Review Case AI Insights and turn quality signals into useful follow-up

Updated Wednesday, October 28, 2026 · Marcel Gravelle

AI Insights brings stored customer-experience and quality analysis into the Case you are reviewing. It helps a support team decide where to look more closely: an unclear explanation, a difficult customer journey, or a missing next step. In a B2B environment, these signals are most useful when combined with the customer’s business impact, technical dependencies and the actual conversation. Use them to focus a thoughtful review and agree a concrete improvement.

At a glance

For: support supervisors, quality reviewers, team leads and agents reviewing Case quality. Applies to: Supportbench Core.

Start here: Open an eligible Case → AI Insights

Understand the analysis you are opening · Check availability before searching for the tab · Follow a complete Case-quality review · Interpret the customer-experience labels · Read the QA scale and individual dimensions · Worked example: a complex integration Case · Build a fair and useful review routine · Troubleshoot missing or unexpected insights

Understand the analysis you are opening

The tab reads analysis already stored for the Case. Opening it or choosing Refresh stored analysis does not run a new AI evaluation, regenerate a summary or rescore the conversation. Refresh retrieves the currently stored result again.

Information	What it represents	How to use it
Customer experience: CSAT	The stored AI interpretation of the customer experience, with a reason where available.	Compare the interpretation with the customer’s actual words and outcome.
Customer experience: CES	The stored AI interpretation of customer effort.	Look for repeated requests, unclear instructions or difficult handoffs.
Quality assurance: QA score and summary	The available Case-level quality assessment.	Choose specific interactions to inspect before drawing a conclusion.
QA breakdown	Available quality dimensions such as explanation, validation and follow-up.	Turn a broad concern into a precise review question.
Survey results elsewhere in the Case	Feedback submitted through a survey workflow.	Keep customer answers distinct from an AI inference.

AI CSAT is not a submitted customer survey

A label such as Positive Experience is an AI assessment. It does not mean the customer submitted a positive survey response, and a negative label is not a verified statement of the customer’s satisfaction. Review actual feedback separately.

Check availability before searching for the tab

Open the Case you are authorized to review. Confirm its identity and customer context.
Look for AI Insights among the Case tabs. Its availability depends on the AI capability, the relevant role permission and an available stored Case summary.
Do not assume every closed Case has the tab. Closing a Case alone is not proof that analysis is stored or that your role can view it.
If a colleague can see it and you cannot, ask an administrator to compare the applicable AI/role access and your Case visibility.
If analysis has not been stored, follow your account’s AI-processing and support process. The tab itself provides no generation command.

The current visibility path is based on capability and stored analysis rather than a universal rule that the Case must currently be closed. Follow the available tab and result for the Case instead of relying on an older instruction to open “any closed Case”.

Follow a complete Case-quality review

Open the Case and read enough of the conversation to understand the request, business impact, technical constraints and outcome.
Select AI Insights. Wait for the stored analysis to load.
Read the Customer experience section: CSAT, its available reason under Why, and CES. Note which claims need supporting conversation evidence.
When present, read Quality assurance, its QA score and Summary. Review the dimensions under Breakdown rather than reacting only to the overall label.
Return to the relevant Activities and compare the assessment with what the agent and customer actually said. Include earlier context and external dependencies that may explain the interaction.
Choose a specific next step: clarify a reply template, coach a follow-up habit, correct an article, or investigate a process handoff. Assign responsibility and a review date through your normal team process.
If you expect a newer stored result, choose Refresh stored analysis and inspect the response. This rereads saved data; repeated clicking does not request a fresh AI run.
Record the human-reviewed conclusion in the appropriate internal location. Distinguish it from the AI’s wording.

Interpret the customer-experience labels

Measure	Stored value and displayed label	Interpretation
AI CSAT	0: Negative Experience; 1: Positive Experience	A binary stored interpretation, with a reason where present. Other or unavailable values need investigation rather than an invented conclusion.
AI CES	0 or 1: Effortless; 2: Quite simple; 3: Neutral; 4: Somewhat difficult; 5: Very challenging	An effort interpretation. Higher displayed difficulty should prompt a review of the steps and handoffs the customer experienced.

A technically correct solution can still require substantial customer effort. For example, a customer may have repeated the same diagnostic information after several handoffs. The useful follow-up is to investigate the handoff and knowledge process, not to assume the final agent caused every difficulty.

