User Guide
Everything you need to know about ClarityCX — a product of SONKAY
1.Overview
ClarityCX (a product of SONKAY) is a CX intelligence platform that analyses customer conversations across all channels — phone, chat, email, bot, and social — using four specialised AI analysts and a unified 12-category taxonomy. Audio is transcribed to text first, then every conversation's transcript text is analysed, scored, and categorised so you can spot cost recovery opportunities, coach agents, and catch emerging issues before they escalate.
- •Four AI analysts: Coaching, Opportunity, Efficiency, and Sentinel — each focused on a different analytical lens
- •Unified 12-category taxonomy (Billing & Payments, Delivery & Fulfilment, Product Quality, etc.) applied consistently across every channel
- •Rand-impact quantification for every insight and alert, so you can prioritise by financial return
- •Automatic anomaly detection (running hourly) that flags sentiment spikes, topic surges, CSAT drops, and compliance failures
- •All analysis is transcript-text based — audio recordings are transcribed first and the text is then analysed; no analysis is performed on the audio waveform itself
2.How ClarityCX Works: Insight → Action → Outcome
Every part of ClarityCX serves one of three stages: it surfaces what is actually happening in your conversations (Insight), tells you exactly what to do about it (Action), and shows whether the outcome moved (Outcome). The detailed sections below explain each feature individually — this is the story they all fit into.
See what's actually happening
Every conversation is transcribed, then the transcript text is scored for sentiment, CSAT, quality, and compliance, and categorised — so patterns, anomalies, and cost leakage surface themselves instead of hiding in call recordings.
Know exactly what to do
Each insight arrives with a recommended action and its estimated rand impact, and every root cause is tagged agent- or company-controlled — so you know whether to coach, fix a process, or escalate.
Track the impact
Action tracking shows whether last week's flagged insights were actioned or went OVERDUE, the Agent Performance page tracks sentiment trends across your coaching cycle, and reports quantify the estimated rand impact of what changed.
- •Insight — every conversation is transcribed, scored, and categorised; the Dashboard, Conversation detail views, Early Warning alerts, and Process & Cost Insights surface what is actually happening across all channels
- •Action — each insight comes with a recommended next step and rand impact: coaching priorities on Agent Performance, recommended actions on Process & Cost Insights, and severity-ranked alert responses in Early Warning. Your Coaching analyst on the AI Analysts page can also suggest coaching targets from your data
- •Outcome — the Reports page tracks whether flagged insights moved from 'identified' through 'in progress' to 'resolved' (flagging stale ones OVERDUE), Quality Scorecards show team averages over time, and the Agent Performance page tracks an individual agent's sentiment trend
- •The stages repeat as a weekly loop — this week's report flags what to fix, and next week's action tracking shows whether it moved
3.Dashboard
The Dashboard is your command centre. It aggregates every analysed conversation into a single view with KPI tiles, trend charts, contact-reason breakdowns, active alerts, and top recovery opportunities — updating as new conversations are processed.
- •KPI tiles show total conversations, average CSAT, negative sentiment rate, average quality score (the average quality score across completed conversations in the selected date range, out of 100, with a period-over-period trend), and Recovery Identified at a glance
- •Recovery Identified is a modelled annual impact estimate, shown in compact currency (e.g. R27.9m) with the sublabel 'estimated annualised cost impact'. It represents the estimated annual cost of the identified patterns if they continued for a year — it is not an invoice amount and it is not derived from the configured cost-per-minute setting
- •Sentiment trend chart tracks positive, neutral, and negative sentiment over time
- •Contact reason distribution breaks down conversations by the 12-category taxonomy
- •Active alerts and top recovery opportunities appear in side panels for immediate action
- •If you serve multiple end-clients, the client filter at the top scopes the whole dashboard to one client's conversations
4.Client Segmentation (BPO)
If your contact centre serves multiple end-clients (a BPO model), you can label conversations, agents, data sources, and alerts by client so dashboards and reports can be scoped to one client at a time. Client segmentation is optional — the client tools only appear once you have created at least one client, and every view continues to show ALL of your data until you choose a client filter.
- •Create and manage clients in Settings → Clients — give each end-client a name, an optional colour for chart and filter consistency, and internal notes; inactive clients are hidden from filters but keep their historical data
- •The client selector appears on the Dashboard, Conversations, Early Warning, Reports, Agent Performance, Audio Upload, and Text Ingest pages once at least one client exists — choose a client to scope that view to the client's conversations
- •Reports can be exported client-scoped — apply the client filter before exporting so the PDF or CSV covers only that client's data
- •Views show all data until a client filter is chosen — nothing is hidden by default
- •Data sources can carry a default client, so every conversation ingested from that source is labelled automatically at ingestion
5.Conversations
The Conversations page lists every ingested conversation with its channel, direction, sentiment, status, and key scores. Click any conversation to open the detail view, which shows the full transcript with per-segment sentiment, quality and compliance scores, AI summary, coaching notes, root cause attribution, and topics.
