This guide explains, in plain language, how Velaro decides what a conversation was about (its theme or topic), how it decides how the customer felt (its sentiment), how those separate answers are combined in your reports, and every control you have to shape the results. It also walks through each report in the Enterprise Reports package, explains what Year in Review and the New-agent ramp report are, and states the limits of the data so you can describe a number to your own leadership accurately.

If you only need a quick answer, start with the section called "The short version", then jump to the section you need.

The short version

• Every conversation gets a topic (also called a theme) when it closes. Velaro picks it from your own topic list. If nothing on your list fits, Velaro keeps its own short suggestion instead of throwing the chat away, and shows it to you as a Suggested topic.
• Every conversation also gets a sentiment score when it resolves. Topic and sentiment are worked out separately, by separate steps, and they never change each other. Reports put them side by side.
• A topic you set by hand is never overwritten, not by new chats, not by a re-check of older chats, and not by merging or renaming.
• You shape topics by editing your list, loading an industry starting list, merging, renaming and grouping topics, accepting or dismissing Suggested topics, and optionally re-checking older chats.
• Enterprise Reports is a package of operations reports built on top of these labels. It is off by default and Velaro turns it on for your account.
• Everything here is derived from chat conversations only, in English for the word-level signals, and only from the date each signal began to be recorded.

Part 1: How a topic is chosen

A topic answers the question "what was this conversation about?" Velaro works it out once, when the conversation closes. It never delays or blocks the close of the chat, and if anything goes wrong while classifying, the conversation still closes normally and you simply see it without a topic until the next backfill or re-check.

What Velaro looks at

To choose a topic, Velaro looks at a small, deliberately limited set of things from the conversation:

1. The first substantive messages the visitor typed. Velaro skips greetings and filler and reads up to the first three real visitor messages. It does not read the whole transcript to choose a topic, which keeps the cost per conversation very small and keeps the topic anchored on what the visitor came for rather than where the chat wandered.

2. The visitor's pre-chat survey answers, if your widget collects any.

3. The page the visitor was on when they started the chat.

4. The name of the website or group the chat came through.

It does not read agent replies to decide the topic, and it does not use sentiment to decide the topic.

Step one: pre-chat answer pass-through

If you ask visitors to choose a topic in your pre-chat survey, for example a dropdown called "What do you need help with?", Velaro treats that answer as the strongest signal available, because the visitor told you directly.

When the visitor's pre-chat answer is an exact match for a topic on your list, Velaro uses it as the conversation's topic straight away, with full confidence, and does not ask the AI at all. This is called pass-through. It is the cheapest and most accurate way to get clean topics, which is why the advice in the section on growing and improving your topics starts with the pre-chat form.

If the pre-chat answer does not exactly match anything on your topic list, for example the visitor typed free text, or your dropdown wording differs from your topic names, Velaro does not force it. The answer is simply given to the AI as one more clue in step two.

Tip: keep the options in your pre-chat dropdown identical to the names on your topic list, and pass-through will do the work for those chats.

Coming soon: choose which answer sets the topic, and track more than one topic

Two additions are on the way. Until they are released for your account, nothing on this page changes and nothing described here appears in the app.

• Use this answer as the chat theme. In the survey builder, each question will have a switch with this name. Turn it on for the question that tells you what the visitor needs, and its answer sets the conversation's topic. For a multiple-choice question you can also pick the topic each answer maps to, so the wording on your form does not have to match your topic names. An answer with no mapping still works if it matches a topic name. If nothing resolves, the usual pass-through and AI steps run as before.
• Multi-topic conversations. A chat can raise more than one subject. With multi-topic turned on for your site, the first resolved topic stays the main topic and up to two more can be recorded as secondary topics. Chat Intelligence will have an "Include secondary topics" switch to show them. The main topic, and every number you see today, stay as they are unless you turn that switch on.

Both are off until your site is switched on for them. Ask your Velaro contact when they are available for you.

Step two: the AI classifier

When there is no exact pre-chat match, a fast, low-cost AI model reads the clues listed above together with your topic list and returns two things: the topic it believes fits best, and how confident it is.

Your topic list is a hint for the AI, not a fixed rulebook. The AI is told to prefer a topic from your list. It is also told what to do when nothing fits, which leads to the next three outcomes.

