BlogsIdeas for teams building with agentic AI
All articles
Conversation Intelligence

How to Measure CSAT and NPS From Call Recordings With AI

AI can identify satisfaction signals in calls, but sentiment is not a survey response. Learn how to connect recordings with CSAT and NPS without creating misleading metrics.

Illustration for “How to Measure CSAT and NPS From Call Recordings With AI”

Understanding how to measure CSAT and NPS from call recordings with AI starts with one distinction: AI can extract an explicit survey answer or estimate satisfaction from conversation signals, but these are not equivalent. A reliable programme keeps reported scores separate from inferred indicators and uses recordings to explain what drives customer experience.

What CSAT and NPS actually measure

Customer satisfaction, or CSAT, usually measures satisfaction with a particular interaction, service or purchase. A survey asks customers to rate their experience. CSAT is commonly calculated as satisfied responses divided by valid responses, expressed as a percentage, with the definition of “satisfied” determined by the survey scale.

Net Promoter Score, or NPS, measures stated likelihood to recommend an organisation, product or service. On its standard zero-to-ten scale, respondents are grouped as promoters, passives or detractors. NPS is the percentage of promoters minus the percentage of detractors.

These measures answer different questions. A caller might appreciate an agent’s help while remaining unwilling to recommend the organisation because of repeated service failures.

A recording can contain evidence relevant to both measures, but a positive conversation does not automatically establish either score.

Three ways AI can use call recordings

Before selecting a platform, decide which measurement task you need. Calling every output “AI CSAT” makes it difficult for decision-makers to know what they are seeing.

ApproachWhat AI doesAppropriate reporting label
Extract explicit answersIdentifies a response to a clearly asked CSAT or recommendation questionRecorded survey response
Analyse experience signalsIdentifies frustration, appreciation, effort and unresolved concernsConversation-derived experience indicators
Estimate survey outcomesUses a validated model to estimate how a caller might answer a surveyPredicted CSAT or predicted recommendation category

The first approach captures something the customer actually said. The second interprets the interaction. The third makes a prediction that needs testing against real survey responses.

If a caller says “ten out of ten” about an agent’s helpfulness, that is not an NPS response. The question, scale and subject of the answer must all match the intended measure.

What AI can learn from a conversation

Speech processing can turn recordings into transcripts and distinguish customer speech from agent speech. Analysis can then identify evidence about the experience, subject to recording quality, language coverage and contextual limitations.

Useful signals include:

  • Stated satisfaction: explicit appreciation or dissatisfaction directed at a specific experience.
  • Resolution status: whether the customer confirms the issue is resolved or leaves with an unanswered concern.
  • Customer effort: repeated explanations, transfers, confusion or difficulty completing a task.
  • Emotional trajectory: whether frustration eases, persists or increases during the call.
  • Relationship context: references to recurring failures, lost trust or intentions to leave.

These signals should remain distinct. “Thank you for trying” can be polite without indicating satisfaction. A calm caller can still be dissatisfied, while an upset caller may be satisfied with the agent’s handling of a difficult situation.

For background on the underlying analysis, see what speech analytics can reveal from customer calls.

A defensible measurement workflow

Define the outcome before analysing calls

Specify whether you want to understand satisfaction with the interaction, resolution of the issue or the wider customer relationship. Document the survey wording, response scale, eligible population and reporting period.

Also define what counts as insufficient evidence. A model should be allowed to return “unknown” rather than assign a confident-looking score to every call.

Connect recordings with actual survey responses

For prediction, you need examples where the conversation and a subsequent survey response can be linked appropriately. The link should identify the relevant interaction without exposing unnecessary personal information.

Timing matters. A survey answered after further contact may reflect a later experience rather than the recorded call. Likewise, relationship NPS may be influenced by pricing, product reliability or past experiences that are absent from the recording.

Survey respondents may also differ from non-respondents. A model trained on respondents does not automatically represent everyone who calls.