- •Each conversation shows predicted CSAT, handle time, quality score, and compliance score
- •Failure demand and repeat contact flags highlight conversations that indicate process problems
- •Click 'Run Analysis' on any pending conversation to trigger transcription and AI analysis, or let the automatic queue processor handle it
- •The detail view includes a Handle Time Analysis card that compares the conversation's handle time against the channel-specific AHT target, with an AI explanation of the variance and contributing factors tagged as agent-controlled or company-controlled
- •The queue status card shows a live progress bar with pending, processing, complete, and error counts
- •Live analysis updates — open a conversation while it is analysing and a progress indicator shows the current stage (queued, transcribing, or analysing) with the elapsed time and the note 'Analysis in progress — this usually takes under a minute.' The view refreshes itself every 3 seconds and renders the full analysis (quality score, sentiment, root cause, coaching notes) automatically the moment the status becomes Complete — no navigating away and back needed. The list view's status badges also refresh live while it is open
- •Channel filter — narrow the list to Phone, Email, or Chat with the channel dropdown (All by default); it combines freely with the search box and the other filters
- •Transcript view colour-codes each speaker segment by sentiment for quick scanning
- •Duplicate uploads are rejected automatically — audio files are fingerprinted by their file hash and text transcripts by a transcript hash, so the same recording or text isn't analysed twice
- •Personally identifiable information (South African ID numbers, payment cards, bank account numbers, phone numbers, and email addresses) is automatically masked in transcripts at ingestion, before any analysis
6.Quality Scorecards
Quality Scorecards aggregate completed conversations by agent, producing a ranked leaderboard with average CSAT, quality, compliance, and handle time. Use this to identify top performers, spot coaching opportunities, and track team-wide averages over time.
- •Agents are ranked by average quality score, with colour-coded indicators for high and low performers
- •Team-wide KPIs show average CSAT, quality, compliance, and AHT across all agents
- •Each agent row shows their total conversation count and per-metric averages
- •Use scorecards to target coaching sessions and measure improvement
7.Process & Cost Insights
Process & Cost Insights surfaces actionable opportunities to recover revenue and improve efficiency. Insights are categorised as cost process (revenue leakage), automation (self-service potential), or conversation quality (agent improvement areas), each with an estimated annual rand impact.
- •Each insight includes a recommended action and an estimated annual rand impact — these are estimates of the potential recovery if the issue is addressed, not verified realised savings
- •Filter by category to focus on cost recovery, automation, or quality improvements
- •Track insight status from 'identified' through 'in progress' to 'resolved'
- •The aggregate recovery dashboard shows total potential savings across all insights
- •Marking an insight 'resolved' records that the fix was implemented — the estimated impact figure stays an estimate; validating the actual saving requires your own before/after measurement
8.Early Warning
The Early Warning system automatically scans your analysed conversations (every hour) for anomalies — sentiment spikes, topic surges, CSAT drops, SLA breaches, volume anomalies, and compliance failures. Each alert includes severity, estimated rand impact, and a recommended response so you can act before issues escalate.
- •Alerts are categorised by type and severity (critical, warning, info)
- •Acknowledge active alerts to track that they've been seen and are being addressed
- •Filter by status: active, acknowledged, resolved, or all
- •Total rand impact of active alerts is shown in the summary cards
- •Each alert card shows its fired-at timestamp (e.g. 'Fired 12 Oct, 09:14')
- •The active alerts feed intentionally does not follow the Dashboard date range — active alerts remain visible until they are acknowledged or resolved, regardless of the period selected elsewhere
9.Real-Time Assist
Real-Time Assist is an optional paid add-on that sends supervisors live-call alerts while the conversation is still happening. Your telephony platform (Amazon Connect Contact Lens or Genesys Cloud) streams live events to ClarityCX; your alert rules evaluate each event and, when a rule fires, the alert reaches your supervisors in seconds — while the customer is still on the line.