The three possible outcomes

A topic on your list. If the AI picks a topic that is on your list and is reasonably confident, that topic is saved on the conversation. This is the normal case, and for a well-built list it is the vast majority of chats.

Other, with a suggested topic. If nothing on your list fits, or the AI's confidence is too low to trust, Velaro files the conversation under Other and keeps the AI's own short topic (two to four words) as a suggestion attached to that conversation. Velaro keeps the suggestion instead of discarding it, because "Other" with no detail tells you nothing, while "Warranty transfer" repeated forty times tells you exactly what topic to add. These suggestions feed the Suggested topics panel described below.

Noise or Uncategorized. This is reserved for conversations that had no real content: a bare greeting, a test message, a visitor who typed nothing of substance. It does not mean "the AI could not decide". If a conversation has real content that merely does not fit your list, it is Other with a suggestion, never Noise.

A conversation that was missed or abandoned before the visitor wrote anything can still be classified from its pre-chat answers and page, so those chats are not invisible in topic reports.

A manual topic is never overwritten

A supervisor or administrator can correct the topic of a single conversation by hand. Velaro records that the topic was set by a person. From then on:

• The automatic classifier will not change it, even if the conversation is classified again for any reason.
• Re-checking older chats (reclassify) skips it.
• Backfill skips it.

The only way a manually set topic changes is that a person changes it again. If you merge or rename the topic itself, the conversation follows the new name, because the label changed, not the decision.

Manual corrections are written to your audit trail with who made them and what the topic was before.

When the topic is chosen

The topic is chosen at conversation close, for chats that go through the normal close path. Chats that closed before topic classification was running on your account have no topic until you run a backfill. Past conversations are never silently re-labelled when you change your list. See "What happens to past conversations" below.

Part 2: Shaping your topics

You control the topic list, and the list is where nearly all of the quality comes from. You find it under Reports, Chat Intelligence, Topics. Making changes requires the Intelligence Suite on your account and a supervisor or administrator role.

Your topic list

Add, rename or remove topics at any time. A list can hold up to sixty topics. Aim for fewer than that. Most teams get the clearest reports with between eight and twenty topics that a new agent could tell apart without a training session. A list of sixty near-duplicates produces noisy reports and makes the AI's job harder, because it has to choose between topics that overlap.

Good topic names are short, written the way your customers and agents talk, and mutually distinct. "Billing question" and "Invoice help" are one topic wearing two names. "Billing" and "Cancel my account" are two different needs.

Changes you make to the list apply to new conversations from that moment on.

Industry starting lists

If you do not want to start from a blank page, load a starting list for your industry and edit it. Starting lists are available for Insurance, Healthcare, Retail, E-commerce, Financial Services, Technology, Government and Manufacturing. A starting list is only a starting point. Rename things into your own vocabulary, delete what does not apply, and add what is specific to you. If you have not loaded a template and have no list, you begin with a blank list and Velaro will rely more heavily on suggestions until you build one.

Merge, rename, retire and group, with no AI cost

These four actions change how your reports present history. They do not re-run any AI.

• Merge combines two topics into one. Every past conversation that had either topic now shows under the merged name immediately.
• Rename changes the label everywhere, past and future, immediately.
• Retire removes a topic from the list so new conversations are no longer filed under it, while keeping your history intact and visible.
• Group rolls related topics up under a group name, for example Billing, Invoices and Refunds under "Money". Velaro suggests groups based on your own topic names and counts. Those suggestions are produced without AI.

Because these actions are instant and free, use them freely. Do not reclassify just to fix a naming problem.

Suggested topics: how your list grows

Whenever a conversation lands in Other, the short topic the AI proposed is kept. The Suggested topics panel (Reports, Chat Intelligence, Topics) groups those suggestions, shows how many conversations each one covers, and shows example chats so you can judge them.

For each suggestion you can:

• Accept it as a new topic on your list. Matching conversations are counted under it.
• Merge it into a topic you already have, when it is a different wording of something existing.
• Rename it before accepting, when the wording is awkward.
• Dismiss it, when it is noise or not worth tracking.

A good rhythm is a five-minute review of the panel once a week. The suggestions with the highest counts are, by definition, the topics your customers raise most that your list does not yet cover.

