AI conversation & client analytics

Two AI-written reports sit next to every chat: AI conversation analytics judges one conversation (how the client behaved, how the manager performed, what to do next), and AI client analytics builds a portrait of the whole client across every channel. Both are generated by a language model from your data, cached, and recomputed on rules described below.

In the app

• Open a chat on Active chats and use the toolbar toggle Show transcript / Hide transcript / AI analytics, or the client card tabs AI <agent> / AI conversation analytics / AI client analytics.
• Settings: avatar → Settings → AI conversation analytics settings (/profile/aiConversationAnalytics) and AI client analytics settings (/profile/aiClientAnalytics).

Who can see and change it

Action	Who
Read a report	Anyone with the Analytics access permission (Admin, Sales Lead, Manager and Analyst by default — see Team & access)
Open the settings pages	Same permission
Save the settings	Admins only — deliberately not delegable, because Analytics access is granted to managers by default

Reports are not billed — they draw nothing from the message balance and do not appear on the Analytics dashboard.

The conversation report

Panel title AI conversation analytics, two tabs: AI analytics and KPI. Export downloads the current report; Recalculate report forces a rebuild ("AI will re-analyze the data and update this report. It may take a few seconds.").

AI analytics tab

The blocks you switched on under Output blocks, plus Script data — the labeled blocks your AI prompt (for example your call script) asks for, such as "CRM", "Budget" or "Next step", or a default set when the prompt names none:

Block	Contents
General conclusion	Short summary plus Conversation tags (from your tag vocabulary, e.g. AI sales, Price, Objection, Payment, Prepayment)
Client evaluation	Lead status (Hot / Warm / Cold …), Lead quality, Target fit, Worth the time, Sales potential, Loss risk, Recommended action (Process urgently, Needs follow-up, Monitor)
Manager evaluation	Strengths, Mistakes, Work level (Novice → Proficient)
Red flags	The thresholds you enabled, each with its severity — Critical, Attention, Moderate
Lost opportunities, What to do better	Missed openings and concrete improvements
Recommended next step, Task for manager, Task for supervisor	Actionable follow-ups

The report only judges the manager when it can tell the manager's words from the client's. When the only material is a call transcript whose manager/client split could not be determined, Manager evaluation, What to do better, Lost opportunities, Task for manager and Task for supervisor are left out. A split is undetermined when the recording was never split into turns, or when the AI role split failed and the roles were only guessed from who spoke first. General conclusion, Client evaluation and Recommended next step still appear, because they describe the conversation and the client rather than the manager. Calls stored before guessed roles were marked are still graded as before.

KPI tab

KPI	Meaning
Call duration, Manager talk time, Client talk time, Talk ratio	From the call recording and metadata (calls only). Manager talk time, Client talk time and Talk ratio stay empty when any call in the chat has guessed speaker roles (the AI role split failed and the roles were inferred from turn order), since a guess can swap the two
First response	Time to the manager's first reply
Messages	Message count in the conversation
Key conversation points	The theses the model extracted
Conversation result	Product, Deal amount, Prepayment, Fixation
Conversation sentiment	Positive / Neutral / Negative
AI recommendations	Free-text advice
Criteria scores and Weighted score	One score per Manager evaluation criteria row, combined by the weights you set. "Not assessed" means "This conversation gave the manager no chance to demonstrate these criteria, so it is not scored." Every criterion is also "Not assessed" when no part of the conversation can be attributed to the manager: a call transcript that was never split into manager and client, or whose roles were only guessed from who spoke first, is not used to grade the manager

When a conversation report is (re)computed

Reports are cached per conversation. The cache is used unless one of these makes it stale:

Trigger	Effect
A call ends	The report is rebuilt automatically in the background — this is the "Analytics update: After every conversation" in the settings. Binotel and UniTalk calls always trigger it; AI calls (Calls) and Ringostat calls trigger it only when the account has saved its analytics settings. Chat channels do not auto-trigger
The report is older than the last call, or the chat has been silent for 1 hour and the report predates that silence	The next open recomputes it. Any new message resets the one-hour window
Recalculate report	Immediate rebuild; also invalidates the client's cached reports
Any change to the analytics settings	Every cached report on the account becomes stale (the cache key includes a fingerprint of the configuration)

If the settings cannot be loaded when a call ends, the automatic build is skipped rather than producing a report on default settings. Every build also stores a snapshot on the chat, which the lead card and the chat list indicator read. The Quality score on Quality Control is separate: a scheduled AI reviewer scores finished handoffs on its own (see How the Quality score works).

Once the account has saved its analytics settings, each conversation report is also written as a note on the client, so the lead card and the agents see it right after a call. A rebuild replaces that chat's previous note instead of adding another one.