- •Available as an add-on on Growth and Enterprise plans (R7,500/month ex VAT, including 50 monitored agents; additional monitored agents approximately R100/agent/month, ex VAT) — if your subscription doesn't include it, the Settings → Real-Time Assist card shows how to request it; your ClarityCX provider enables the add-on for you
- •Rules are off by default — nothing fires until you create and enable a rule
- •Sentiment floor — fire when live sentiment drops to or below a score you set (e.g. -0.5)
- •Consecutive negative — fire after a set number of consecutive negative events on a single live contact
- •Keyword / category rules — fire when the live contact matches a category your telephony platform tags (e.g. retention-risk)
- •Webhook setup — your telephony admin points Amazon Connect (EventBridge) or Genesys Cloud notifications at the webhook endpoint shown on the Real-Time Assist settings card, signing events with the signing secret provided for your account
- •Alert fatigue controls — each rule has a cooldown (no repeat alert for the same rule + agent within the window, default 10 minutes) and a daily alert cap per rule + agent (default 20)
- •When a rule fires, the alert appears in Early Warning with a [Real-Time] prefix and severity badge, is routed per the rule's routing options (a fixed email address, the affected agent's team leader, and/or the client's ops manager), and appears in the live alert surfaces for eligible users
- •Optional BPO scoping — a rule can be scoped to a single end-client and/or a single team so it only fires for that scope's live contacts
10.Team Leader Live Alerting
When a Real-Time Assist alert fires, supervisors and leaders see it immediately inside the app — no need to watch the Early Warning page or wait for an email. Three surfaces keep live alerts in front of the people who can act while the call is still in progress.
- •Alert bell — the Live Alerts bell in the navigation shows a count of unacknowledged live alerts; open it to see each alert's agent, severity, and message, acknowledge it, or jump straight to the filtered Early Warning view
- •Live toasts — while the app is open, a new alert pops up as a toast notification in the bottom-right corner with View and Acknowledge actions; each alert is shown once per session
- •Critical-alert chime — critical-severity alerts play a short chime; use the speaker icon next to the bell to mute or unmute it (your choice is remembered on this device)
- •Action queue — the 'Real-Time Alerts — needs action' panel at the top of the Early Warning page lists all unacknowledged live alerts, oldest first, so a team leader can work through them while the calls are still live; acknowledging an alert also acknowledges its linked Early Warning record
- •Who sees live alerts — leadership roles (owners, admins, executives) see all live alerts in your organisation; team leaders see alerts for their own team, and ops managers see alerts for their client's contacts. Roles without leadership scope don't see live alerts
11.Reports
The Reports page has two tabs. The Operations Intelligence Report is an auto-generated ops review that turns your conversation analysis into an actionable briefing for your selected date range. The Custom Report Builder lets you build your own report by choosing dimensions, metrics, and filters, then export it as a branded PDF or CSV.
- •Operations Intelligence Report — an auto-generated briefing with seven sections: trend metrics (deltas vs the previous period), topic distribution, root cause breakdown, agent leaderboard, agent spotlight narratives, estimated cost impact, and action tracking
- •Topic distribution — a clickable horizontal bar chart showing the % of conversations by primary topic; click any bar to drill into the individual conversations driving it
- •Root cause breakdown — a pie/donut showing the agent vs company vs mixed attribution split, plus a per-topic stacked view, so you can see whether to coach agents or fix processes; each segment drills into the conversations behind it
- •Agent leaderboard — a table ranked by quality score with colour-coded indicators (green ≥85, amber 70–84, red <70), showing CSAT, compliance, sentiment, top topic, and derived strengths/development areas; click a row to drill into that agent's conversations
- •Agent spotlight — auto-generated narratives for your best performer, most-improved agent (when a previous period exists), and the agent who needs attention, compiled from coaching notes and root cause drivers — not hardcoded
- •Trend metrics — every key metric shows a delta vs the previous reporting period (green = improvement, red = deterioration, grey = flat). If no previous-period data exists, it says so rather than showing fake deltas
- •Estimated cost impact — a transparent, defensible cost model, shown in your tenant currency (rand by default): (failure_demand × avg_handle_cost) + (repeat_contacts × avg_handle_cost × 2) + (low_quality_conversations × churn_risk_value). avg_handle_cost = avg AHT × cost_per_minute ÷ 60. The churn-risk value is adjustable (default R500) so you can model different assumptions. The full formula breakdown is shown so you can defend the number to a client. This is an estimate of the potential cost of these patterns over the period — not a verified realised saving
- •Action tracking — the feedback loop. Last week's briefing flagged insights to fix; this week's report shows whether they moved from 'identified' to 'actioned' or 'resolved'. Insights still sitting at 'identified' are flagged OVERDUE with a red badge. Use the status dropdown on each Process Insight to mark it Identified → Actioned (In Progress) → Resolved
- •Branded PDF export — the Operations Intelligence Report exports as a ClarityCX-branded PDF (logo + 'A product of SONKAY' header, tenant name and date range footer, 'Confidential — prepared by ClarityCX')
- •Custom Report Builder — choose a date range (this week, this month, last 30 days, last quarter, custom), group by up to two dimensions (Topic, Agent, Sentiment, Root Cause, Channel, Direction, Status), select metrics (Conversation Count, Quality Score, CSAT, Compliance Score, AHT, Sentiment Score, Failure Demand %, Repeat Contact %, Estimated Cost Impact), and apply optional filters (by topic, agent, sentiment, root cause, channel, quality range)
- •Custom report output — render on-screen as a dashboard (bars, donut, stacked bars, or tables depending on the grouping), export as a branded PDF, or download raw data as CSV. All charts and tables drill down to the individual conversations behind each group
- •Save & load templates — save a report configuration (dimensions, metrics, filters, date-range type) with a custom name and re-run it with one click from the 'Load template' dropdown. Templates are private to your user account
- •Date range and client filter — apply a date range and, if you use client segmentation, a client filter; the whole report (both tabs and the PDF export) covers only that scope
- •Tenant selector (platform owners) — platform admins and owners can switch between tenants from the header to view any tenant's report; tenant users are automatically scoped to their own tenant
12.Briefings
The Briefings page turns your conversation analysis into an AI-drafted stakeholder briefing you can edit and share. Generate a draft for a date range, review and edit it, then publish it to a share link for operations and executive stakeholders.