Re-checking older chats

If you add a topic and want older conversations to be considered against it, you can re-check them. Velaro shows an estimated cost before it runs, and a limit on the number of conversations applies to each run. Re-checking:

• Only touches conversations that are filed under Other, or all conversations if you choose that scope.
• Never changes a topic that was set by hand.
• Is optional. You never need it for merging, renaming, retiring or grouping.

Backfill for conversations that were never classified

Conversations that closed before topic classification was active on your account have no topic at all. A backfill classifies them, newest first, up to the limit you set for the run. It is a one-time or on-demand step. It processes a conversation whether or not the AI finds a topic that fits; conversations with no classifiable content simply stay uncategorized.

If a backfill reports that it stopped early, the usual cause is a problem with the AI service rather than with your conversations, and it is safe to try again later.

Ask for tidy-up ideas

The Analyze topics action makes a single low-cost AI call over your topic counts and your suggested topics. Personal details are removed first. It shows the estimated cost before it runs, saves the result, and can run once a day. It proposes merges, renames and groups. Nothing changes until you accept each proposal. Proposals are limited to names that already appear in what was sent, so it cannot invent topics from nowhere.

The same controls elsewhere

Everything in the Topics page can also be done from the Velaro CLI and from the Velaro MCP tools, for teams that manage settings from scripts or assistants. The behavior is identical.

What happens to past conversations

Past conversations keep the topic they were given. Changing your list never silently relabels history. The exceptions are the actions that are about names rather than decisions: merge, rename, retire and group change how a label is shown, not whether the conversation was judged to fit. Re-checking is the only action that asks the AI to decide again, and it always shows a cost first.

Part 3: How sentiment works, and how it is separate from topics

Sentiment answers a different question from topic: "how did the customer feel?" It is worked out by its own step, at its own moment, and it does not depend on the topic.

The conversation score

When a conversation resolves, an AI model reads the transcript and produces:

• A score from 1 to 10.
• A label: Positive, Neutral or Negative.
• The customer's primary intent.

Roughly, 9 to 10 means clear satisfaction or thanks, 7 to 8 means a generally positive interaction where needs were met, 4 to 6 is a neutral factual exchange, 2 to 3 means frustration or complaints, and 1 means a hostile or highly escalated conversation worth following up on. No setup is needed beyond having the Satisfaction report on your account.

There is exactly one overall sentiment label per conversation. It is the verdict on the whole chat. It is not a minute-by-minute arc.

Contact-level sentiment

Every contact keeps two running scores that update each time one of their conversations is scored:

• The all-time average, where every conversation counts equally. Use it for long-term relationship assessment.
• The recency-weighted score, where recent conversations count for more than old ones. Use it to see how the relationship feels right now.

When the recency-weighted score is higher than the all-time average, the customer has been happier lately. When it is lower, recent contact has been worse than their history, which is an early churn signal. A trend arrow tells you whether the contact is improving, stable or declining, and contacts who are negative and declining are flagged as at risk. A confidence badge tells you how many scored conversations stand behind a number: under five is directional only, five to fifteen is reasonably representative, and over fifteen is solid enough to act on.

Message-level sentiment for the Enterprise Reports

Some Enterprise Reports need to know how the visitor felt at a particular moment, for example just before a handoff from the bot to a person. For that, Velaro records a derived sentiment signal for each visitor message when a conversation closes, using a fixed English word list rather than AI. This is separate from the overall conversation score. It is covered in the section on limits.

Other classifications that run separately

Topic and sentiment are two of several independent labels. Each is its own step, and none uses another's result as input:

• Quality score (QA). The AI scores the conversation against your rubric, and supervisors can review and override. See your Quality QA settings.
• AI field extraction. Reads each resolved conversation and fills fields such as line of business, intent and state. With automatic tagging turned on, those become conversation tags.
• Customer effort. An inferred 0 to 100 score from how the chat behaved (repeats, re-asks, transfers, waits).
• Promise detection. Finds commitments agents made in chat.
• Keyword extraction. Surfaces the most common words and phrases, with growth.

How the labels combine in reports

Reports combine labels by joining them on the conversation. Each conversation carries its topic, its sentiment, its QA score and so on, and a report simply groups by one label and measures another. Examples:

• "Sentiment by topic" groups conversations by topic and shows the share positive, neutral and negative in each.
• The topic radar ranks topics by volume, growth, transfer rate and handle time.
• The agent-by-topic skill-gap matrix groups by agent and topic and shows quality and sentiment together.
• The coaching queue combines the QA score with whether a supervisor has reviewed the chat.

