What should AI agent pricing include in 2026?
Price the outcome, but bill the work. A defensible AI agent price has three parts: a setup fee for discovery, integration and guardrails; a monthly platform fee covering hosting and monitoring; and a usage or outcome component tied to a metric the client already tracks. Avoid promising revenue. Tie the variable piece to resolved conversations, booked appointments or hours saved, and cap it so a good month does not create a surprise invoice.
Which pricing model fits which buyer?
Small local businesses usually prefer a flat monthly fee because it is easy to budget. Mid-market teams accept per-resolution pricing when the baseline is measured before launch. Enterprises want a platform fee plus committed usage and a clear overage rate. Outcome-only pricing sounds attractive but shifts risk to you; use it as a pilot structure with a floor, not as the default contract. Whatever the model, publish the unit and the measurement method in writing.
What are the minimum guardrails before launch?
Five layers: scope limits so the agent only answers in-domain questions, retrieval grounding so answers cite your content, refusal rules for pricing and legal or medical advice, escalation to a human with context attached, and logging of every conversation. Add rate limits and a kill switch. Test with 50-100 adversarial prompts before go-live, including attempts to extract instructions or invent policy. If the agent cannot say it does not know, it is not ready.
Who owns the agent and its data?
The buyer should own the prompts, knowledge base, integration code and conversation logs; the vendor should own nothing that cannot be exported. Ownership matters most at renewal and exit. Insist on a documented export path, retention settings and deletion on request. If the agent runs on a third-party model, the client still owns the surrounding system. Write this into the contract, not the sales deck.
How do you govern an agent after launch?
Run a monthly review with four numbers: containment rate, escalation rate, resolution time and cost per conversation. Sample 10% of transcripts weekly for tone and accuracy, and review every escalation tagged as a failure. Keep an incident log with named owners and response times. Set a review date for model or prompt changes, and require approval before expanding the agent's scope. Governance is a calendar, not a document.
What does a 90-day agent rollout look like?
Weeks 1-2: baseline the current process and pick one workflow. Weeks 3-5: build retrieval, integrations and guardrails; test with real transcripts. Weeks 6-8: pilot with a small user group and a human in the loop. Weeks 9-12: widen scope only if the numbers hold, then agree pricing and support terms. Hold back 20% of budget for the fixes that only appear in production. Agents improve with curation, not with bigger launches.
Frequently asked questions
Should I charge per conversation or per resolution?
Per resolution aligns better with value but needs a clean definition. Per conversation is simpler to meter. Many teams start with a flat fee and add usage later.
How much should guardrails cost?
Treat guardrails as 15-25% of build effort, not an afterthought. They reduce incident, escalation and support costs, and they make pricing defensible.
Can clients switch model providers later?
Yes, if prompts, retrieval configuration and logs are portable. Design for export from day one and avoid provider-specific features that cannot be replicated.
Want the full playbook?
The free 2026 AI Revenue Playbook turns these shifts into a plan with costs, scripts and pricing templates.