- •Generate a draft — click 'Generate New Briefing', choose a date range (defaults to the last 7 days), and the AI drafts a briefing from the completed conversations, top insights, and active alerts in that window
- •A generated briefing contains an executive summary (financial impact, CSAT trends, risks, recommendations), operations action items (coaching targets, process fixes, next steps), and a key-metrics snapshot (conversation count, average CSAT, negative sentiment rate, quality/compliance averages, failure demand and repeat contact counts) captured at generation time
- •Drafts are editable — open a briefing to edit the title, executive summary, and operations actions before sharing
- •Publish & share — publishing turns the draft into a public share link that anyone can open in a browser without logging in. Treat the link as confidential: anyone who has it can read the briefing, so only share it with people who should see the data
- •Delete removes the briefing and its share link — the public link stops working once the briefing is deleted
- •Access — briefings can be created and edited by tenant admins (and platform staff); figures in a briefing are a snapshot at generation time and won't update as new data arrives
13.Survey Import
The Survey Import page lets you attach your customers' actual post-interaction survey scores to their conversations, replacing the AI-predicted CSAT with the real customer-rated score.
- •CSV upload — upload a CSV with the columns external_id, csat_score, verbatim. The external_id must match the conversation's external ID from your telephony platform; scores outside 1–5 are clamped to that range
- •Manual entry — search for a conversation on the page and enter its survey score (1–5) and the customer's verbatim comment individually
- •Matching — survey rows are matched to conversations by external_id (manual entries can also match the conversation's internal id); rows that can't be matched to a conversation are skipped and counted in the import result
- •Once imported, the actual score and verbatim replace the AI prediction for that conversation and are labelled as survey-sourced everywhere CSAT appears (Dashboard, Reports, Quality Scorecards, Conversation Detail)
- •Who can import — admins and analysts (Ingest Only users can attach surveys to the conversations they uploaded themselves)
- •Importing survey data does not re-run analysis — it only attaches the score and verbatim to the existing conversation record
14.AI Analysts
ClarityCX includes four conversational AI analysts you can chat with directly (sidebar label 'AI Analysts', page heading 'AI Agents'). Each analyst specialises in a different analytical domain and can query your conversation data to answer plain-English questions, generate insights, and recommend actions — effectively giving you an on-demand data analyst for each area of your operation.
- •Coaching Agent — answers plain-English questions about agent performance, CSAT, quality scores, compliance, and AHT. Use it to identify which agents need coaching, what specific behaviours are driving low scores, and which quality criteria are underperforming (available on all plans)
- •Opportunity Agent — answers questions about root causes, topic trends, and estimated rand impact across all channels. Instead of manually filtering dashboards, you can ask 'What are the top 3 drivers of negative sentiment this week and what is their financial impact?' and get a specific, structured answer (available on all plans)
- •Efficiency Agent — identifies repetitive tasks and conversation types suitable for self-service or automation, helping you reduce failure demand and deflect avoidable contacts (Growth and Enterprise plans)
- •Sentinel Agent — discusses emerging risks, active anomalies, and proactive responses. Ask it to summarise what needs urgent attention and what the recommended response is for each alert (Growth and Enterprise plans)
- •On Starter, the Efficiency and Sentinel agents appear locked with an upgrade prompt — upgrading to Growth unlocks them
- •Speed to insight — agents bypass dashboard filtering and return summarised, strategic answers in seconds, shifting your role from hunting for data to interpreting actionable outcomes
- •Executive-ready output — all agents respond in professional markdown with clear headers, bolded metrics, and specific numbers, making it easy to copy findings directly into briefings, stakeholder reports, or operational action plans
- •Cross-channel visibility — agents access data across all channels (phone, chat, email, bot, social), giving you a unified view without needing to check each channel separately
- •Credit usage — chatting with the built-in analysts does not consume your ClarityCX analysis credits (1 credit = 1 conversation analysis); custom agents are the only AI chats that bill credits (see Custom Agents)
15.Custom Agents
Enterprise customers can build their own AI agents tailored to their business from the Custom Agents page. You define the agent's persona and focus area by writing instructions — the agent automatically inherits access to your full conversation analysis data (sentiment, topics, process insights, and rand impact), so no tool configuration is needed. Custom agents are a configurable extension of the built-in analyst suite, created from the front end rather than in code.
