RevelirQA performs quality assurance to the standard of a human quality analyst. Revelir calibrates it with each customer until it reaches over 95% accuracy, as validated by that customer's own QA analysts on a sample of evaluations. Covering 100% of conversations comes after that, as table stakes most auto QA tools already claim. This standard matters here because one question decides whether agents or policies are to blame: did the agent follow policy correctly? If sentiment falls in conversations where agents applied your SOP correctly, the policy is the cause. These drops usually cluster in one contact reason across many agents. If sentiment falls where specific agents departed from policy, often across several contact reasons, the agent is the cause.
TL;DR
- RevelirQA performs QA to the standard of a human quality analyst. Revelir calibrates it with each customer to over 95% accuracy, as validated by that customer's own QA analysts on a sample of evaluations. Covering 100% of conversations comes after that, as table stakes most AutoQA tools already claim.
- Sentiment alone cannot assign blame. Read the sentiment arc (start versus end of the conversation) next to a policy-adherence score on the same ticket.
- Drops concentrated in one contact reason, among agents who followed policy, point to the policy. Drops concentrated in specific agents, across contact reasons, point to the agents.
- Coaching fixes agent causes. It cannot fix a policy that works as written and still frustrates customers.
What does falling customer sentiment in support tickets actually tell you?
Before you can assign a cause, you need to be clear about what the sentiment number captures. Customer sentiment analysis reads unstructured conversation data and turns it into customer insights a team can act on [google.com]. Sentiment analysis typically uses natural language processing, lexicon-based rules or machine learning models to read the ticket text.
The trap is that end-of-ticket sentiment blends two separate things:
- Why the customer contacted you. Someone whose refund was declined arrives unhappy, whoever answers.
- How the conversation was handled. This covers tone, accuracy, and whether the agent applied the right policy.
The sentiment arc separates them. It compares sentiment at the start of a conversation with sentiment at the end. A customer who arrives angry and leaves neutral was handled well, even though the final score looks poor. A customer who arrives neutral and leaves angry tells you something broke inside the conversation. Tickets marked "resolved" often hide exactly this pattern.
How do you tell whether agents or policies are causing the drop?
The sentiment arc tells you where things went wrong. Policy adherence tells you why. Cross the two signals and four patterns appear:
| Policy adherence | Sentiment arc | Likely cause | First action |
|---|---|---|---|
| Agent followed policy | Worsened | Policy or process | Take the evidence to the policy owner |
| Agent departed from policy | Worsened | Agent knowledge or behaviour gap | Coach on the specific miss |
| Agent departed from policy | Improved or held | Workaround that hints at policy friction | Ask whether the workaround should become the policy |
| Agent followed policy | Improved or held | Working as intended | No action |
The third row is the one teams most often miss. When agents bend a rule and customers end up happier, that is often the earliest sign the rule needs to change.
The reasoning works like diagnosing skids on a highway. If one driver skids on every road, look at the driver. If every driver skids on the same bend, look at the bend. When the problem shows up across everyone in one place, the cause is the place. When it follows one person everywhere, the cause is the person. The same is true of sentiment drops.
Why is this diagnosis hard to run on sampled tickets?
The table above only works if you have enough scored conversations in each slice. Sample size is where traditional review runs short. Klaus reported in its AutoQA announcement that, on average, only 2% of customer service interactions are manually evaluated.
A small sample can still show an individual agent's habits over time. This question is different, because you have to split the data twice: by agent and by contact reason. Once a small sample is divided across many agents and many contact reasons, most slices hold only a handful of tickets. A policy problem in a mid-volume contact reason can stay invisible. The limit here is sample size, not the reviewers' judgment.
What does a step-by-step diagnosis look like?
Once you know what data the diagnosis needs, the workflow itself is short:
- Tag every ticket with its contact reason and with sentiment at the start and at the end.
- Score policy adherence against the current SOP for that contact reason. A QA scorecard grounded in your policies catches a wrong refund amount, incorrect eligibility criteria or a wrong escalation path.
- Segment declines by contact reason first, then by agent within each reason.
- Apply the four-pattern table to each segment where sentiment fell.
- Read the evidence before acting. Treat a score as a hypothesis. The quoted exchange and the SOP passage it was measured against are the proof.
- Recheck after the fix. Did the sentiment arc recover in that contact reason, or for that agent?
RevelirQA is built around this workflow. It tags each ticket with sentiment, contact reason and recurring issue type. It then scores the ticket against the customer's own SOPs, which it retrieves via RAG before each evaluation. Every score comes with a trace showing the model, the prompt, the documents retrieved and the reasoning.
What should you change once you know the cause?
A diagnosis only helps if the fix matches it, and the two causes have different owners:
- Agent cause: coach on the exact QA metric the agent missed, quoting their own exchange. Then track that agent's scores over the following weeks.
- Policy cause: send the policy owner (ops, product or compliance) the contact reason, the volume, the sentiment arcs and examples where agents applied the policy correctly. If the rule is required, for example by regulation, change how it is explained rather than the rule itself.
- Mixed cause: when agents read an unclear SOP in different ways, rewrite the SOP first and coach second.
- AI chatbots: the same logic applies. A bot that follows policy while sentiment falls points to the policy. A bot that gives wrong answers points to its configuration or knowledge base.
In our view, teams reach for coaching too quickly because it is the lever CX controls. Teaching an agent to deliver an unpopular policy more warmly is the wrong fix when the policy itself is driving the drop.
Frequently Asked Questions
Can agents and policies cause a sentiment drop at the same time?
Yes. A vague policy causes inconsistent handling, so both show up together. Fix the policy wording first, then coach.
Is CSAT enough to find the cause?
No. CSAT measures the outcome, and only for customers who answer the survey. It does not show whether the agent followed policy.
What is auto QA?
Auto QA is automated quality assurance for customer service conversations. AI scores them against a QA scorecard instead of relying only on manual review of sampled tickets. RevelirQA performs AutoQA at the standard of a human quality analyst, which far fewer tools can show.
How accurate is RevelirQA?
Revelir calibrates it with each customer to over 95% accuracy. When the customer's own QA analysts review a sample of its evaluations, they judge more than 95 out of 100 correct.
Does RevelirQA score AI chatbots?
Yes. It applies the same scorecard to human and chatbot conversations on any chatbot platform that produces a conversation log via API.
Does RevelirQA discipline agents automatically?
No. Results go to the QA lead, who can override any grade, and agents can contest their scores.
About Revelir AI
Revelir AI is a Singapore-headquartered company, founded in 2025, that builds RevelirQA, an AI quality assurance platform for enterprise teams worldwide. RevelirQA scores conversations against each customer's own SOPs and QA scorecard, and every score carries an auditable reasoning trace. Revelir calibrates it per customer to over 95% accuracy, as validated by the customer's QA analysts. It connects to any helpdesk via API, including Zendesk and Salesforce, and runs in production at Xendit and Tiket.com.
If sentiment is slipping and you cannot yet tell whether to coach your team or rewrite your policies, see how RevelirQA separates the two at https://www.revelir.ai/.
