Inferred CSAT vs surveys: measuring satisfaction on every conversation

Inferred CSAT estimates satisfaction on every conversation from what was said, while surveys capture only the few customers who answer; the strongest programs use inferred scores for coverage and surveys as ground truth.

Published Oct 20, 2026 · Updated Oct 20, 2026 · Author: Contact Center AI Index

Primary research hub: VoC and CX analytics · Entities: Level AI, iCSAT, NiCE, Observe.AI, Medallia

Inferred CSAT (customer satisfaction) is a score that a model assigns to a conversation based on what the customer and agent actually said, so it covers every interaction. A post-contact survey covers only the customers who answer it, often a small and skewed minority: researchers at Dialpad report that "only up to 8%" of callers leave a CSAT response, and they tend to be those with strongly positive or negative experiences (ACL 2023). The practical answer is to use both: inferred scores such as Level AI's iCSAT for coverage and root cause, and surveys as the ground truth you calibrate against.

What is inferred CSAT and how is it calculated?

A traditional CSAT score asks customers to rate their satisfaction, usually on a 1 to 5 scale, right after a contact. NPS (Net Promoter Score) asks how likely they are to recommend you on a 0 to 10 scale and subtracts the share of detractors (0 to 6) from the share of promoters (9 and 10), as Bain explains.

Inferred CSAT skips the question. A model reads the transcript (and sometimes audio features) and predicts how satisfied the customer was. Approaches differ:

• Composite models. Level AI's iCSAT page says its score blends three parts: a sentiment model with "temporal weights" and a range of detected emotions, a customer-effort model (repetition, time, transfers, customer actions), and whether the issue was resolved. Level AI's platform page names the underlying model Veridia.
• Survey-trained predictors. Academic and industry teams train models to predict the survey score a customer would have given. Kanchinadam and colleagues, for example, predict 1 to 10 post-call survey scores from speech-to-text transcripts using a graph neural network (arXiv, 2021).
• LLM annotation. A large language model is prompted to rate satisfaction on defined axes, such as overall, agent, outcome and effort (arXiv, 2026).

Why do post-contact surveys miss most customers?

Two problems compound each other.

1. Sparsity. Few customers respond. The Dialpad team's ACL 2023 paper puts the share at up to 8% of callers.
2. Response bias. The same paper says respondents "tend to be customers with strongly positive or negative experiences." The quiet middle, and often the customer who gave up, are missing.

The effect on the headline number can be large. In a July 2026 preprint, Hong and Potteiger used GPT-4.1 to annotate about 9,000 support conversations at a global consumer-goods firm. Reading every contact gave "a full-census 2.91 against the surveyed 3.62 on a five-point scale" (arXiv 2607.11026). In that dataset, the survey average overstated satisfaction across the whole customer base. It is one firm and a preprint, so treat it as a warning to check your own data, not a universal correction factor.

Inferred CSAT, surveys and sentiment: what does each one measure?

How accurate is inferred CSAT compared with what customers say?

Published evidence is thin, and vendor pages generally do not include accuracy numbers. Level AI's iCSAT page says iCSAT can be more trustworthy than biased surveys but gives no supporting accuracy data. The independent evidence we found:

• In the 2026 preprint, LLM-annotated overall, agent and outcome scores correlated with customers' own ratings at about 0.65 before adjustment. Disagreements clustered in a small set of divergent sessions, and excluding the severe ones raised the overall correlation to 0.811. The authors found the extra axes did not improve prediction, and said their value lies in "attribution and coverage."
• The ACL 2023 paper optimizes for precision on the lowest score (CSAT = 1), because finding the angriest customers is the operational goal. The abstract reports the model runs in production across millions of calls per month but gives no headline accuracy figure.

The takeaway: inferred CSAT tracks real satisfaction reasonably well in published work, but not perfectly. Its main value is seeing the customers surveys miss and explaining why a score is low.

Which platforms offer inferred CSAT or real-time sentiment scoring?

• Level AI: iCSAT on 100% of interactions with evidence and reasoning (iCSAT page); live call sentiment for managers through Manager Assist (Agent Assist page). In a Level AI case study, Smartsheet reports a 12% iCSAT increase "in a few months" and internal CSAT moving from 2.95 to 3.31 between February and July (vendor-published).
• NiCE: Interaction Analytics scores sentiment "for both the agent and the customer" and offers an add-on called "NiCE AI Models for Customer Satisfaction" (NiCE).
• Observe.AI: a June 2025 update to GenAI Insights added "turn-by-turn sentiment shifts and the reasoning behind those changes" (KMWorld).
• Medallia: AI-powered emotion analysis across conversations, aligned with survey and journey data (Medallia).

Our broader comparison is in best Voice of the Customer analytics for contact centers.

How do you validate an inferred CSAT model before trusting it?

1. Join scores to surveys. For every conversation that has a survey response, put the inferred score next to the real one.
2. Measure agreement by segment. Check correlation overall, then by channel, queue, language and call length. Weak segments are where the model needs work.
3. Read the disagreements. Pull 50 cases where the model and the customer disagree most. The 2026 preprint found errors concentrated in a small divergent tail; yours may too.
4. Check the low end. Measure how many customers who rated 1 or 2 the model also flagged. Missing angry customers is the costly error.
5. Compare respondents with non-respondents. If inferred scores for non-respondents are lower, your survey average is probably flattering you.
6. Re-check quarterly. Products, policies and customer mix change. Recalibrate when agreement drifts.

Can inferred CSAT actually raise CSAT and NPS?

Measurement alone does not raise scores. Inferred CSAT helps when it points to fixable causes: repeat contacts, transfers, unresolved issues or a broken process. Because composite models such as iCSAT break the score into effort and resolution, they tell a team which lever to pull. NPS is a different case. It measures loyalty to the brand, so price, product and delivery problems move it as much as service does. Inferred CSAT can show which support contacts create detractors, but it cannot replace a relationship survey. A useful pattern is to report three numbers side by side each month: survey CSAT with its response rate, inferred CSAT across all contacts, and the gap between them. A widening gap is a sign that your survey sample is drifting away from your real customer base. Pair the findings with QA and coaching, covered in our guide to the best CX intelligence and analytics software.

Limits of this guide

• No vendor publishes iCSAT accuracy on the pages cited. Level AI, NiCE, Observe.AI and Medallia describe their methods but give no validated accuracy figures there.
• The research is narrow. The 2026 study is a single-firm preprint, the 2021 paper is a preprint, and the ACL 2023 paper reports a response-rate ceiling from one vendor's customer base. None is a cross-industry benchmark.
• Case study numbers are vendor-published. Smartsheet's iCSAT figures come from Level AI's site.
• Survey response rates vary by channel, industry and survey design, so your own numbers may differ from the up-to-8% figure.
• Product names change. Observe.AI's insights capability is cited from June 2025 coverage because its product page did not load.