- •Enterprise-only — custom agents are available exclusively on the Enterprise plan; Starter and Growth customers see an upgrade prompt
- •Build from the front end — give your agent a name, a short description, and detailed instructions describing its persona, focus area, and how it should answer (no code or tool configuration required)
- •Automatic data access — every custom agent inherits your full CX dataset (sentiment breakdown, top topics, channel mix, process insights, and rand impact) so it can answer questions grounded in your real data
- •Credit cost — creating a custom agent costs 100 ClarityCX analysis credits, and each query to that agent costs 2 credits; both costs are shown in the builder and the chat panel before you act
- •Your credit balance is shown in the sidebar of the Custom Agents page and updates after each creation or query
- •Manage your agents — select an agent to chat with it, or delete it from the list; the query count on each card shows how many times an agent has been used
- •Use cases — build an escalation risk predictor, a churn-cost analyst, a compliance-gap finder, or any specialist tailored to your operations; the agent answers in clear markdown with specific numbers and rand amounts
16.Text Ingest
Text Ingest lets you manually add chat, email, bot, or social conversations for analysis without needing an audio file. Paste the conversation text, specify the channel and direction, and the system will analyse it using the same AI pipeline as audio conversations.
- •Supports chat, email, bot, and social channels — no audio required
- •Each ingested conversation goes through the full AI analysis pipeline
- •Uploads are batched — you can ingest up to 100 conversations per submission, with per-item feedback showing which succeeded and which failed
- •Transcripts are screened for personally identifiable information (ID numbers, payment cards, bank account numbers, phone numbers, emails) and masked before they are stored or analysed
- •Duplicate text transcripts are detected and skipped within your organisation
- •If your credit balance is at zero, ingestion is paused until you top up — each ingested conversation consumes 1 analysis credit when it is analysed
- •Ingested conversations appear in the Conversations list alongside audio ones
17.Integrations
The Integrations page is where data sources for ClarityCX are set up with your provider. Supported source types include Genesys, Avaya, Cisco, Twilio, Amazon Connect, RingCentral, Talkdesk, Salesforce, Zendesk, and HubSpot, plus manual upload for one-off imports.
- •Each data source shows connection status, sync frequency, and conversations ingested
- •Configure sync frequency as realtime, hourly, or daily depending on your needs
- •Error messages surface connection issues so you can troubleshoot quickly
- •A platform listed here is a supported source type — being selectable in the dropdown does not mean data is flowing; a source only ingests conversations once it is connected and its status shows 'connected'. Working integrations are set up with your ClarityCX provider
- •Manual upload (Audio Upload and Text Ingest) is always available as a fallback for ad-hoc imports
18.Audio Upload
Audio Upload lets you upload call recordings for automatic transcription and AI analysis. Upload an audio file, and the system transcribes it, runs the full analysis pipeline, and makes the conversation available in the Conversations list with all scores and insights.
- •Supports .mp3, .wav, .ogg, and .m4a files up to 25 MB per file, with batch upload supported
- •Per-file feedback — every file in a session shows its own status, and failed files list the reason (unsupported format, too large, duplicate, or upload failure) so nothing fails silently; failed files can be fixed and retried
- •Duplicate recordings are rejected automatically — each audio file is fingerprinted by a hash of its bytes, so the same recording isn't analysed twice
- •Transcripts are screened for personally identifiable information and masked before they are stored or analysed
- •Each upload creates a conversation record and triggers the analysis pipeline
- •Uploaded conversations are analysed automatically by the queue processor — no manual action needed
- •Track processing status from 'pending' through 'transcribing' and 'analyzing' to 'complete'
- •Once complete, the conversation appears in the Conversations list with full insights
19.Automated Analysis & Queue Processing
ClarityCX automatically processes pending conversations through a batch queue system, so you don't need to manually trigger analysis on every conversation. Whether conversations arrive via audio upload, text ingest, or telephony integration, they are picked up and analysed automatically with controlled concurrency to ensure stability under high load.
- •The queue processor runs every 5 minutes. By default it picks up 3 pending conversations per run and processes 2 at a time, but both the batch size (1–6) and concurrency (1–5) are configurable per tenant in Settings under Processing Throughput
- •New conversations from uploads, text ingest, or integrations are analysed automatically — no manual action required
- •Failed analyses are automatically retried up to 2 times before being left in 'error' status for manual review
- •The queue status card on the Conversations page shows a live progress bar with pending, processing, complete, and error counts, and auto-refreshes every 10 seconds while the queue is active
- •Manual 'Run Analysis' is still available on individual conversations for immediate processing when needed
- •Default throughput is approximately 864 conversations per day (3 per run × 12 runs/hour × 24 hours). At the maximum batch size of 6, throughput reaches up to 1,728 conversations per day — capped to stay under 2,000/day and avoid API rate limits
20.Performance Targets & Throughput
Admins can set benchmark targets for key contact centre metrics and configure processing throughput directly in Settings. These targets define your performance goals and serve as the baseline for comparing actual results across the platform, while throughput settings control how fast conversations are processed.