The client report

AI client analytics — "Based on all interactions" — with two sub-tabs:

• Teamwork and progress: Current status (Hot / Warm / Cold lead), Funnel stage, Next, Total tasks (completed / in progress / overdue), Active managers, Deal probability (High ≥ 70, Medium ≥ 40, else Low), then Client status history, Manager work quality (strong communication / stable performance / needs supervision), Client tasks, Key team actions, Problem points, Next steps, and AI-Veronika recommendations. Formulate action plan generates a Client action plan.
• Interactions overview: General client description, Total interactions (+N this week), Average rating, Needs attention, Last activity, an Interaction chronology, Activity by day and hour heat map, Communication channels, Main topics and issues, Insights and recommendations, Tasks and statuses.

Date presets: Last 7 days, Last 30 days, Last 90 days, Custom range.

Rule	Detail
Availability	Only for standard client profiles or chats linked to a unified client — otherwise "Client analytics unavailable"
Recompute	No automatic staleness rule. A client report is served from cache until you press Recalculate report, change the analytics settings, or refresh a conversation report for that client. The "Daily at 02:00" label on the settings page describes the intended schedule, not a rebuild you can rely on
Timezone	Part of the cache key — switching your timezone forces a rebuild
Lead card fields	Deal probability, Manager performance quality and Overall lead score on the lead card and chat list come from the last stored snapshot; a fresh 30-day report only fills gaps. A conversation report in which no criterion could be assessed (for example a call with guessed speaker roles) clears the stored Manager performance quality to "—" instead of keeping an older number

Report language

Both reports are written in one language, resolved in this order:

1. The Response language saved in AI client analytics settings.
2. Otherwise the one saved in AI conversation analytics settings (Ukrainian or English).
3. Otherwise the most common reply language of your active assistants.
4. Otherwise Ukrainian.

The Automation block on the client settings page shows the Response language the reports will be written in — the saved one, or the resolved fallback when nothing is saved yet.

Settings

Both settings pages save automatically ("Changes saved automatically") and have Reset all settings. The Test run buttons are placeholders ("will be available in a future release").

AI conversation analytics settings

Section	What you set
Conversation data sources	Transcript, Audio / call recording, Call metadata, Communication channel, Lead card, Previous contact history, Manager messages, Post-conversation notes
Analysis context	Analysis type (Call / Chat / Email), Response language, Report format (Short structured / Detailed), Consider client history
Manager evaluation criteria	Weights for Need identification, Client qualification, Solution presentation, Objection handling, Next step, Response speed / structure, Tone & script compliance — the total must equal 100% or the page will not save
What the system determines	Lead status, Lead quality, Target client / fit, Sales potential, Manager strengths, Manager mistakes, Loss risks, Missed opportunity, Next step
Thresholds & red flags	No next step, Need not identified, Budget / timeline not mentioned, Client has negative sentiment, Manager responds too slowly, No relevant offer / demo / test — each with severity Critical / Attention / Moderate
AI prompt	Extra instructions, such as your call script, max 8 000 characters
Output blocks	Which blocks the report shows
Conversation tags	Up to 20 tags; duplicates are rejected (the whole list goes into every prompt)

AI client analytics settings

Section	What you set
Data sources	Chats, Calls, Leads, CRM tasks, Manager notes, Funnel stages, Client ratings, Social & messengers, Email, Status history
Analysis period & client merging	Period (Last 30 / 90 days / All time), Merge by (Phone / Email / Client ID), Deduplication window (15 min – 3 h), all contacts vs completed only
What we calculate	Total interactions, Average rating, Last activity, Interaction timeline, Heatmap, Communication channels, Main topics and issues, Tasks and statuses — each with its formula hint
Manager performance quality	Response speed, Response completeness, Issue closure, Has next step, Active managers count, Client status change, Tasks done / overdue
AI prompt, Output blocks	As above

Use cases

• Coach a new manager — open their handed-over chats, read Manager evaluation and Criteria scores, and set the criteria weights to match what you coach on.
• Prioritise a morning — sort by Lead status = Hot and Recommended action = Process urgently across yesterday's conversations.
• Call floor QA — every Binotel/UniTalk call gets a report automatically (AI and Ringostat calls too, once the settings are saved), so the lead card carries the analysis without anyone opening a chat.

Test it

1. On AI conversation analytics settings, set weights that sum to 100% and save — the page confirms "Settings saved"; try 90% and confirm the validation message.
2. Open a chat that has a few messages and switch the toolbar to AI analytics — the report builds in a few seconds and shows Updated ….
3. Write one more message in that chat and reopen the report immediately: it is served from cache. Wait an hour (or press Recalculate report) and confirm the timestamp changes.
4. Change one output block in the settings and reopen any report — it rebuilds, because the settings fingerprint changed.

See also

• Analytics & Quality Control — the separate Quality score for human managers.
• Chats & client analytics — the client card where the reports live.
• Team & access — the Analytics access permission.
• Balance & billing — why reports are free.