When an AI voice agent marks a call resolved and the customer rings back the next day, what should that do to your first call resolution figures? Andreas Kromann, CEO and founder of Spirelight, explains how to read the second call.
A customer calls an office supplier for a copy of their August invoice. The voice assistant finds it, says it has been sent to the email address on the account, and the customer says thank you. The dashboard marks the call resolved. The next morning, the same customer calls again.
Should the first call lose its resolved status? Most first call resolution (FCR) reports would say yes, because they count any call back within a set window as a repeat. But the honest answer depends on what the second call was for, and with a voice assistant there is no agent to ask how the first call went.
Why repeat calls are hard to read on a voice agent
As Joseph Kim argued in CSM’s recent article on resolution rates, avoiding a handoff to a human doesn’t prove the agent did the right job. Containment, the share of calls the automated system finished without a transfer, is a weaker test than first call resolution, which also asks whether the customer had to come back.
On the phone, the gap between the two is easy to miss. A caller who has given up sounds much like one who is satisfied: they say thanks and hang up. A voice agent can also tell a caller something has been sent when its request to the billing system failed, or act on a number it misheard. VAmoS Bench, a voice agent benchmark published in July 2026, grades each simulated call against what the agent did in the backend as well as what it said, which is how it catches an agent that claims to have changed a customer’s card without updating the record.
Your contact centre doesn’t have a grader reading the backend after every call. It has the customer’s next contact, and that is worth reading carefully.
Five versions of the next morning’s call
Before you look at the second call, write down the job from the first one in a single sentence: give the customer a usable copy of the August invoice. Defined that way, an email leaving the system is only one piece of evidence that the job was done.
Now imagine five versions of the next morning.
“I still haven’t received the invoice.” The delivery log shows the request failed. The assistant said “sent” when the email had only been requested. The original need is still open, and the cause is an action the agent confirmed before it was complete.
“You sent me the wrong invoice.” The customer read out invoice number 4015. In the recording, “forty fifteen” was transcribed as “forty fifty”, and the assistant sent invoice 4050, an older invoice on the same account. The need is still open, and this time the cause is hearing.
“The email arrived, but the attachment won’t open.” The assistant did its part. The customer still can’t use the invoice, so the request is unfinished, but the fix belongs to whoever runs the document system.
“I got August’s. Could you send July’s as well?” The first job was done. Same account, same document type, same channel, but this is a new request.
“I’m calling about my invoice.” This could be any of the above, or a question about the amount. The right label for now is unclear, followed by a clarifying question. Guessing doesn’t make the dashboard more accurate.
The first three are failed resolutions, each with a different owner: the voice agent’s confirmation logic, its speech recognition and the document system. The fourth is a new request, and the fifth can’t be judged yet. A raw repeat call count treats all five the same way.
Why a raw repeat call rate gets this wrong
Contact centres don’t agree on what a repeat call is. In one Call Centre Helper reader thread, a team found its repeat figures didn’t match a colleague site’s: one counted two calls within 24 hours as a repeat, the other needed three. An agent at one of the “big 4” wireless carriers described a three-day rule: “even if a customer we helped calls back within 3 days for a completely different reason, it is still held against us.”
Apply that rule to the July invoice and the voice agent is marked down for good service. Apply a 24-hour window, and a customer who notices two days later that the invoice never arrived is never counted.
Matching causes problems too. A telephone number can help you find a record, but shared numbers and family accounts mean it shouldn’t decide on its own whose request the second call is about. Link contacts through the verified account and the request, under your existing access rules, and send anything ambiguous to a person to check.
A five-step repeat call review
You can run this with the systems you already have and a few hours of a team leader’s time each week.
- Pick one call type and one window. Start with a single recurring request, such as invoice copies, and a fixed window, such as seven days, which lines up with the bounce-back measure in Marija Jovanović’s earlier CSM article. Only review requests whose full window has passed. A request closed yesterday hasn’t had the same chance to come back as one closed last week.