- •CSAT target — desired customer satisfaction score on a 1-5 scale (default 4.5)
- •SLA adherence — target percentage of conversations meeting service level (default 80%)
- •Quality score — target agent quality score out of 100 (default 85)
- •Compliance score — target compliance score out of 100 (default 95)
- •Average handle time — per-channel AHT targets (in seconds) set separately for phone (default 300), chat (600), email (1,800), bot (60), and social (900); the general AHT target serves as a fallback for any channel not explicitly customised
- •Negative sentiment — maximum acceptable negative sentiment rate (default 10%)
- •Batch size — number of conversations picked up per queue run, configurable 1–6 (default 3)
- •Concurrency — number of analyses processed simultaneously within a batch, configurable 1–5 (default 2)
- •Daily throughput is capped at 1,728 conversations (6 × 288 runs/day) to stay under 2,000 and avoid API rate limits
- •Targets and throughput settings are saved per tenant and can be updated at any time from the Settings page
21.Billing & Credits
ClarityCX operates on a prepaid credit model: 1 credit = 1 conversation analysis. Each plan includes a monthly credit allocation, and when your balance runs out, analyses pause until you top up — no overage charges. All prices are ex VAT; checkout charges the VAT-inclusive amount.
- •Three plans: Starter (R6,500/month ex VAT — 9,000 credits / 9,000 conversations), Growth (R14,000/month ex VAT — 36,000 credits / 36,000 conversations), Enterprise (R42,000/month ex VAT — 135,000 credits / 135,000 conversations)
- •Seats count team-member logins, not agents being analysed: Starter includes 3 logins, Growth 10, Enterprise 25. Additional logins beyond the plan are invoiced manually at R150/user/month ex VAT
- •One-time setup/onboarding fees (ex VAT): Starter R0 — no setup fee, self-serve onboarding; Growth R7,500 — waived entirely on a 6-month commitment; Enterprise R25,000–R35,000 — 50% off on a 12-month commitment
- •Prepaid only — 1 credit per conversation of up to 10 minutes, then 1 credit per started 10-minute block; analyses pause at zero, no overage charges
- •Plan credits do not roll over — a fresh allocation is added and your cycle usage resets when a plan payment is confirmed for the new cycle
- •Plan changes and top-ups are processed via PayFast checkout; credits are added and plans activated once the payment is confirmed — not before
- •Credit top-up packs — 5,000 credits (R3,750), 10,000 credits (R7,000), 40,000 credits (R26,000), 100,000 credits (R55,000), all ex VAT. Top-up credits are valid for 60 days from the date of purchase — unused top-up credits lapse after that, and the Credit History shows an expiry entry explaining the balance drop. They are priced above the subscription per-credit rate to incentivise subscribing
- •Credit balance, credits used this cycle, remaining credits, top-up expiry dates, and the full transaction history are visible on the Billing page
- •Automated email warnings are sent to your primary contact at 80% and 95% credit usage, and when credits are exhausted
22.How Credits Are Used
Your ClarityCX subscription and top-ups buy analysis credits — the prepaid balance shown on your Billing page. The AI compute behind each analysis runs on the platform side and is included in your plan; it is not something you monitor or purchase separately. Here is exactly what consumes your analysis credits.
- •Conversation analysis — each analysed conversation consumes 1 analysis credit at processing time (audio transcription and the AI analysis itself are platform-side compute included in your plan)
- •Custom agents (Enterprise) — creating a custom agent consumes 100 analysis credits, and each query to a custom agent consumes 2 analysis credits; costs are shown before you act
- •Ingestion is credit-checked — audio uploads, text ingest, and API ingestion are paused when your balance reaches zero (402 'Insufficient credits'), so a zero balance stops new analyses rather than creating a bill
- •What does NOT consume analysis credits — chatting with the built-in AI analysts (Coaching, Opportunity, Efficiency, Sentinel), report generation, digest generation, and alert processing run on platform-side compute included in your plan
- •Top up via the Billing page if your balance reaches zero, or wait for your next plan allocation once the new cycle's payment is confirmed
23.How CSAT Is Calculated
ClarityCX measures customer satisfaction (CSAT) on a 1-5 scale. There are two sources of CSAT data — AI-predicted and actual survey — and the platform uses the best available source for each conversation. Understanding how CSAT is obtained and scored helps you interpret the metrics correctly.
