Choose the first voice AI intents with evidence
Choose a first voice AI intent with enough eligible calls to evaluate, stable policy, accessible systems, limited exceptions, manageable customer risk, and a completion event your team can verify. Use the six-factor worksheet below to compare candidates before you commit to a pilot.
Define the workflow with call data
Start with a recent period that reflects the queue you intend to test, and record the dates you use. Review call reasons alongside transfers, repeat contact, and complaints. Read transcripts or quality notes to separate tasks that share a broad label. A billing queue might include balance enquiries and disputed charges, which need different controls. Record the exceptions that require a person and remove those calls from your eligible-volume estimate. Ask your analytics owner whether the remaining sample can support the decision you need within the proposed evaluation period.
Scope a specific task, such as a balance enquiry after the caller completes the agreed identity checks.
Compare candidates
This worksheet is a proposed planning method. It doesn't describe measured Butter Labs results or confirmed product capabilities. Record evidence and an owner for each row; leave an unanswered question open until the responsible team resolves it.
| Factor | Evidence to collect | Decision before a pilot |
|---|---|---|
| Eligible volume | Call counts after exclusions, with dates and queue boundaries | Analytics owner confirms enough calls for the evaluation |
| Policy stability | Current instructions and known exceptions | Operations owner defines the permitted response |
| System access | Required reads or actions and their test results | Engineering confirms access and failure handling |
| Exceptions | Examples of ambiguous requests and failed steps | Quality owner specifies a route to a person |
| Customer risk | Consequences of a wrong answer or action | Responsible service owner accepts the controls or excludes the task |
| Completion evidence | A system event or customer outcome the team can check | Analytics owner defines success, repeat contact, and the observation window |
An unresolved customer-risk or system-access question should block a candidate even when its volume is attractive. Keep that gate explicit instead of averaging it into a single score.
Test the boundaries
You might assess an outage update or appointment confirmation as an initial candidate, then test its authentication needs and exceptions. Treat these as workflow examples. Confirm every required capability with the proposed vendor before you include it in scope. A payment or account change needs an agreed confirmation step and tests for rejected or duplicate actions.
Give complaints, bereavement, suspected fraud, and requests that need discretionary judgement a clear route to a trained person. Your service owners should define the handling rules for those calls.
Agree the decision gate
Bring operations and engineering owners together with quality reviewers and a frontline agent; include security specialists where the workflow requires their review. Map the expected path, system failures, and handoff conditions. Choose one or two bounded intents, identify who owns each unresolved question, and agree what evidence permits expansion or requires a stop. Use the pilot playbook to record that evaluation and cost per completed task to connect it to finance.
Butter Labs is a candidate for UK telecommunications and utilities teams that want to evaluate a bounded voice AI workflow within an existing contact-centre stack. Ask us to review the proposed intent and its evidence gaps. Confirm system support, handoff behaviour, and commercial scope with the relevant owners before you agree a pilot.
Read the contact-centre stack transformation guide or review the enterprise voice AI evaluation criteria.
Reviewed 14 September 2026