First-call resolution means the customer’s issue was settled during the first contact without avoidable follow-up. For a Zanus AI call center evaluation, define what settlement means for each call type. A conversation ending politely does not prove that the requested work happened.

Set a resolution rule for each reason

For an opening-hours question, an accurate answer may complete the request. For a booking change, resolution requires the correct calendar change. For an issue needing specialist review, a successful transfer can be useful while the underlying case remains open.

Write these rules with the operators who do the work. Decide how to treat calls that contain two requests. Marking the whole call resolved when one task remains open inflates the result and leaves the customer chasing.

Collect evidence beyond the call log

RequestEvidence of completion
InformationThe approved answer addresses the question
Booking amendmentThe correct event is changed and confirmed
CallbackThe promised contact actually occurs
Service caseThe responsible team completes the case work

Keep unresolved but properly routed calls visible. Routing them correctly is an operational success without being first-call resolution.

Look for repeat contacts with the same cause

Link later calls to the original request or case ID. A customer might ring from another number or reply by email, so a phone-number-only check can miss repeat work. Set a review period that fits the service rather than declaring success immediately after hang-up.

Sample calls marked resolved and ask staff to inspect the destination record. Include a test in which an action fails after the assistant has started speaking. The outcome should reflect the saved result, with an exception routed for repair.

Report resolution without hiding workload

Calculate resolved eligible first contacts divided by eligible first contacts, using your written definitions. Report transfers, repeat contacts, reopened cases and manual repairs beside that rate. Keep excluded calls visible and explain the exclusion rule.

Compare similar reasons and complexity levels before judging AI against staff. An agent answering simple hours enquiries is not facing the same work as a specialist handling complaints. Expand the deployment when the team can trace each claimed resolution to an actual completed customer request.

Questions and answers

Is transferring a caller the same as resolving the request?

No. Track a completed handoff separately. Count resolution only when the underlying request meets the completion rule for that call type.

Defining What "Finished" Looks Like Before the First Call Arrives

Resolution rules written after a deployment has gone live tend to reflect what the system already does well rather than what the customer actually needed. A more defensible approach is to draft the completion criteria for each call type before measuring anything, then test those criteria against real call outcomes rather than adjusting them to improve the reported rate. This section proposes the operational detail a team would need to do that work rigorously.

Start by listing every reason a customer might call. Group them not by how the assistant handles them but by what would have to be true in a downstream system for the customer's need to be met. An enquiry about operating hours is satisfied when the caller receives an accurate, complete answer and does not need to call again for the same information. A cancellation request is satisfied when the relevant record reflects the cancellation and any downstream processes that depend on that record have received the updated status. A complaint requiring a manager's judgment is not satisfied by a promise to call back; it is satisfied when that judgment has been delivered and the customer has been informed. Writing these distinctions in plain language, with examples, gives the team a shared reference that survives staff turnover and is auditable by someone who was not involved in drafting it.

Where a single call contains more than one distinct request, record each separately. A caller who asks for a booking change and also reports a billing discrepancy has two open items. Resolving the booking change does not close the billing item, and marking the call resolved because one item was handled understates the real workload while overstating effectiveness. Consider whether your call logging structure can accommodate multiple request codes per interaction, because without that granularity the resolution rate becomes an approximation rather than a measurement.

A Worked Scenario: Tracing One Enquiry From Receipt to Confirmed Completion

The following is a fictional illustrative scenario, proposed to help a team design their own verification steps.

Imagine a caller contacts the AI assistant to reschedule an appointment from one day to another. The assistant confirms the new time verbally and ends the call politely. From the call log, the interaction appears complete. A reviewer checking only the call transcript would mark this resolved.

Now suppose the reviewer also checks the calendar system. The appointment record still shows the original time. The assistant encountered a write error after confirming verbally, but the session ended without flagging the failure to the caller. The customer believes the change happened. The business believes the change happened. Neither is correct.

This scenario illustrates why the call transcript is a necessary but insufficient source of evidence. A practical verification step would be: within a defined period after the call ends, retrieve the destination record independently and compare its state to what the assistant confirmed. Where those two things do not match, the call should be reclassified as unresolved and an exception record created for human review. The exception record should capture the call identifier, the expected system state, the actual system state, the time gap between call end and system check, and the staff member who reviewed it. That record becomes evidence both for repairing the customer's situation and for diagnosing whether the mismatch is a rare edge case or a pattern requiring a process change.

