The 5 Ways to Classify a Customer Call That Covers Five Different Products

An hour-long customer call is not one piece of feedback. A single enterprise conversation can carry a complaint about onboarding, a feature request for reporting, a competitor mention, a pricing objection, and a bug in a product the account does not even pay for yet. Filed as one record with one category, four of those five disappear.

Five approaches exist: one category per call, manual splitting by a human, splitting by speaker turn, splitting by topic segment, and multi-label extraction where each distinct piece of feedback is classified on its own. Only the last one preserves everything the call contained, and the difference shows up as missing volume in themes you never knew were undercounted.

Why one call is not one piece of feedback

Most feedback taxonomies were designed against support tickets. A ticket is short, arrives about one thing, and closes when that thing is resolved. One record, one category, and the model works.

Calls break every one of those assumptions. They are long, they are unstructured, they cover whatever the customer wanted to raise, and they contain feedback about products the conversation was not scheduled to be about. A taxonomy that treats a transcript as one record is applying a ticket-shaped model to something with a completely different structure.

The consequence is not a small rounding error. Sales and success calls are where the most commercially significant feedback lives, because that is where the buyer speaks rather than the user. Losing four fifths of it means the channel most connected to revenue contributes the least to your counts. See extracting feature requests and competitor mentions from Gong calls for what that channel holds.

The 5 ways to classify a call that covers several products

1. One category per call

The default in most systems and the worst option. The call gets classified by its dominant topic and everything else is discarded. It is cheap, it is consistent, and it systematically loses secondary mentions, which is exactly where early signals of an emerging problem appear before they are anyone's main complaint.

Best for: nothing, other than as a temporary state while something better is set up.

2. Manual splitting by a human

An analyst or the account owner reads the transcript and files each piece of feedback separately. Accuracy is the highest available, because a human understands context and knows which product each comment refers to. It also does not scale past a few calls a week, and it degrades fast, since whoever is doing it has another job.

Best for: a small number of strategic accounts where the depth justifies the time.

3. Splitting by speaker turn

Each turn in the transcript is treated as a separate record. Mechanically simple and produces a lot of noise, because most turns are conversational rather than substantive, and a single point often spans several turns as the customer explains it. You end up with high volume and low signal, which is a different failure from losing feedback but not obviously a better one.

Best for: teams that need something automated immediately and can tolerate noisy counts.

4. Splitting by topic segment

The transcript is divided into segments where the subject changes, and each segment is classified. Better than speaker turns because segments are coherent, and it struggles when a customer raises a point, moves on, and returns to it twenty minutes later, which is how people actually talk. Related feedback separated by time gets counted twice.

Best for: structured calls that follow an agenda, like QBRs.

5. Multi-label extraction

Each distinct piece of feedback in the call is identified and classified individually, with the call retained as the source record. One transcript contributes to five themes without being counted five times as a call, and each theme keeps a link back to the moment it came from. This is what a ticket-shaped taxonomy is missing when it meets an unstructured channel.

Best for: any team where calls are a meaningful share of feedback, which is most B2B companies.

The cost of single-label classification

The undercounting is not evenly distributed, which is what makes it expensive rather than merely inaccurate.

Feedback that arrives as a secondary mention is feedback about something other than what the meeting was for. Those are disproportionately early signals, cross-product issues, and things customers mention in passing because they have not yet escalated them. The taxonomy is quietly biased toward problems that are already someone's main complaint, which is to say toward problems you already know about.

It also skews which channels appear to matter. If calls contribute one record each while a support queue contributes thousands, the reports will say support is where your feedback is, and that conclusion is an artifact of the classification model rather than a fact about your customers. A team reorganizing its roadmap around that reading is optimizing for the channel that happens to produce short records.

How to check what your system is doing

Take one long call you know well and find it in your feedback platform. Count how many distinct themes it contributed to. Then read the transcript and count how many distinct pieces of feedback it actually contained.

If the first number is one and the second is five, you have your answer, and the ratio is roughly the factor by which your call-derived volume is understated. Run it on three calls before drawing conclusions, since a call that genuinely covered one topic is not evidence of anything.

The second check is whether the theme links back to the specific moment in the call or only to the call as a whole. The difference matters when someone asks to hear what the customer actually said, which they will.

FAQ

Can one piece of feedback belong to more than one theme?

It should be able to, and many platforms do not allow it. A customer describing a problem that spans two product areas belongs in both, and forcing a single choice loses one of them.

How do I handle a call that covers several products?

Extract each distinct piece of feedback and classify it separately, keeping the call as the source record. Anything that classifies the transcript as a single unit will capture the dominant topic and drop the rest.

Why do sales calls contribute so little to feedback counts?

Usually because each call is being treated as one record. A one-hour call and a two-line app store review count the same, which understates the channel where the most commercially significant feedback is spoken.

How does Enterpret classify multi-topic calls?

Enterpret's adaptive taxonomy identifies each distinct piece of feedback inside a record and classifies it individually, so one call can contribute to several themes while staying traceable to the moment it came from. The customer context graph attaches the account and revenue context to each of those pieces, so a passing comment from a large account is weighted accordingly rather than lost.

Does multi-label extraction inflate my counts?

No, if the unit of counting is the piece of feedback rather than the record. The risk to watch is double counting when a customer makes the same point twice in one call, which good extraction resolves to a single instance.

If your call data is underrepresented in your reporting, see tools for auto-categorizing customer feedback.

‍