Validate against unseen examples

Assess the model on conversations it did not learn from. Compare predictions with actual responses and examine where it overstates or understates satisfaction.

For predicted CSAT, check whether customers assigned a similar likelihood of satisfaction actually report satisfaction at a comparable rate. For predicted NPS categories, inspect confusion between promoters, passives and detractors, rather than relying only on the final aggregate.

Review results across languages, call reasons, channels and relevant customer groups. An acceptable overall result can conceal unreliable estimates for a particular segment.

Keep human review in the loop

Use reviewers to examine ambiguous calls, disputed labels and examples where predictions diverge from survey answers. Review the supporting transcript and, where appropriate, the recording rather than accepting a summary alone.

For broader implementation lessons, the guide to common mistakes when introducing AI into a contact center explains why ownership and workflow design matter alongside technology.

Report measured and inferred results separately

A useful dashboard should make the source of each result immediately visible. Separate:

  • Survey CSAT and NPS: answers given by customers, with response coverage clearly stated.
  • Predicted outcomes: model estimates, with their coverage and validation limitations.
  • Conversation evidence: the themes and excerpts that help explain an experience.

Do not fill missing survey responses with predictions and present the combined result as ordinary survey NPS. That changes the meaning of the measure and can create a false improvement when the model or coverage changes.

For an aggregate predicted NPS, estimated promoter and detractor proportions can be calculated from model outputs. However, this remains a model-based estimate, not observed NPS, and should carry a clear label and an explanation of uncertainty.

Trend comparisons also require consistency. Changes in survey wording, call mix, supported languages or model versions can move a result without any underlying change in customer experience.

Use the findings to choose improvements

The strongest value often lies in explaining dissatisfaction, not producing another headline score.

Suppose callers express appreciation for agents but repeatedly describe an unresolved billing problem. Coaching agents to sound warmer may not address the cause. The appropriate action could be to review billing communications or the process for resolving disputes.

CX Insights brings together sentiment, compliance, performance and Auto QA across voice, email and chat. Its sentiment intelligence tracks customer emotion and agent empathy by journey stage, product and team, providing context for investigating experience drivers. That context should not be treated as a direct substitute for survey-based CSAT or NPS.

Assign each recurring issue an owner and check whether subsequent customer feedback supports the intended improvement. Avoid using inferred satisfaction alone to rank agents: policy constraints, case difficulty and customer history can all shape the outcome.

Governance questions before rollout

Call recordings may contain sensitive personal information. Confirm that your proposed analysis fits the organisation’s applicable privacy obligations, customer notices and recording practices.

Ask who can access recordings, transcripts and extracted evidence; how long each is retained; and whether access and changes are auditable. For public-sector organisations, also consider whether satisfaction and recommendation measures are appropriate proxies for service quality, fairness or accessibility.

Finally, establish who approves model changes and who can challenge a result. A useful measurement system must be explainable enough to support decisions—not merely produce a score.

Frequently asked questions

Can AI calculate NPS without asking the customer a question?

AI can estimate a recommendation category if there is a suitably validated model, but it cannot establish the customer’s actual NPS response from sentiment alone. Report the result as predicted NPS and keep it separate from survey NPS.

Is sentiment analysis the same as CSAT measurement?

No. Sentiment analysis interprets emotional language or vocal signals, while CSAT records satisfaction with a defined experience. Sentiment can help explain a satisfaction result, but politeness, frustration and satisfaction do not map neatly onto one another.

Can AI extract a CSAT answer spoken during a call?

Yes, provided the recording contains a clear question and an identifiable answer. The system must distinguish the customer’s response from agent speech and confirm that the response refers to the intended question and scale.

Should predicted scores replace customer surveys?

Not by default. Surveys provide direct feedback and a reference for checking predictions, while recordings offer context and broader experience signals. Using both preserves the distinction between what customers reported and what a model inferred.

See CX Insights in action

Find out how it fits your team’s workflow.

Contact Us

Offices

| | |