Carrying this scenario further: the customer, having received no confirmation that anything went wrong, arrives at the original appointment time. They now call again. This second call should be linked to the first through a case or request identifier rather than only through the caller's phone number. If the customer calls from a different device or contacts through another channel, a phone-number-only match will miss the connection. The second call then appears as a fresh first contact, the original failure disappears from the repeat-contact data, and the resolution rate remains artificially clean. The practical counter to this is to ask, during the second contact, whether the caller has been in touch recently about the same matter, and to record the answer before the assistant or agent proceeds.

Structuring the Evidence Audit for Different Request Types

Different call types require different evidence sources and different reviewers. Grouping them by verification method makes the audit manageable rather than ad hoc.

Informational requests

For calls where the customer asked a factual question, the evidence question is whether the answer given was accurate and complete at the time of the call. This requires a reference document or approved answer set against which the assistant's response can be compared. Where no such reference exists, accuracy cannot be measured systematically. Before deploying an AI assistant on informational call types, a team should therefore assemble and approve those reference documents, version them so that outdated answers can be identified retroactively, and establish a process for updating them when the underlying facts change. A reviewer auditing an informational call should record which version of the reference document was current at the time of the call and whether the assistant's answer matched it.

Transactional requests

For calls where the customer requested an action in a system, the evidence is the system record after the call. The review process should specify how long after the call the check occurs, who performs it, and what tolerance exists for timing delays in system updates. If a system update takes several minutes after a call ends, a check performed immediately may show a pending state that resolves on its own. The review window should account for that. Where a transactional request spans multiple systems, each system should be checked independently, because a successful write to one system does not guarantee a successful write to another.

Routed or escalated requests

For calls transferred to a human team, the resolution question splits into two parts. The first is whether the transfer itself was completed correctly: did the right information travel with the call, did the receiving team member have what they needed, and was the customer kept informed about what would happen next? The second is whether the underlying request was eventually resolved, and over what timeframe. These two outcomes should be reported separately. A well-executed transfer is a meaningful operational result and should be credited as such, but it is not the same as the customer's request being closed. Conflating the two makes the AI assistant appear more capable than the end-to-end process warrants.

Designing an Acceptance Exercise Before Expanding Deployment

Before broadening the scope of an AI call center deployment, a team can conduct a structured acceptance exercise using observable evidence rather than reported metrics. The following steps are proposed as a framework, not a prescribed sequence.

Select a sample of calls from each active call type. Choose calls that were marked resolved in the system and calls that were marked unresolved or escalated, in roughly equal proportion. For each call in the sample, retrieve the relevant downstream record and compare its state to the resolution rule written for that call type. Record whether the two match. Where they do not match, note the type of mismatch: was the action not taken at all, was it taken incorrectly, or was it taken but not reflected in the record within the expected timeframe?

Include in the sample at least one call type where an action fails partway through. If your environment allows it, simulate a condition where the assistant has begun processing a request and the downstream system becomes temporarily unavailable. Observe what the assistant communicates to the caller, what record is created, and what the recovery path looks like. The outcome a team should look for is that the customer is informed promptly, the failure is recorded in a way that enables follow-up, and a human reviewer is notified before the customer has to call again. Whether a particular system behaves this way should be verified in your own environment rather than assumed from documentation.

After completing the sample review, calculate the agreement rate between the system's resolution classification and the independent record check. A high agreement rate, where the system marks calls resolved that the record check also confirms as resolved, supports expanding the deployment. A low agreement rate, or a pattern of mismatches concentrated in one call type, is a signal to investigate the classification logic and the downstream integration before expansion.

Handling Exception Cases and Staff Responsibilities

No call routing or resolution process eliminates exceptions. The operational question is not whether exceptions will occur but who is responsible for each type and what the expected response time is. Documenting this in advance prevents the situation where an exception sits in a queue because each team believes another team owns it.

Consider the following categories of exception and the staff responsibilities that could be assigned to each.

