What Is 100% CSAT Coverage?

One hundred percent CSAT coverage sounds like every customer completed a survey.

That is not what it means.

You cannot force every customer to rate a support conversation, and you should not imply that an AI prediction came from the customer. The practical definition is simpler:

Every eligible closed conversation has a satisfaction signal—either a customer-submitted rating or a clearly labelled AI prediction.

That is 100% coverage without a 100% survey response rate.

Response rate is not coverage

Three metrics are commonly blurred together:

Survey response rate

The percentage of eligible survey requests that produced a customer rating.

Ratings received ÷ eligible survey requests

Surveyed CSAT

The percentage of submitted ratings that meet the organization’s definition of positive.

Positive ratings ÷ all submitted ratings

CSAT coverage

The percentage of eligible conversations with either a surveyed or predicted satisfaction signal.

(Conversations with surveyed CSAT + unrated conversations with predicted CSAT) ÷ eligible closed conversations

Consider a team that closed 1,080 conversations last month:

• 80 are excluded — bot-only threads the customer never engaged with, merged duplicates, and conversations with too little content to judge — leaving 1,000 eligible.
• 250 customers submit a rating, and 225 of those ratings are positive.
• Survey response rate is 25%; surveyed CSAT is 90%.
• AI evaluates the 750 unrated conversations.
• Satisfaction coverage is 100% of eligible conversations — which is about 93% of everything that closed.

Each figure answers a different question, and none should replace the others.

Why coverage matters

Survey participation is usually incomplete, and that creates two operational problems: most conversations can never be inspected through the CSAT dashboard, and the customers who respond may not resemble the ones who stay silent. In one study of more than 170,000 customer-service chats, only 16.2% of sessions received a rating, and rated sessions skewed more positive than the unrated majority — the positivity bias we unpack in our field notes on CSAT response bias.

The production-oriented pCSAT paper reaches a similar conclusion for call centres: low participation biases the average and hides the conversations that most need attention. Concretely, the miss looks like this — an angry customer closes the chat, never rates, never gets a recovery follow-up, and churns two weeks later, invisible to a dashboard that only counts responses.

Coverage gives teams a way to inspect the silent majority while preserving the special status of direct customer feedback.

The two-signal model

A reliable CSAT system has two sources.

1. Surveyed CSAT

This is the customer’s explicit rating. It is first-party feedback and the source of truth for that conversation.

Surveyed ratings are particularly valuable because they:

• Express the customer’s chosen judgment
• May include a written explanation
• Create labels for model calibration
• Reveal praise or dissatisfaction the transcript alone may not make obvious

2. Predicted CSAT

This is an AI-generated estimate based on the conversation. It is used only where a customer rating is unavailable.

Predicted CSAT can help teams:

• Find likely dissatisfaction among nonrespondents
• Compare experience patterns across issue types
• Review a representative set of agent work
• Detect changes before enough survey responses accumulate
• Prioritize conversations for human review

A prediction is useful precisely because it fills what would otherwise be a blank — but it remains an inference, and every rule below treats it as one.

Five rules for honest coverage

Rule 1: Actual ratings always win

If the customer submitted a rating, that is the number. Never overwrite it with the model’s view.

Rule 2: Label every prediction

“Predicted CSAT,” “AI-inferred,” “estimated satisfaction” — any label works, as long as nobody could mistake the score for something the customer chose.

Rule 3: Keep the sources separable

Managers should be able to filter surveyed-only, predicted-only, and all covered conversations, so a combined number can never hide where its evidence came from. These first three are the same trust rules that make survey data interpretable despite response bias; the next two are specific to running coverage.

Rule 4: Define eligibility

Not every closed item deserves a satisfaction score, and coverage should be calculated against eligible conversations rather than inflated by scoring meaningless records. A reasonable starter policy excludes:

• Conversations with fewer than two customer messages
• Bot-only threads the customer never engaged with
• Merged or duplicate tickets
• Spam and conversations closed with internal notes only

Be honest about the consequence: after exclusions, “100% coverage” means 100% of the conversations that qualify — in the worked example above, 1,000 of the 1,080 that closed. Publishing the exclusion count alongside the coverage number is what keeps the claim credible.

