Buy an outcome, not a demo.
Managed AI can remove operating work. It cannot remove the need to define who is responsible.
An agentic service combines software that can take actions with an operating promise: resolve support requests, qualify leads, process documents or maintain records. The provider may supply models, integrations, monitoring and human fallback. Your procurement job is to discover exactly which of those responsibilities are included.
1. Make “done” a contractible event
Start with a narrow service boundary. “Automate customer support” is too broad. “Handle delivery-status questions for domestic orders, using the order system and approved policy, with escalation for refunds” is something you can test. Define eligible inputs, excluded cases, authorised actions and evidence of completion.
A closed ticket is not necessarily a resolved issue. For outcome-based billing, ask whether customer abandonment counts, whether reopened cases are charged again, and how repeat contacts are attributed. Set a review window and a dispute process for incorrectly classified outcomes. Otherwise a provider can optimise its dashboard while your customers do extra work.
Products such as Intercom Fin and Salesforce Agentforce illustrate different entry points: a support-focused AI offering versus agent capabilities within a broader business platform. Neither product name tells you the full operating model. Implementation, human support, integrations and usage allowances depend on the purchased package and agreement.
2. Compare commercial units carefully
| Billing unit | Potential advantage | Question to ask |
|---|---|---|
| Seat or monthly licence | Predictable base spend | What usage and integrations are excluded? |
| Conversation or action | Spend follows activity | Do retries and failed actions count? |
| Verified resolution | Closer alignment with value | Who verifies resolution and handles disputes? |
| Managed retainer | May include operational support | Which staffing and service levels are guaranteed? |
Model the entire month using your own traffic distribution. As a hypothetical example, 4,000 verified outcomes at £0.80 each cost £3,200. Add a £1,000 platform charge and 60 hours of internal review at £30 per hour, and the operating total is £6,000 before implementation, tax or exceptional support. These figures are assumptions for planning, not vendor prices.
Compare that total with a baseline that includes quality. Faster handling is not a saving if incorrect answers increase complaints or create remediation work. Ask for volume tiers, overage alerts, spending caps and the cost of a seasonal peak. Avoid minimum commitments based only on a vendor's estimated automation rate.
3. Draw a responsibility map
Assign an owner for knowledge accuracy, access permissions, model changes, exception queues and incident response. Your business normally remains responsible for its policies and customer relationships. The provider should state who investigates failures, how quickly they notify you, and what logs you can inspect without buying an additional tier.
Distinguish platform uptime from service quality. An available API can still generate incorrect actions. Use separate measures for availability, successful task completion, escalation latency and unauthorised changes. Require a rollback plan when prompts, models or integrations change, and preserve a tested manual route for critical work.
Review the data-processing agreement, subprocessor list, retention controls and cross-border transfer terms. Ask whether your data is used for model training, which settings govern that use, and whether staff can access conversation content. A security certificate is useful evidence, but it does not answer whether the configured workflow respects your permissions.
4. Pilot the exceptions, then practise leaving
Run a time-boxed pilot on representative cases, including ambiguous requests, missing records, angry customers and systems that time out. Agree success criteria before launch. Use independent review of a random sample rather than accepting the provider's automated success label as ground truth.
Keep the first scope reversible. A draft response is easier to correct than a changed account balance. Expand only after the service handles exceptions reliably and your team can operate the fallback without the vendor present. A successful pilot includes evidence that humans receive the right context when the agent stops.
Finally, test export and deletion. Can you retrieve conversation history, configurations, audit records and knowledge sources in usable formats? Which identifiers survive migration? Negotiate transition support and deletion confirmation before renewal pressure makes those questions urgent.
Write the resolution definition, responsibility map and exit checklist before requesting a quote. A serious provider should welcome all three.
Explore current product scope directly at Intercom Fin and Salesforce Agentforce. Obtain written commercial terms for your use case.