Software capability is not operational ownership
An agentic service uses AI agents to perform parts of an ongoing business process, often with integrations, monitoring and human escalation included. Customer support, lead qualification and document operations are common starting points. The offer might be a software platform your team operates, a managed service run by the supplier or a combination. Those arrangements carry very different responsibilities.
Intercom Fin and Salesforce Agentforce illustrate product approaches to deploying agents in customer-facing and business workflows. They are not interchangeable outsourced operations. Ask who configures policies, updates knowledge, reviews failed interactions and staffs escalations. Never infer that a product’s agent capability includes a fully managed human service.
A sensible first scope is narrow: answer delivery questions using an approved order API, or classify incoming supplier documents and route exceptions. Avoid an open-ended promise to “replace the support team.” Defined inputs, permitted actions and exclusion rules make both the contract and the pilot easier to evaluate.
Understand what the meter counts
Suppliers may bill per seat, conversation, action, credit or resolved issue. Some combine a platform subscription with consumption charges and onboarding fees. Check current official pricing and obtain a written quote for your expected traffic. An action-based price can look inexpensive until a single customer request requires several searches, updates and follow-ups.
Outcome pricing only aligns incentives when the outcome is defined well. Does “resolved” mean the customer confirmed success, the ticket was closed or the conversation went quiet? Is a reopened ticket charged again? Are spam, test messages and human handoffs excluded? Request a billing export that lets you reconcile charges against your own records.
| Basis | Potential advantage | Question to ask |
|---|---|---|
| Seat | Predictable team budget | Which usage limits apply? |
| Action / credit | Tracks system activity | How many units does a full task use? |
| Resolution | Closer to business value | What qualifies, and who verifies it? |
| Managed retainer | Operational support included | What volume and staffing are covered? |
As an illustrative calculation, 4,000 billable outcomes at £0.80 each cost £3,200 before platform fees and internal labour. If only 3,200 outcomes pass your quality review, the usage cost is £1 per accepted result. Add the cost of correcting failures before comparing the service with your existing process. These figures are an example, not a supplier quotation.
Make accountability visible
Ask for a responsibility matrix covering knowledge maintenance, access control, incident response, approvals and customer complaints. Availability commitments describe whether the service is reachable; they do not necessarily guarantee answer quality. Define a separate quality measure, a reporting cadence and the remedy when performance falls below your agreed threshold.
Review data-processing terms, subprocessor lists, model-provider arrangements, retention and international transfers. Establish whether your content can be used for training and which settings control that use. Require prompt notice of material changes to models or processing locations. Your procurement, security and legal teams should review commitments that affect regulated or sensitive data.
Human escalation needs more than an email address. Specify operating hours, queue ownership, response targets and the context passed to reviewers. For refunds, account changes or financial commitments, define transaction limits and approval rules enforced by the underlying system. The supplier should be able to explain how duplicate actions and interrupted requests are handled.
Buy evidence before scale
Run a time-boxed trial on one channel and a clearly defined case category. Compare against a baseline for correct resolution, handling time, customer satisfaction and cost per accepted result. Randomly audit successful cases as well as failures: automated closures can hide unresolved customers. Include adversarial requests and incomplete records in testing.
The advantages of buying are faster deployment, packaged integrations and access to operational expertise. The disadvantages are supplier dependency, less flexibility and potential opacity around decisions. Building offers greater control but moves engineering, monitoring and staffing costs onto your team. Compare these complete operating models rather than licence fees alone.
Before expanding, test your exit plan. Can you export transcripts, knowledge, configuration and audit records in usable formats? Can staff resume the workflow if the service fails? Agree deletion steps, transition assistance and contract termination conditions. A dependable service should make it possible to leave without losing the process it helped you run.