Because the labels are independent, a conversation can have an excellent topic match and a negative sentiment (the customer was correctly identified as asking about a late delivery, and was upset), or a poor topic fit and a positive sentiment. This is a feature: it lets you ask "which topics make customers unhappy?" without the answer being tangled up in how the topic was chosen.

It also means an improvement to your topic list improves topic reports without changing a single sentiment number, and vice versa.

Part 4: Growing and improving your topics

Here is a practical routine for a team that wants clean, trustworthy topic reports.

1. Start with a short list. Load the industry starting list closest to you, delete anything that does not apply, and rename the rest into your own words. Eight to twenty topics is plenty.

2. Use a topic dropdown in the pre-chat survey with options that match your topic names exactly. Those chats are classified by pass-through, at no AI cost and with full confidence.

3. Review Suggested topics weekly. Accept the high-count suggestions, merge near-duplicates into existing topics, and dismiss the noise.

4. Merge and group, do not multiply. If two topics read the same in your reports, merge them. Group related topics so leadership can see the big picture and still drill down.

5. Retire topics you no longer use instead of deleting history.

6. Correct individual mistakes by hand. A manual correction sticks forever and is audited.

7. Re-check older chats only when it is worth it. Do it after adding a major new topic, and read the cost estimate first.

8. Watch the Other share. If more than a small slice of conversations lands in Other, your list is missing topics. If Noise is high, check for bot test traffic or very short conversations.

Multi-topic chats and the survey theme anchor

Some conversations genuinely cover more than one subject, and a post-chat survey can be tied to a topic so that survey results are reported by topic. Support for tracking more than one topic in a conversation and for anchoring survey results to a topic is being built and is not yet part of this guide's description of what is live. Until it is announced as available, each conversation has one topic, the one described in Part 1, and survey results follow that single topic.

Part 5: The Enterprise Reports package

Enterprise Reports is a package of operations reports for leaders who need to compare programs, coach agents and spot service problems. It is built on the labels described above: topics, sentiment, QA scores and the other classifications.

Where it is and how to get it

Enterprise Reports is off by default. It is not part of a standard plan and is not bought from a menu. Velaro turns it on for your account. Each report in the package is switched on separately, so you may have some reports and not others. Ask your Velaro contact to enable the package or a specific report.

When it is on, you find it under Reports, Enterprise Reports, with a tab for each report. When a report is off for your account, it is hidden in the menu, and the server also refuses requests for it, so it cannot be reached by a direct link.

Some reports build on features you already use, such as conversation reports, quality scoring, sentiment reports, theme classification or the Chat Intelligence report. If one of those is not on your account, the dependent report will not show data until it is.

The reports

1. Program, website and platform breakdown. One report that compares every website (program), team and platform side by side: chats, missed chats, transfers, wait time, first reply time, QA score, QA coverage and negative sentiment. Use it to see which program is struggling without running seven separate reports. You can group by website, team, platform, device or entry page.

2. Skill routing by entry point. Not a report page but a routing capability in Skill routing, Requirements. A requirement can apply only when the chat starts on a page whose address contains some text, when a query-string parameter has a value, when a custom value matches, or when a pre-chat answer matches. Requirements with no condition apply to every chat as before. The simulator accepts a page address and custom values so you can test before turning a requirement on. Skill routing narrows which agents on the team can take the chat; the team itself still comes from the website or workflow.

3. Individual agent scorecard. Each agent by name: chats handled, resolution rate, transfers, handle time, first reply, AI score, supervisor-reviewed score and sentiment impact. Sentiment impact is the agent's net sentiment (positive minus negative) compared with every other human-handled chat, in percentage points. Because the platform records one sentiment label per chat, this is a comparison to the team, not a start-to-end arc.

4. QA coverage and coaching queue. Shows the share of resolved chats that the AI scored and supervisors reviewed, by week and by website, and lists low-scoring chats nobody has reviewed, per agent. A supervisor review in Quality QA takes a chat out of the queue. Coverage below 100 percent usually means a rubric samples fewer than every chat, a monthly scoring cap was reached, or the chat had too few messages. The report tells you which settings apply.