- •Predicted CSAT — when no survey data is available, the AI analyses the conversation transcript and predicts a CSAT score (1-5) based on sentiment, agent performance, issue resolution, and customer language cues. This is an estimate, not a customer-rated score
- •Actual CSAT (Survey) — when you provide post-interaction survey data via the Survey Import page, the real customer-rated score replaces the AI prediction for that conversation. The customer's verbatim comment is also stored and displayed alongside the score
- •Source priority — if a conversation has an imported survey score, the actual CSAT is used everywhere CSAT appears (Dashboard, Reports, Quality Scorecards, Conversation Detail). Conversations without survey data continue to use the AI-predicted score
- •CSAT methodology — in Settings you can define which survey ratings count as 'meeting' your CSAT target: Top-Box (only 5-star ratings count), Top-2-Box (4-star and 5-star ratings count), or Average Score (the mean of all survey scores is compared to your target)
- •CSAT target adherence — the percentage of conversations whose survey score meets your chosen methodology threshold is compared against your CSAT target adherence percentage (e.g., 80% of conversations scoring 4 or 5)
- •Importing survey data — use the Survey Import page to upload a CSV (columns: external_id, csat_score, verbatim) or enter individual survey responses manually. Survey data is matched to conversations by their external ID from your telephony platform
24.How Sentiment Is Rated
Sentiment is an independent AI judgment read from the words of the transcript, not a formula derived from CSAT or QA. All ClarityCX analysis is text-based: audio is transcribed first and only the transcript text is analysed — nothing is detected from the audio itself. Understanding what sentiment measures — and what it deliberately does not measure — is essential for interpreting it correctly next to your other scores.
- •Text-based only — the AI reads the transcript wording (the customer's language, phrasing, and word choices) and judges sentiment from the evidence in the text. It never analyses the audio itself — only the words on the page
- •What sentiment measures — how the customer's language reads: frustrated, happy, or calm and measured. It reflects how the customer comes across in the words used, not whether their issue was resolved
- •How it's scored — the AI reads the entire transcript and produces two values: an overall label (positive, neutral, or negative) and a numeric sentiment score from -1 (very negative) to +1 (very positive), where 0 is neutral. Each transcript segment also gets its own sentiment label and score, so you can see where the language shifts mid-conversation
- •Per-segment colour-coding — the transcript view colours each speaker segment green (positive), grey (neutral), or red (negative), letting you scan a long call and spot the moments where the wording turned
- •Independent of CSAT and QA — sentiment is not calculated from CSAT or quality score. All three are produced separately by the AI and measure different things: sentiment = how the customer's language reads, CSAT = how satisfied the AI estimates the customer is with the outcome, quality = how well the agent followed the rubric
- •Neutral sentiment with low CSAT and low QA is valid and common — a customer can use calm, measured wording (neutral sentiment) while the agent fails to resolve the issue (low CSAT) and misses rubric steps (low QA). The customer came across as calm even though the outcome was poor
- •Why the separation matters — sentiment alone would never reveal a calm-but-unresolved call. Splitting the three scores is exactly how ClarityCX surfaces failure-demand and repeat-contact patterns that hide behind calm, neutral wording
- •It's an AI judgment — sentiment is an estimate inferred from the transcript, not a measurement; treat it as directional evidence and check the transcript itself when a score matters
- •No hard-coded thresholds — the AI's sentiment judgment is stored directly; there is no fixed score-to-label formula in the platform. The negative sentiment target in Settings (default 10%) sets your acceptable rate of negative-sentiment conversations, not how the label itself is assigned
25.First Contact Resolution (FCR)
ClarityCX predicts First Contact Resolution (FCR) for every analysed conversation. FCR measures whether a customer's issue appears to have been fully resolved on their first contact, without the need for a follow-up interaction. This is one of the most important operational metrics in a contact centre — high FCR rates correlate with higher CSAT, lower operating costs, and reduced repeat contact volume.
- •AI-predicted estimate — the AI analyses the conversation transcript to determine whether the customer's issue appears to have been resolved during this interaction, considering resolution cues, agent confidence, and the absence of escalation or follow-up indicators. It is an inference from the transcript, not a verified outcome — confirm with your own records when a specific call matters
- •Confidence score — each FCR prediction includes a confidence score (0-100) indicating how certain the AI is about its assessment. Scores above 80% are high-confidence predictions
- •FCR is closely related to the 'repeat contact' flag — a conversation flagged as a repeat contact is unlikely to be an FCR, since the customer has been in touch before about the same issue
- •FCR appears on the Dashboard as a percentage rate, on individual conversation detail pages as a resolved/unresolved indicator, and in Reports alongside other KPIs
- •Use FCR to identify agents or topics with low resolution rates — these are prime targets for coaching, process improvement, or knowledge base updates
26.Root Cause Attribution
Every analysed conversation includes a root cause attribution that classifies the primary driver of the conversation's outcome as agent-controlled, company-controlled, or mixed. This is the AI's evidence-based judgment from the transcript — use it to distinguish between coaching opportunities (things the agent can improve) and systemic issues (things the company needs to fix through policy, process, or product changes).