Rule 5: Monitor confidence and accuracy

Once a month, compare predictions against customer ratings that arrived after the prediction was made, and track the agreement rate. Investigate when agreement drops below your established baseline, and re-check deliberately after any change to the model, scoring rubric, product, or support workflow. The full workflow — holdout sets, disagreement review, subgroup checks, drift monitoring — is in our guide to validating predicted CSAT.

What 100% coverage does not mean

It does not mean:

• Every customer answered
• AI knows exactly what every customer felt
• Survey response rate no longer matters
• Predicted and surveyed CSAT are interchangeable
• Every score should be used in agent compensation
• Human review is unnecessary

Coverage is an operating tool that expands what the support team can see — nothing more, and nothing less.

How to report it

A clear weekly report might show:

Metric	Result
Closed conversations	1,080
Excluded (bot-only, duplicates, too little evidence)	80
Eligible conversations	1,000
Customer ratings	250
Survey response rate	25%
Surveyed CSAT	90%
AI-scored unrated conversations	750
Predicted positive rate	84%
Satisfaction coverage	100%

Notice the gap between the two quality numbers: surveyed CSAT is 90%, but the predicted positive rate on unrated conversations is 84%. That six-point difference is positivity bias made visible in your own data — the customers who bothered to rate were happier than the silent majority, exactly the pattern in the research cited earlier. A stable gap of a few points is normal; treat it as the honest correction to your headline score rather than a problem to fix. What deserves investigation is movement: if the gap widens month over month, either the unrated experience is deteriorating or the model is drifting, and a sample of low-predicted conversations plus recent prediction-versus-rating disagreements will tell you which.

The report may then show the most common drivers of low predicted scores, recent DSAT comments, and conversations where the customer rating disagreed with the prediction. That is more actionable than presenting a single blended percentage.

From signal to action

Once every eligible conversation has a signal, the score stops being a monthly average and starts pointing at work. An illustrative chain:

Predicted DSAT turns out to be concentrated in billing-refund conversations handled by agents in their first 90 days. QA pulls fifteen of those conversations and finds a saved reply that promises the wrong refund window. The macro gets corrected, the new agents get one coaching note, and next month’s predicted DSAT in that queue is the check on whether the fix worked.

None of that required a single additional survey response. It required a signal on every conversation, connected to conversation evidence, QA attributes, and someone whose job is to act on what the signal surfaces.

The one-sentence coverage report

If you adopt one habit from this page, make it this one. Instead of reporting “CSAT was 90%,” write:

“Of 1,000 eligible conversations, 250 customers rated us (90% positive) and AI scored the remaining 750 (84% predicted positive) — 100% coverage, with the 6-point gap under review.”

Every number in that sentence is honest, every source is labelled, and nobody has to ask what the score is hiding.

Supportman produces that view out of the box: predicted CSAT on every unrated eligible conversation, kept visibly distinct from customer-submitted ratings.

Explore 100% CSAT coverage →

Frequently asked questions

Can you really achieve 100% coverage?

Yes, if every eligible closed conversation has either a customer rating or a valid prediction. Conversations without enough information should be excluded transparently rather than force-scored.

Does prediction replace the survey?

No. Surveys supply direct feedback and calibration labels. Prediction extends coverage to the conversations that customers do not rate.

Is 100% CSAT coverage the same as 100% QA coverage?

No. CSAT coverage means every eligible conversation has a satisfaction signal — how the customer likely felt. QA coverage means every conversation was evaluated against a quality rubric — how well the agent performed. The two work together (predicted DSAT is a good way to choose conversations for QA review), but a conversation can score high on satisfaction and still fail QA, or the reverse.

Can predicted CSAT be used for coaching?

Yes—as a signal for selecting and understanding conversations. Coaching should use repeated patterns, conversation evidence, and human judgment rather than one inferred score.