Sentiment Analysis — How It Works

Velaro automatically analyzes every conversation for customer sentiment when it resolves. An AI model reads the full transcript and scores it on a 1-10 scale, classifies it as Positive, Neutral, or Negative, and detects the customer's primary intent.

This happens automatically — no setup required beyond enabling the Satisfaction report in your subscription.

Scoring Scale

Each conversation gets a sentiment score from 1 (extremely negative) to 10 (extremely positive):

• 9-10: Customer expressed clear satisfaction or gratitude
• 7-8: Generally positive interaction, needs were met
• 4-6: Neutral — factual exchange without strong emotion
• 2-3: Customer expressed frustration or complaints
• 1: Hostile or highly escalated — follow-up recommended

Per-Contact Tracking — Two Scores

Every contact maintains two parallel sentiment scores that update automatically when a conversation resolves:

All-Time Average — equal weight per conversation. Formula: (previousScore × count + thisScore) / (count + 1). Every interaction counts equally. Use this for long-term relationship assessment and compliance reporting.

EWMA Score (Recency-Weighted) — Exponentially Weighted Moving Average with α=0.3. Formula: 0.3 × thisScore + 0.7 × previousEWMA. Recent conversations count 30%; older history carries 70%. Use this to understand how the relationship feels right now.

When they differ: If the EWMA score is higher than the all-time average, the customer has been happier recently than their long-term history suggests — the relationship is improving. If the EWMA is lower, recent interactions are worse than historical — potential churn risk worth acting on.

Trend Direction (Momentum)

Each contact has a trend indicator that tells you whether sentiment is moving in the right direction:

• ↑ Improving — the EWMA increased by more than 0.5 points in the last scored conversation
• → Stable — the EWMA moved less than ±0.5 points
• ↓ Declining — the EWMA dropped by more than 0.5 points

At-Risk contacts are contacts currently marked Negative (score below 4.0) who are also declining. These are your most urgent follow-up candidates.

Confidence Levels

The more conversations scored, the more reliable the average:

• Low (fewer than 5 conversations): directional only — a single conversation can swing the score significantly
• Medium (5-15 conversations): reasonably representative
• High (15+ conversations): statistically meaningful, safe for automated decisions and performance reporting

New contacts start with no score and show "—" until their first conversation is scored. No assumptions are made.

Report Views

The Satisfaction report has four views:

Overview — Aggregate sentiment distribution with donut and bar charts, KPI cards, and top customer intents. The report email includes a comparison to the prior period of equal length (e.g., this week vs. last week) so you can see whether sentiment is trending up or down.

Breakdown — Per-conversation or per-message detail table. Click any row to open the conversation. Filter by sentiment type, source (agent, bot, or contact), and date range.

By Contact — Per-customer dashboard. Shows both the all-time average and the current EWMA score side by side, plus the last conversation score, trend indicator, confidence badge, and channel breakdown. Use the Trend dropdown to quickly filter to Improving, Stable, Declining, or At-Risk contacts. Click any contact to drill into their full conversation history with a mini trendline chart.

By Agent — Per-agent sentiment leaderboard for the selected date range. Shows each agent's average score, positive rate, delta vs. the team average, and a horizontal bar chart so you can spot outliers at a glance. Also includes a Bot vs. Human comparison panel showing whether bot-handled conversations score better or worse than agent-handled ones.

Channel Breakdown

See how satisfaction varies across channels — Web chat, SMS, WhatsApp, IVR, Email, and more. Channels with lower scores may indicate channel-specific friction (like a confusing IVR flow vs. fast web chat). Use the "Exclude bot conversations" toggle on the By Contact tab to isolate human-agent interactions only.

Agent Sentiment Analysis

The By Agent tab lets managers answer questions like:

• Which agents consistently deliver the highest-rated conversations?
• Which agents are below the team average with enough volume to be statistically meaningful?
• Do customers feel better after talking to a human vs. going through the bot?

Each agent is shown with a confidence badge. Agents with fewer than 5 conversations are marked Low Confidence — avoid making performance decisions on small samples. Use the High Confidence threshold (15+ conversations) for coaching conversations.