5. Sentiment by program, agent and topic. The negative share of conversations by website, team, agent or topic, with the change against the previous period of the same length. This is where the labels from Part 3 meet.

6. Keyword intelligence and topic radar. Top keywords and phrases with growth, plus a radar that ranks topics by volume, growth, transfer rate and handle time, per program. Use it to spot a topic that is growing and expensive.

7. Automatic tagging. AI field extraction reads each resolved chat and fills line of business, intent and state. With automatic tagging on, the line of business and intent also become conversation tags, so agents stop typing them. Turn on the option to create tags in the AI Field Extraction settings.

8. Weekly executive brief. A Monday-morning email in your time zone with program comparison, QA coverage, the coaching queue, the topic radar with service opportunities and, if you have the Intelligence Suite, the health score and volume forecast. Use the test-send on the Weekly brief tab to check delivery without waiting for Monday.

9. Bonus insights. The QA coverage gap (manual versus AI review, with the reviewer hours it would take), next-best-fix opportunities, an early warning for escalation risk, the agent-by-topic skill-gap matrix, handle time versus quality outliers and root causes of abandoned chats. Every dollar figure shows its assumptions so you can change them to your own.

10. Operations Briefing. One run that produces every package report from a single load of the data, plus one short AI summary that is cached and size-limited. It reports its own AI cost so the summary is never a surprise.

11. Repeat-contact leakage. Of the chats that closed in the range, how many came back from the same contact within a return window (1, 3, 7, 14 or 30 days). It counts only resolved chats from identified contacts that ended at least one full window ago, and reports how many were too recent to judge. It shows how many came back on the same topic, how long they took to return, whether the first chat was closed by a bot or a person, and the split by agent, topic and program. A return is a signal, not proof the first chat failed, because people also come back with new questions; the same-topic figure is the closer hint. Visitors who never identified themselves cannot be matched.

12. Demand versus coverage heatmap. A day-by-hour grid of the chats that needed a person against the agent hours actually available, in the time zone shown on the page. Chats the bot handled alone are shown separately. This report ignores the From and To dates and uses the last 1 to 12 whole local weeks you choose, averaged per week. Coverage means agent hours in the Available status, measured from status changes; Busy and Away do not count. Cells are marked as gaps when chats arrive and nobody is available, or when pressure is higher than an agent can handle at the handling time and concurrency shown on the page. It also lists the biggest gaps, after-hours volume and a next-week forecast. It does not say how many sales were lost.

13. Bot false containment and handoff regret. Claimed containment is the share of bot-touched chats the bot closed with no handoff, which is what a bot dashboard would claim. True containment removes the bot-only chats where, within 72 hours, the same contact started another chat that reached a person or was about the same topic. Handoff regret counts chats handed to a person that then ended badly: negative sentiment, a negative visitor message just before the handoff, a QA score below your low-score threshold, or missed or abandoned. Some handoffs are policy, such as refunds or account changes, and are not a bot failure.

14. Inferred customer effort index. Each chat is scored from 0 to 100 by adding signals: transfers, reopens, repeated visitor messages, details the visitor had to give twice, a long wait, an unusually long exchange, a return within seven days and clearly frustrated messages, capped at 100. The page shows the median and the 90th percentile (never the average), the share of chats that could be scored, how much each signal adds, and the split by topic, program and team. Higher means more effort. It is inferred from behavior and is not a survey, so it must not be compared with a customer effort survey score.

15. Promise ledger. Commitments agents made in chat, such as "I will follow up" or "someone will call you", whether a follow-up appeared, which are overdue, and by agent. Only messages written by human agents are read. A promise is Kept when an agent sends another message in the same conversation at least five minutes later, or the same contact gets a new human-handled conversation, before the deadline. Past the deadline with neither, it is Overdue. Managers and administrators see the promise sentence, with emails, phone numbers and order numbers replaced by placeholders, and can Dismiss a promise that was kept another way. Other users see counts and the split by agent only. Detection uses fixed phrases, so some real promises are missed and a few listed ones are not promises. Phone calls and personal email are invisible, so Overdue means check, not broken.

