Inferred CSAT

Last updated: October 2, 2026

Inferred CSAT is Unwrap's read on how your customers actually felt in a support interaction, without ever sending them a survey.

For every support conversation we ingest, Unwrap reads the transcript and assigns:

  • a label for each individual customer message, so you can see satisfaction rise and fall turn by turn across the conversation

  • a conversation-level label: how the customer likely felt walking away from the interaction

Because it is inferred from the conversation itself, Inferred CSAT covers 100% of your support volume which have at least one customer message, not just the small fraction of customers who respond to a post-chat survey.

The labels

Every customer message and every conversation gets exactly one of six labels:

  • Very Satisfied | Explicit strong positive language ("amazing", "thank you so much", "perfect"), and/or an enthusiastic confirmation that the issue is fully resolved.

  • Satisfied | Confirms the issue is resolved, or uses mild appreciative language ("thanks, that works"). No unresolved frustration remains.

  • Neutral | Transactional or matter-of-fact: describing the issue, asking a clarifying question, supplying requested info, acknowledging next steps. No sentiment either way.

  • Dissatisfied | Mild-to-moderate frustration or unmet expectations ("this is frustrating", "I already tried that", "I still don't understand").

  • Very Dissatisfied | Strong negative emotion: explicit anger, profanity, threats to cancel or escalate, outrage at a repeated failure to resolve.

  • Insufficient Signal | Genuinely no basis to judge satisfaction: a bare greeting, a one-word menu click-through, a request to be connected to an agent, and nothing else.

A few things worth knowing about how these labels are applied:

  • Weak or mixed signal still gets a real label. "Insufficient Signal" is reserved for messages (or conversations) with truly nothing to go on — not for messages where the sentiment is simply subtle.

  • A flat, unemotional conversation is Neutral, not Insufficient Signal. If the customer described their issue, request, or any substantive information anywhere in the conversation, the conversation-level label will be Neutral at worst. A conversation is only Insufficient Signal overall when the customer never said anything substantive at all.

  • The conversation-level label is not simply the last message. It answers: considering the interaction as a whole, which single label best captures how the customer ultimately felt?

Where you'll see it

On each piece of feedback

Support conversations show an Inferred CSAT chip in the feedback card header. Hover the chip to read the AI's one-sentence justification for that label.

Open the full conversation and each customer message carries its own colored marker on the transcript rail, with its own justification — this is where the turn-by-turn view becomes useful. You can see, for example, a customer who opened Very Dissatisfied and left Satisfied.

As a filter

Inferred CSAT is a standard Unwrap field, so it works anywhere fields work:

  • Filter your feedback, Explore taxonomy, boards, and digests to just Very Dissatisfied conversations

  • Break charts down by Inferred CSAT to see which issues drive dissatisfaction

  • Use it as a filter on the Support Performance / Support IQ page via the one-click label chips in the sidebar

As the DSAT Rate metric

DSAT Rate is the headline metric built on Inferred CSAT:

DSAT Rate = (Very Dissatisfied + Dissatisfied) ÷ all labeled conversations

It appears as a metric box on the Support Performance page, as a sortable column when you break performance down by team or agent, and as a chart metric you can trend over time or break down by any field.

Note on the denominator: DSAT Rate's denominator is all labeled conversations, including those labeled Insufficient Signal. This keeps the metric stable, making it the share of your total support volume that went badly, not the share of a filtered subset.

As the Support Recovery Flag

Unwrap also derives a Support Recovery Flag (Yes/No) from the per-message labels. A conversation is flagged Yes when:

  • the conversation has at least 3 customer messages, and

  • at least one of the first 3 customer messages was Very Dissatisfied, and

  • the conversation-level label ended up Satisfied or Very Satisfied

In other words: the customer arrived furious and left happy. Filtering to Support Recovery = Yes surfaces your best save stories, which is useful for coaching material and for recognizing agents who turn hard conversations around.

CSAT Protocol: teaching the rubric about your business

The standard rubric knows what frustration looks like in general. It doesn't know that in your business, "I'll just use the web version" is a churn signal, or that your enterprise customers are polite even when they're furious.

A CSAT Protocol is a free-text document where you give the grader that context. It is added to the standard rubric, never a replacement for it. The six labels and their core criteria stay intact, and your protocol supplements them.