  • A call where the assistant was unable to identify the request type: a human reviewer determines the correct category, attempts to address the request if it remains open, and records the reason the assistant could not classify it.
  • A call where the assistant confirmed an action that the system record does not reflect: a human reviewer contacts the customer to correct the situation, records the mismatch, and flags it for technical review.
  • A call where the customer was transferred but the receiving team has no record of receiving it: the receiving team supervisor and the AI deployment team jointly investigate the transfer path and the customer is contacted within a defined period.
  • A call where the same customer has made contact more than once about the same underlying issue within the review period: a senior reviewer assesses whether the original resolution was genuine or whether the resolution rule needs revision.

For each exception category, the record should capture who reviewed it, what action was taken, how long the review took, and whether the underlying resolution rule was adequate or needs amendment. Over time, the pattern of exceptions is more informative than any single resolution rate, because it shows where the process is genuinely working and where it is producing the appearance of resolution without the substance.

Distinguishing the Customer, the Request and the Responsible Party

Resolution tracking works correctly only when these three things are recorded separately and linked rather than collapsed into a single record. The customer is the person or organisation whose underlying need must be met. The request is the specific action or information sought in a given interaction, which may be only one of several things the customer needs. The responsible party is the team or individual accountable for ensuring the request reaches its completion rule.

When a customer calls on behalf of someone else, for instance a carer arranging a service for another person, the customer identity and the beneficiary of the service are different. Resolution of the request means the beneficiary's situation has changed appropriately, not only that the caller received a polite response. Call records should be structured to capture this distinction where it is operationally relevant, because a resolution check that looks only at the caller's record may miss whether the action was applied to the correct individual.

Similarly, when a request involves a professional judgment, such as a clinical team member reviewing a case or a financial specialist approving an adjustment, the AI assistant's role is to gather information and route correctly, not to make the judgment. The resolution rule for that call type should specify that the professional review must occur and must be documented before the request is counted as resolved. Administrative process and professional decision-making are separate steps, and the measurement framework should keep them separate rather than treating the handoff as the endpoint.

Connecting Resolution Data to Operational Decisions

A resolution rate is useful only if it changes decisions. Before investing effort in measuring it, a team should identify which decisions the measurement is intended to inform. Candidate decisions might include whether to add a new call type to the AI assistant's scope, whether to adjust the transfer threshold for a particular request category, whether to revise the training material the assistant uses for informational calls, or whether to change the staffing arrangement for human review of exceptions.

For each of those decisions, identify what pattern in the data would justify making the change and what pattern would suggest waiting. Without that prior agreement, a resolution rate becomes a reporting exercise rather than a management tool, and the team ends up debating what the number means rather than acting on it.

Report the resolution rate alongside the volume of calls excluded from the calculation, the volume of repeat contacts within the review period, the volume of exceptions by category, and the average time between call end and confirmed completion for transactional requests. Those figures together give a more complete picture of operational health than the resolution rate alone. When the resolution rate rises while repeat contacts also rise, the measurement framework probably needs revision. When the exception volume falls while resolution holds steady, the process is likely maturing in a meaningful way. Keeping all of these figures visible in the same report, rather than burying some in supplementary tables, makes it harder for any single metric to create a misleading impression of performance.

Anchor the expansion decision not to a rate meeting a target but to the team's demonstrated ability to trace each claimed resolution to a confirmed completed record, respond to each exception category within a defined timeframe, and update resolution rules when the evidence shows they are producing inaccurate classifications. Those three capabilities, rather than any single reported figure, are what allow a deployment to grow without accumulating invisible unresolved customer requests beneath an attractive-looking summary number.

Sources and next steps

This is an independent Sysevo planning guide, not a report of a hands-on Zanus AI test. Use the official Zanus AI Front Office overview to confirm current product scope and request a demonstration of the workflow described here. The scenarios and acceptance criteria above are proposed evaluation steps.

For the broader topic, read Zanus AI call center fit. If you are comparing providers, explore Sysevo AI voice agents and request a Sysevo demo using the same scenarios. Sysevo publishes this guide and offers its own business software.

Related workflow guides

For current Sysevo subscriptions and usage options, compare plans and pricing.