- Write a short record for each pair. Note the job in one sentence, what the assistant promised, what the logs show actually happened, what the customer wanted next time and who owns any fix. For the broken attachment, that reads: usable copy of the August invoice; assistant said it was sent; delivery log confirms it went out; customer can’t open the attachment; billing support to send an accessible copy.
- Label the next contact. Use four labels: same need (still unresolved); related next step, such as asking how to submit the invoice to an expense system; new request; and unclear. The original request sets the boundary. If the customer asked for an invoice to upload to an expense system and the format won’t upload, it is the same need. For every same-need case, add the cause: hearing, action, answer (the agent gave wrong information) or downstream system.
- Have two colleagues label the same cases independently. Where they disagree, settle what “done” means for that job against the original request before you turn any of it into a target.
- Report by request, with raw counts. Three calls about one missing invoice are three contacts but one unresolved request. Report contacts separately as workload. Show the unclear cases and any contacts you couldn’t match, and break results down by language and line type. In τ-Multilingual, a September 2026 benchmark of 4,500 simulated voice agent calls, task completion stayed within 3.2 points of English in Spanish, Brazilian Portuguese and Hindi, but fell by 14.7 points in Korean and 8.4 in Mandarin. A blended figure would have hidden that.
What the labels do to the numbers
Here is an invented week with round figures. The voice agent takes 1,000 invoice requests and transfers 100, so containment is 90%. Of the 900 it closed, 180 are followed by another contact from the same customer within seven days. Count all 180 as failures and the agent’s first call resolution is 72%, with the 100 transfers counted as not resolved by the agent. If staff resolved some of those on the same call, the contact centre’s own FCR is higher, so judge the transfers separately on whether each handoff was right.
Now label the 180. Say 72 (40%) are the same need, 36 (20%) are related next steps, 54 (30%) are new requests and 18 (10%) are unclear. First call resolution is 828 of 1,000 (83%) if none of the unclear cases is a failure, and 810 (81%) if all of them are. The range is more honest than a single figure, and it tells you exactly which 18 contacts to check next.
One thing the review can’t do is prove satisfaction from silence. A customer who never calls back may have solved the problem some other way, or given up.
Turn confirmed failures into test calls
Each failure you confirm is also a test case. Write down what the caller wanted, what they said and what the backend should show at the end of a good call, then run those scenarios before every change to a prompt, model or speech provider.
Failures that started with hearing, like the “forty fifteen” invoice, need audio to reproduce. A transcript shows what the agent got wrong, but not the accent or line noise that caused it, so you need recordings that sound like your callers, on phone lines like theirs. Real customer calls carry personal data, so check what your consent and privacy terms allow before you test on them.
At Spirelight we record voice agent test audio for this job: private test sets voiced by native speakers and built around a team’s callers, languages and channels, kept out of any training data we deliver so the scores stay honest. The speech side has metrics of its own, from accuracy on numbers to how quickly the agent answers, and we cover those in a guide to voice agent evaluation.
A resolved label is a decision made with the evidence available at the time. The customer’s next call can add evidence that changes it. A good review lets the team correct the label and say why, while leaving work that was actually finished marked as done.
Questions about repeat calls and AI voice agents
How do you measure first call resolution for an AI voice agent?
Link each request the agent closed without a transfer to the same customer’s next contact on any channel within a fixed window, such as seven days. Label each next contact as same need, related next step, new request or unclear. Only same-need cases count as failed resolutions, and unclear cases are reported as a range rather than guessed.
How do you reduce repeat calls from a voice assistant?
Fix the cause your review found. If the agent misheard, have it read back invoice and account numbers before it acts, and test it on audio that sounds like your callers. If it confirmed actions that hadn’t finished, have it check the result in the backend before telling the caller. If customers keep coming back with the same related next step, add the answer to the end of the call.
About the author
Andreas Kromann is the CEO and founder of Spirelight, a Danish company that records and transcribes speech data for voice AI teams, including test audio for voice agents. SpeechData.ai, a Spirelight company, licenses conversational speech datasets in 60 languages and dialects.
This is a Sponsored Article.