Conversely, a friendly conversation does not prove the customer’s technical issue was resolved. Compare the stated solution, customer confirmation, outstanding tasks and actual support outcome.

Read the QA scale and individual dimensions

Stored QA score	Displayed label
0	Not Applicable
1	Poor
2	Needs Improvement
3	Satisfactory
4	Good
5	Excellent

Not Applicable is not a failed score. Missing QA data is also different from a measured low result. The Case-level analysis uses the available Case QA record; it is not a promise that every Activity was separately evaluated or that the current tab has just recalculated the score.

Dimension	Useful human review question
Grammar	Could errors make the instruction difficult to follow or change its meaning?
Tone	Does the response fit the customer’s situation and the organization’s communication standard?
Politeness	Is the interaction respectful, including when setting limits or requesting more information?
Product Knowledge	Are the explanation and proposed steps technically correct for the relevant product/version?
Empathy	Does the agent acknowledge the customer’s actual business impact without inventing commitments?
Explanation	Can the intended reader understand and perform the next steps?
Language	Is the wording clear and appropriate for this reader’s knowledge and language context?
Validation	Did the interaction establish the problem and relevant facts before assuming a cause?
Clear Follow-Up	Does the customer know the next action, responsible person and expected update?
Solution Quality	Does the proposed resolution or path forward address the verified issue and constraints?

These questions are review recommendations. They help a supervisor investigate the displayed dimensions; they are not a claim that the AI follows a customizable checklist supplied through this tab.

Worked example: a complex integration Case

A business customer’s integration issue required several teams and a vendor dependency. AI Insights shows a difficult customer experience and a weak Clear Follow-Up assessment. The supervisor reads the conversation and finds that the technical investigation was necessary, but two handoffs did not tell the customer who owned the next update.

Separate unavoidable technical complexity from avoidable communication gaps.
Confirm the actual handoff messages and customer requests before accepting the assessment.
Agree a practical habit: each handoff states the responsible team, next diagnostic step and next update expectation.
Improve a shared reply template or handoff checklist and link the relevant reviewed knowledge where useful.
Review a later sample of comparable Cases to see whether the follow-up became clearer. Do not compare a simple password request with a multi-vendor incident as if their effort were equivalent.

The outcome is a specific operational improvement supported by conversation evidence. It is more useful than telling an agent merely to “improve the AI score”.

Build a fair and useful review routine

Select a representative sample of Cases and include different complexity levels, customer impacts and channels. Do not examine only the lowest scores.
Have reviewers calibrate a few examples together so they use the same interpretation of quality and technical correctness.
Read the underlying conversation before using an AI assessment in coaching. Discuss uncertain or disputed interpretations with the people who know the Case.
Separate individual habits from system problems such as missing knowledge, unclear ownership or delayed engineering dependencies.
Use the KCS process to correct reusable knowledge gaps found during review. A stronger article can improve many future interactions.
Use aggregate reports or scorecards for trends, keeping their metric definitions and populations distinct from this one Case’s panel.
Limit access and distribution to the appropriate review audience. Customer conversations and internal coaching details do not belong in public articles.

Record what changed and what success will look like: clearer next steps, fewer repeated requests, or more consistent use of a verified solution. These are operational goals to measure, not guaranteed performance claims supplied by the feature.

Troubleshoot missing or unexpected insights

Problem	Checks and corrective action
AI Insights is missing	Check Case access, AI capability, the relevant role permission and whether a stored Case summary exists. Closed status alone is insufficient.
The tab says analysis is unavailable	Check the result and retry later when appropriate. An administrator may need to investigate processing or access; the tab cannot generate missing analysis.
Refresh shows the same result	Refresh stored analysis rereads saved data. It does not start a new evaluation.
QA is absent but customer-experience information exists	The available stored Case data may not include a Case-level QA record. Missing data is not a zero-quality judgment.
A dimension says Not Applicable	The mapped QA value is zero. Review applicability and available evidence rather than treating it as Poor.
AI CSAT disagrees with a survey	They are different evidence sources. Read the actual customer feedback and the AI reason in context.
An assessment misses a technical constraint	Check the conversation and external dependency evidence, record the human-reviewed conclusion and use your normal AI/support feedback process.
The aggregate report has different values	Check scope, date basis, eligible population and report processing. This panel is a stored analysis of one Case.
The score looks outdated after later messages	The tab shows stored results. Check the relevant processing lifecycle with an administrator; do not assume the display automatically rescores every new message.

Was this article helpful?