1. What this dashboard is for
To follow the performance of ClaudIA's "Agentes Especialistas" (Specialist Agents): how much of the volume goes through them, how much they resolve on their own, how much goes to human support, how the customer rates it, and where the tools are failing.
2. Three rules that apply to the whole dashboard
1. Only conversations that are already finished get in. An ongoing conversation has no outcome — counting it as resolved would inflate retention.
2. The most recent days read low. The Langfuse load is daily, so the day's agentic activity has not fully arrived yet. Do not read the last bar or the last point as a real drop.
3. Each metric has its own denominator. Retention and handover are about the conversations with a defined outcome; GenCX is about the scored conversations (today, roughly half of them); "Erros" (Errors) counts events, not conversations. That is why the numbers do not match one another.
Every definition question is answered in the tile itself. Each chart has an ⓘ icon next to the title with the definition of the metric, the denominator used and whether or not it responds to the "Especialista" (Specialist) filter.
3. Filters
| Filter | What it does | Note |
|---|---|---|
| "Data da conversa" (Conversation date) | Period selection | Default: last 30 days |
| "Agrupar por" (Group by) | Granularity of the time axis | Hour, Day, Week or Month (default: Week) |
| "Tag" | Conversation tag | |
| "Nível de resolução" (Resolution level) | N1 (ClaudIA) or N2 (human) | |
| "Motivo de resolução" (Resolution reason) | Reason recorded at closing | |
| "Teste A/B" (A/B test) | Key of the gradual rollout experiment | |
| "Especialista" (Specialist) | Filters by sub-agent — it sits inside the "Especialistas" (Specialists) section, not in the top bar | Applies only to the specialist, tool and tool error tiles. See the question in section 6 |
4. Glossary
| Term | Meaning |
|---|---|
| "Agente Especialista" (Specialist Agent, or sub-agent) | Agent with a specific scope that the supervisor triggers inside a ClaudIA conversation |
| "Ativação" (Activation) | Share of ClaudIA conversations that went through the Agentes Especialistas |
| N1 / "Retenção" (Retention) | Conversation resolved by ClaudIA, with no human |
| N2 / "Transbordo" (Handover) | Conversation transferred to human support |
| Tool | Tool that the specialist calls to look something up or to execute something |
| CSAT | Rating given by the customer: positives ÷ responses received |
| GenCX | A 1 to 5 score generated by a model that reads the conversation — it is not the customer answering |
| "Erro" (Error) | Occurrence of an agentic failure (a single conversation may have several) |
| Median / P80 / P90 | Typical case / the value that 1 in every 5 exceeds / the one that 1 in every 10 exceeds |
5. What is in each section
"KPIs principais" (Main KPIs)
| Tile | What it shows |
|---|---|
| "Ativação" (Activation) | How much of ClaudIA's conversations went through the Agentes Especialistas. The numerator is the same number as the "Conversas" (Conversations) card, which is why it is not repeated |
| "Retenção" (Retention) | Conversations resolved by ClaudIA (N1) over the ones with a defined outcome. Below it, how many they were |
| "Transbordo" (Handover) | Conversations transferred to a human (N2) over the ones with a defined outcome. Below it, how many they were |
| "Conversas" (Conversations) | ClaudIA conversations that went through the Agentes Especialistas in the period |
| CSAT | Positive ratings over ratings received. Below it, how many responses |
| "Conversas com erro" (Conversations with an error) | Conversations with at least one agentic error, and how much that represents of the period's total |
| "Ativação por período" (Activation by period) | The line is the activated share; the bars are how many conversations ClaudIA handled |
| "Tempo até a primeira resposta" (Time to first response) | How long the customer waited, in seconds. Bottom line = median; top line = P80 |
"Retenção e Transbordo (N2)" (Retention and Handover)
| Tile | What it shows |
|---|---|
| "Volume de conversas — agêntica" (Conversation volume — agentic) | Conversations per period that went through the specialists |
| "Retenção e transbordo — agêntica" (Retention and handover — agentic) | Split of each period between resolved by ClaudIA (N1) and transferred to a human (N2) |
| "Motivo do transbordo — agêntica" (Handover reason — agentic) | Why the conversation went to the human. "Repasse automático" (Automatic handoff) is the default handoff, which records no reason |
CSAT
| Tile | What it shows |
|---|---|
| "Taxa de resposta do CSAT — agêntica" (CSAT response rate — agentic) | How many of the conversations handled by ClaudIA answered the survey |
| "CSAT — agêntica" (CSAT — agentic) | CSAT of each period in the line; in the bars, how many responses back the number |
| "Notas de CSAT por período" (CSAT scores by period) | Distribution of the 1 to 5 scores given by the customer |
| "Notas de GenCX por período" (GenCX scores by period) | Distribution of the 1 to 5 scores generated by the model that reads the conversation |
"Especialistas" (Specialists)
| Tile | What it shows |