The team average is shown as a reference line on the bar chart so individual performance is always in context.

Bot vs. Human Comparison

The By Agent tab includes a Bot vs. Human panel that splits sentiment by how the conversation was handled:

• Agent — conversations assigned to a human agent
• Bot — conversations handled fully by automation
• Unassigned — conversations that resolved without assignment

High bot positive rate = your self-service automation is solving problems effectively. Low bot scores with high agent scores = the bot may be frustrating customers before handoff.

Integration with Health Score

Sentiment is one of four components in your Customer Health Score (alongside NPS, CSAT, and Engagement), each worth up to 25 points. The health score updates daily and appears in the conversation sidebar when agents are chatting with contacts.

Integration with Surveys

Sentiment analysis captures tone from 100% of conversations automatically. Post-chat surveys (NPS, CSAT, Star Rating, Thumbs) capture explicit feedback from the percentage of customers who choose to respond. Both metrics feed into health scoring — they complement each other rather than duplicate.

Report Email Subscriptions

Subscribe to receive scheduled Satisfaction summary emails (daily, weekly, or monthly). Each email includes:

• Total rated, positive, negative, and neutral counts for the period
• Distribution bars showing the positive/negative/neutral breakdown
• Trend vs. prior period — how this period compares to the same length window before it (e.g., this week vs. last week), with a trend arrow and percentage point delta
• Top agent highlight — the highest-scoring agent with sufficient data for the period
• At-risk alert — a count of contacts with declining sentiment, if any
• A methodology note explaining how scores are calculated

Manage subscriptions at Reports → Email Subscriptions.

AI Model

Sentiment uses the same AI model configured for your Copilot. If you need different accuracy for sentiment (for example, a more powerful model for nuanced conversations), your admin can set a separate SentimentAiModel override in subscription settings without affecting your bot.

Automated Follow-Ups

With Email Campaigns enabled, you can create automated drip sequences that trigger based on sentiment:

• Recovery Campaign: Auto-enroll contacts when sentiment drops to Negative. Send personalized follow-ups with escalation to account managers if no improvement after 48 hours.
• Delight Campaign: When sentiment is Positive, request reviews, offer loyalty rewards, or ask for referrals while the customer is happy.

Ask your account manager about enabling Email Campaigns if you'd like to automate sentiment-based outreach.

Copilot — Ask Questions About Sentiment

Your AI Copilot understands how sentiment scores are calculated and can answer questions like:

• "What's the customer sentiment trend this week vs. last week?"
• "Are customers happier or more frustrated right now than their historical average?"
• "Which customers are at risk?"
• "How does each agent compare on sentiment this month?"
• "Show me sentiment over the past 30 days"

Ask Copilot to include agents or at-risk contacts: "Show sentiment trends including agent breakdown and any at-risk contacts."

FAQ

How accurate is AI sentiment? Large language models achieve 85-92% agreement with human raters on customer service conversations. Edge cases include sarcasm, cultural context differences, and very short messages like "ok" or "thanks."

What's the difference between the EWMA score and the all-time average? The all-time average weighs every conversation equally — great for long-term baseline. The EWMA gives recent conversations 30% weight so it responds faster to changes. If a long-term happy customer had a rough last interaction, the EWMA will show that signal before the all-time average does.

Does it cost extra? No. Sentiment analysis uses the same AI model and quota as your Copilot. It's included with the Satisfaction report feature.

Can I see sentiment during a live conversation? Yes — agents see a sentiment badge in the contact info panel during active conversations, showing the contact's rolling score and confidence level.

How often does the EWMA update? It updates in real time each time a conversation resolves and gets scored, typically within a few seconds of the conversation ending.

How do I find contacts most at risk? Go to Satisfaction → By Contact, open the Trend dropdown, and select "At Risk." These are contacts with Negative sentiment who are actively declining — the highest priority for recovery outreach.

How do I enable it? Ask your admin to enable "Satisfaction" in Reports settings, or contact your account manager.