Where to find it

Open the protocol editor from Support IQ settings. You'll see two tabs:

  • Support Protocol — the operating procedure used for agent quality grading

  • CSAT Protocol — the satisfaction-grading context described here

Write your protocol in the CSAT Protocol tab and save. Editing requires Support IQ edit permission.

A few structural differences from Support Protocol, worth knowing:

  • CSAT Protocol is team-wide only. There is no per-taxonomy-group CSAT Protocol. One document applies to every conversation on the team.

  • There is no version history. CSAT Protocol stores only the current text, so a save overwrites what was there before.

What's worth putting in it

Keep it to things the grader cannot infer from the transcript alone:

  • Domain vocabulary — what your product names, error codes, and internal shorthand mean, and which ones signal a serious problem

  • What counts as escalation in your business — e.g. "mentioning their account manager or legal is an escalation, even if the tone stays polite"

  • Tone norms for your customer base — if your customers are habitually terse, or habitually effusive, say so, so the grader doesn't over- or under-read politeness

  • Known pain points — issues your team already knows always frustrate customers, so a flat-sounding mention still reads as dissatisfaction

  • Channel differences — expectations in a live chat differ from a multi-day email thread

What not to put in it: instructions to redefine or renumber the six labels, or to bias the overall distribution ("grade more generously"). The standard rubric wins on those, and attempting it mostly produces inconsistent labels.

When changes take effect

A saved protocol applies to conversations graded from that point forward. It does not retroactively change existing labels.

To apply it to conversations you've already imported, you can re-grade them from Support IQ settings, scoped to the sources and date range you care about.

This is the intended workflow: write the protocol, save it, spot-check newly graded conversations, iterate on the wording, and only re-grade your history once you're happy with the labels.

FAQs

How is this different from a CSAT survey?

A survey measures the customers who chose to respond. Inferred CSAT measures every customer. Survey response rates in support are typically in the single digits and skew toward the very happy and the very angry, so survey CSAT and Inferred CSAT will rarely match exactly. Inferred CSAT is the fuller picture, while survey CSAT remains the customer's own self-reported score.

If you also send Unwrap your survey scores, you can keep both: survey CSAT stays its own field, and you can compare the two side by side.

Does it replace agent quality scoring?

No. Inferred CSAT measures how the customer felt. Unwrap's agent grading measures how well the agent handled the conversation. The two frequently diverge; an agent can handle an unfixable problem perfectly and still leave a dissatisfied customer. Both are available on the Support Performance page, and reading them together is more informative than either alone.

Why does a conversation have no Inferred CSAT value?

The most common reasons:

  • The conversation has no customer messages (bot-only or agent-only), so it was never eligible for grading.

  • The conversation predates Inferred CSAT being enabled for your account and hasn't been backfilled yet.

  • Inferred CSAT isn't enabled for your account. If the Inferred CSAT filter chips on the Support Performance page are greyed out, the feature hasn't produced any grades for your team yet.

Can I correct a label I disagree with?

Not on a single conversation. Inferred CSAT labels are read-only, with no per-conversation override.

If labels look systematically wrong for your domain, that's what the CSAT Protocol is for: add the missing context in Support IQ settings, then re-grade the affected conversations. If the labels still look wrong after that, send us examples! That's the signal we use to improve the standard rubric itself.

Can I get historical conversations graded?

Yes. Inferred CSAT can be run over your existing support history from Support IQ settings, scoped to the sources and date range you choose. If you want a large or full-history re-grade, reach out and we'll scope it with you.

Does Inferred CSAT grade the agent's messages too?

No. Only customer messages receive labels. Agent and bot messages appear in the transcript as context. They're what the AI reads to understand the customer's reaction , but they're never scored.

Getting started

  1. Ask your Unwrap contact to enable Inferred CSAT for your account (it is gated per account today, not self-serve).

  2. Confirm your support source is connected and conversations are ingesting with full message history. Inferred CSAT depends on the conversation transcript, so sources that send only a single summary field will produce much weaker labels.

  3. Once grades begin landing, start on the Support Performance page: look at DSAT Rate overall, then break it down by team and by agent.

  4. Filter your Explore taxonomy to Inferred CSAT = Very Dissatisfied to see which product and process issues are driving the worst conversations.

  5. Filter to Support Recovery Flag = Yes for your save stories.