|---|---|
| "Especialistas — latência típica e P90" (Specialists — typical latency and P90) | Median and P90 latency of each specialist. Top 12 by P90 |
| "Especialistas — retenção e transbordo" (Specialists — retention and handover) | Each specialist split between retained (N1) and handed over (N2); the third slice is the conversation that did not match with ClaudIA. Top 12 |
| "Especialistas — acionamentos" (Specialists — triggers) | How many times each specialist was triggered. Top 12 |
| "Especialistas — agêntica" (Specialists — agentic) | Triggers, latency and error rate of each specialist |
| "Conversas por especialista — agêntica" (Conversations per specialist — agentic) | Handover, retention, CSAT and GenCX of each specialist |
Tools
| Tile | What it shows |
|---|---|
| "Tools — volume × latência" (Tools — volume × latency) | Each dot is a tool: calls × average latency, size by the number of errors and color by the error rate |
| "Profundidade da conversa — agêntica" (Conversation depth — agentic) | Messages, tool triggers and distinct tools per conversation, over the period |
| "Especialista × tool — mapa de calor" (Specialist × tool — heat map) | Error rate of each specialist × tool pair. The 80 most called pairs |
| "Tools — volume e latência — agêntica" (Tools — volume and latency — agentic) | Most triggered tools, with average latency and error rate |
| "Falha por tool — agêntica" (Failure by tool — agentic) | Calls, failures and failure rate of each tool |
"Erros" (Errors)
| Tile | What it shows |
|---|---|
| "Erros ao longo do tempo — agêntica" (Errors over time — agentic) | Errors per period. "Erros" (Errors) counts each occurrence; "Conversas com erro" (Conversations with an error) counts distinct conversations |
| "Motivos de falha de tool — ranking" (Tool failure reasons — ranking) | Error reasons sorted by volume. Top 10 |
| "Motivos de falha de tool — agêntica" (Tool failure reasons — agentic) | Table of the reasons. "Conversas afetadas" (Affected conversations) is how many conversations felt the error |
Tickets
| Tile | What it shows |
|---|---|
| "Amostra de conversas com agêntica" (Sample of conversations with agentic activity) | One row per conversation that triggered specialists. The link opens the conversation in the Hub |
6. Frequently asked questions
Retention + Handover adds up to 100%, but it does not match the "Conversas" (Conversations) card. Why? Retention and handover are calculated only over the conversations that matched with ClaudIA and have a defined N1/N2 outcome. The "Conversas" card is the total that went through the specialists. Different denominators.
Why does summing the "Conversas por especialista" (Conversations per specialist) column give more than the total number of conversations? A conversation that went through two specialists counts in both. The sum is of triggers, not of unique conversations.
What is the difference between CSAT and GenCX? CSAT is the customer answering the survey. GenCX is a 1 to 5 score generated by a model that reads the conversation. GenCX covers more conversations (it does not depend on the customer answering), CSAT is the customer's voice.
Why is "Erros" (Errors) higher than "Conversas com erro" (Conversations with an error)? "Erros" counts each occurrence; "Conversas com erro" counts distinct conversations. A conversation with 3 failures adds 3 to one and 1 to the other. The number of distinct conversations matches the card at the top.
Why does last week's activation appear close to zero? The Langfuse load is daily, so the agentic activity of the most recent days has not fully arrived yet. Wait for the day to close before concluding anything.
Where is the specialist filter? I cannot find it in the top bar. It sits inside the "Especialistas" (Specialists) section, right below the section title.
Why does the "Especialista" (Specialist) filter not change the Retention card (nor CSAT, nor the charts by period)? Because those numbers are per conversation, and one conversation may go through more than one specialist — there is no way to attribute it to a single one. The filter applies to the tiles built on specialist triggers: the ones in the Especialistas section, the ones in Tools and the tool error ones. Each tile says in its own tooltip (ⓘ) whether it responds or not.
Do Cloud Humans' internal tools count towards tool failure? No. The "Falha por tool" (Failure by tool) tile excludes the tools that operate the Cloud Humans platform — they are not customer operation and they fail much more often.
What is "(tool não identificada)" (unidentified tool)? Old failures in which the tool name was not recorded. It decreases as the load moves forward.
What is "Repasse automático" (Automatic handoff) in the handover reason? It is the default handoff to a human, which does not record a specific reason.
Does an ongoing conversation show up in the dashboard? No, in no tile. Without an outcome, it would distort retention, handover and conversation depth.
Where do I see a specific conversation? In the Tickets section, tile "Amostra de conversas com agêntica" (Sample of conversations with agentic activity) — the link on each row opens the conversation in the Hub.