How should you plan the budget?
- Start from your current cost per contact, by contact reason, including staff time, tools and overheads.
- Estimate which reasons the agent will resolve, which it will draft for, and which stay with people.
- Add one-off costs: integration, knowledge rewrite, initial testing.
- Add running costs: platform, model usage at expected volume, review time, escalation time, quality sampling, retesting.
- Compare cost per resolved contact, agent against people, per reason. Resolved means no recontact on the same issue.
- Track it monthly per agent, as described in how to track the cost of each AI agent.
An illustrative example
Take a mid-sized retailer considering an agent for order tracking, returns and billing questions.
Order tracking is cheap to automate: short conversations, one read-only integration, clear answers. The cost per resolved contact falls well below the cost of a person. Returns need write access to the returns system and a rewritten policy, so the one-off cost is higher, but volume makes it pay back. Billing questions turn out to escalate often, because many involve disputes. The agent still helps there by collecting the facts and drafting a note, but the saving is small.
The retailer launches tracking and returns, keeps billing at draft depth, and reviews the figures after three months. The cheapest decision in the plan was the one not to automate everything.
What costs appear when things go wrong?
The hidden cost of a cheap agent is the clean-up: refunds issued wrongly, promises the business has to honour, complaints and recontacts. An agent that is cheaper per conversation but generates rework can cost more overall. That is why the measure should be cost per resolved contact, not cost per conversation.
General running costs across all agent types are covered in how much it costs to run an AI agent, and finance-specific costs in how much an AI finance agent costs.
Where Pryme fits
In Pryme, every verdict is filed per agent and per capability, so cost can be attributed to the contacts the agent resolved, held or refused. That makes cost per resolved contact a figure you read from the record. More is in the AI support agent that knows when to say no.
Questions people ask
Is pricing per resolution better than per conversation?
It aligns the vendor's incentive with yours only if "resolution" means the problem stayed solved. Check how it is defined and who decides.
Will the agent's costs fall over time?
Knowledge and testing costs usually fall after the first months; model usage scales with volume. Review effort should fall as the agent earns more depth.
Should we count the cost of people who take handovers?
Yes. Handovers are part of the service the agent provides. Leaving them out flatters the agent.