Year in Review

Year in Review is the next report in the package, a once-a-year look back at a program or team across the year. It is being built and is not yet available on your account. Until Velaro announces it, do not expect to find it in the menu. When it is released it will be switched on the same way as the other reports.

New-agent ramp

The New-agent ramp report is also being built and is not yet available. It is designed to show how quickly newly hired agents reach the team's typical handle time, quality and resolution, so you can see where onboarding is working. It will be switched on the same way as the other reports once released.

How the package is organized

The package behaves like a bundle with one item per report. Switching on the whole package turns on every report. Switching on one item turns on only that report. A package can also be included as part of another, larger bundle, which then switches on everything inside it. In every case a report that is off is hidden and refused by the server, not merely hidden in the menu. Customers cannot enable it themselves; ask your Velaro contact.

Part 6: Limits you should know

Be exact about these when you present a number to others.

• Chat only. Everything here is derived from chat conversations. Phone calls, personal email and anything that happened outside Velaro chat are invisible. A promise kept by phone looks Overdue.
• Dates matter. Some signals can only be recorded while the chat text still exists. Per-message sentiment, the bot-phase history and the frustrated-message part of the effort score are recorded when a chat closes, from the date capture became live on your account. Chats that closed before that show Insufficient data for those parts. They are never estimated, and they cannot be filled in later, because the text has since been cleared.
• Insufficient data means insufficient data. Never describe a figure as available when the page shows Insufficient data.
• English only for word-level signals. Per-message sentiment uses an English word list, not AI, so messages in other languages are left unscored. Overall conversation sentiment is scored by AI and is not subject to the word list.
• One topic and one sentiment label per conversation. There is no start-to-end sentiment arc and, today, no second topic on the same chat.
• A topic list is a hint. The AI can still choose a topic you would not have. Correct it by hand and the correction sticks.
• Inferred is not surveyed. Effort, true containment and repeat contact are inferred from behavior. They point at where to look. They are not proof.
• Anonymous visitors cannot be followed. Anything that needs the same person to return needs an identified contact.
• Large ranges can be truncated. Some reports use the most recent chats in a very large range and say so on the page.
• Time zones. Reports bucket by calendar days in the time zone shown on the page. A date you type is a whole day in that time zone.
• Costs and caps. Re-checks, analyses and the Operations Briefing show an estimated cost before they run, and run within limits.

Part 7: Common questions

Why was this chat filed under Other? Nothing on your list fit, or the AI was not confident enough. Open Suggested topics: the AI's own short topic for that chat is there. Accept it, or merge it into an existing topic.

Why is a chat Noise? It had no real content, such as a greeting or a test message. A chat with real content never becomes Noise.

I changed my topic list. Why did old chats not change? Past chats keep the topic they were given. Merge, rename and group change labels immediately and cost nothing. Re-check is the way to ask the AI to decide again.

I fixed a chat's topic by hand and it changed back. It should not. A topic set by a person is never overwritten by the classifier, a re-check or a backfill. If a merge or rename changed its label, that is the label changing, not the decision.

Why does the pre-chat answer win over the AI? Because the visitor told you directly. An exact match to a topic on your list is used as is.

Is sentiment worked out from the topic? No. They are separate steps and neither uses the other.

Why is a report showing Insufficient data? The signals it needs were not being recorded when those chats closed. They cannot be reconstructed.

Why can I not see an Enterprise Report? The package or that item is not switched on for your account. Ask your Velaro contact.

Can I see Year in Review or the New-agent ramp today? Not yet. Both are being built.

Do these reports cost extra AI usage? Topic and sentiment classification cost a small amount per conversation and are part of normal operation. Optional actions such as re-checking older chats and Analyze topics show an estimate before they run.

Where to look in the app

• Reports, Chat Intelligence, Topics: your list, templates, merges, groups, Suggested topics, re-check and Analyze topics.
• Reports, Satisfaction: sentiment overview, breakdown, by contact and by agent.
• Reports, Keyword Intelligence: top keywords and growth.
• Reports, Enterprise Reports: the package reports, one tab each, when enabled for your account.
• Reports, Email Subscriptions: the weekly executive brief.
• Quality QA settings: the rubric, sampling and supervisor review.
• AI Field Extraction settings: fields and automatic tagging.
• Skill routing, Requirements: entry-point conditions.
• Your pre-chat survey settings: the topic dropdown that powers pass-through.