Reports
How well your brand's AI agent answers, and a browsable list of every answer it has given.
Where this is
Brand → (your brand) → Vex → Reports. One screen now holds the whole answer-quality picture: the summary figures, the conversations the AI carried on its own, and — at the bottom — Answers, the browsable list of every individual reply. That list used to be a separate Responses tab. The tab is gone; nothing went with it.
Everything here is about this brand's AI agent, over the date range you choose in the top right. Opening the screen needs the Vex permission.
The three figures
Successful responses
Of the real questions asked, the share the AI handled — answered from the brand's content, connected a human the visitor asked for, or ran a workflow.
Neutral responses
Messages that asked nothing: greetings, thanks, acknowledgements. Shown as a share of all responses.
Failed responses
Real questions the AI could not answer — missing knowledge, a handoff it chose itself, or a generation error.
AI-handled conversations
Under the figures is the list of conversations the AI carried without a person stepping in. Its title shows what share of all conversations in the range those were.
Conversations that needed no human intervention at all.
One long-running chat is split into segments at least 180 minutes apart, and each segment is counted separately.
The badges on each row break that conversation down into successful, neutral and failed replies.
Selecting a chat ID opens that conversation in the inbox, in a new tab.
Answers
The Answers section at the bottom lists every reply this agent has written, newest first, with its own search, an outcome filter, a Partial answers filter (replies that covered only some of the topics a message asked about) and a date range. Each row carries an outcome badge:
The AI answered the question using content retrieved from the brand's knowledge, or an API tool. (Counts as successful)
The AI handled a message that needed no content — a greeting or a simple pleasantry. (Counts as neutral)
The AI did not have enough information to answer the question. (Counts as failed)
The visitor explicitly asked for a person, and the AI handed over. (Counts as successful — the AI did the right thing)
The AI found nothing relevant, or a tool failed, so it handed over by itself. This one points at a gap in your knowledge. (Counts as failed)
The AI started a workflow that resolved the turn. (Counts as successful)
Something went wrong while the answer was being generated. (Counts as failed)
Inspecting a single answer
Use the Details action on a row to expand it.
Conversation
The visitor's question and the AI's reply, rendered exactly as the visitor saw them — including any product carousel the AI attached. When a workflow's Send to VEX AI step started the turn, the question side is labelled Workflow hand-back instead of Visitor: the text there is the step's instruction, which the visitor never typed.
Details
Six fields summarise the answer at a glance:
| Field | Meaning |
|---|---|
| Outcome | The badge described above. |
| Channel | Where it happened — web chat, Facebook, Instagram, email or phone. |
| Grounded | Whether the reply was built on content found in the brand's knowledge. |
| Model | The AI model that wrote it. |
| Tokens | Total tokens the turn consumed. |
| Latency | How long the answer took to produce. |
Below the six fields, one or two lines explain how the reply came about when the steps alone would not tell you:
- why a turn has no search step — the message was a plain greeting handled without a search; or the search found nothing and the AI judged the reply needed no company facts (small talk, a follow-up); or it found nothing and said so;
- for a message that asked about several topics — how many were found, and which ones the knowledge base had nothing on.
Steps
A numbered timeline of what the AI did before replying. Each step shows a status and opens for detail:
- Knowledge search — the search it ran and the pieces of content it matched, each with its source title and a relevance score. Open one to read the text itself. When the AI searched in a different language from the one the visitor wrote in, the step says so. A message with several topics shows one search per topic, labelled part 1 of 2, part 2 of 2; at the Deep search depth the extra phrasings appear as alternative phrasing; and when the AI's own query missed and it retried with the visitor's exact words, the step says retried with the visitor's own words.
- API tool — an external tool the AI called, with the data sent and received.
- Workflow — a workflow the AI started.
- AI-written message — a workflow message step with Let the AI say it on. Under Raw data, the input shows what you typed and the output what the AI said. A failed one means your typed text went out unchanged, and the error says why — for example
answer_cap(the visitor had used their daily answers) orai_paused.
What these numbers leave out
Improving what you see here
Read the failures first
Filter Answers by Failed and Handoff (no context). Those two are the questions your content does not cover yet.
Open the steps
Expanding an answer shows what the AI searched for and what it found — the fastest way to tell "nothing matched" apart from "the wrong article matched".
Fill the gap
Add what is missing on Brand → Knowledge, or on Brand → Products when the question was about your catalogue. Anything you add becomes searchable straight away.
Decide what happens when it cannot answer
When the AI doesn't know the answer lives on Personality, on the Answers tab. Set it to offer a human if you would rather unanswered questions reach your team than be politely declined.