- •Agent — the outcome was primarily driven by factors within the agent's control (wording and courtesy, empathy, product knowledge, procedure adherence, effort, hold/transfer handling). These are coaching opportunities
- •Company — the outcome was primarily driven by factors outside the agent's control (policy restrictions, product defects, system outages, billing errors, process gaps, staffing, long queue times). These require policy, process, or product changes
- •Mixed — both agent and company factors contributed roughly equally to the outcome
- •Each conversation also lists specific root cause drivers — the individual factors that contributed to the outcome, each tagged as agent-controlled or company-controlled with a brief explanation grounded in the transcript
- •The attribution is an AI inference from the transcript evidence, not a verified fact — review the transcript before acting on a contested call
- •Use the Agent vs Company breakdown on the Dashboard to see what proportion of your negative outcomes are coaching issues versus systemic issues — this helps you decide where to invest improvement effort
- •The Root Cause card on the conversation detail page shows the attribution and all drivers, making it easy to understand why a conversation went the way it did
27.AHT Variance Analysis
Every analysed conversation includes an Average Handle Time (AHT) variance analysis. The platform compares the conversation's actual handle time against the target AHT for its specific channel and uses AI to explain why the handle time was longer or shorter than target, categorising the contributing factors as agent-controlled or company-controlled.
- •Per-channel targets — AHT targets are set per channel (phone, chat, email, bot, social) in Settings, so a 5-minute phone target and a 30-minute email target can coexist without conflict
- •The variance is calculated as actual handle time minus the channel-specific target — positive means over target, negative means under
- •The AI explains WHY the handle time deviated from target, considering factors like unnecessary holds, lack of product knowledge, complex policies, system issues, and customer preparation
- •Each contributing driver is tagged as 'agent' (within the agent's control — coaching opportunity) or 'company' (outside the agent's control — policy/process issue)
- •The Handle Time Analysis card on the conversation detail page shows the actual handle time, variance, target, explanation, and drivers
- •If no duration is available from the source, the platform estimates handle time from the transcript length or segment timestamps
- •Use AHT variance analysis to identify whether handle time issues are coaching opportunities (agent-controlled) or systemic issues (company-controlled)
28.How Costs Are Calculated
ClarityCX estimates the operational cost of each conversation and quantifies the annualised financial impact of process issues and recovery opportunities. Understanding these calculations helps you prioritise actions by financial return.
- •Estimated conversation cost — each conversation's cost is calculated as (duration in minutes × cost per minute). The cost per minute is configurable in Settings under CSAT & Cost Configuration (default: R2.50/minute). This provides an approximation of the operational cost of handling the interaction
- •Cost is calculated automatically — the estimated cost is set during AI analysis based on the conversation's duration (or average handle time if available) and your configured cost-per-minute rate
- •Rand impact on insights — each process insight carries an estimated annual rand impact, modelled per issue type (for example avoidable contacts, repeat contacts, or low-quality patterns). These are modelled estimates of the potential recovery if the issue is addressed — not derived from your configured cost-per-minute setting, and not realised savings
- •Rand impact on alerts — early warning alerts quantify the estimated financial risk of detected anomalies (sentiment spikes, compliance failures, etc.) using the same cost-per-conversation basis
- •Recovery identified — the 'Recovery Identified' figure on the Dashboard (and the total on Process & Cost Insights) is the sum of rand impacts across the process insights in the selected period. It is a modelled annual impact estimate, not an invoice amount, and it is not derived from the cost-per-minute setting
- •Cost per minute is adjustable — update your cost-per-minute rate in Settings to reflect your actual staffing, infrastructure, and overhead costs per agent minute. All cost calculations update automatically on subsequent analyses
29.Recovery Identified: What the Figures Mean
The 'Recovery Identified' tile on the Dashboard (and the total at the top of Process & Cost Insights) represents the estimated annual cost of the inefficient patterns the analysis has detected in your conversations — the money that leaks out of the operation each year if these patterns are left unaddressed. It is the headline number for prioritising: which process fixes, automation projects, or coaching programmes will return the most. This narrative explains exactly where the number comes from, so you can explain and defend it to your own stakeholders.
- •Where the figure comes from — Recovery Identified is the sum of the 'estimated annual rand impact' on every process insight in the period you have selected. Each insight is one detected pattern (for example 'Failure demand — Billing & Payments'), and each carries its own annualised impact figure. The tile is simply their total
- •How each insight's impact is built up — a three-step calculation. Step 1: detect the pattern — the analysis counts conversations matching a known cost pattern (avoidable failure-demand contacts, repeat contacts on the same issue, bot conversations that failed and escalated to a human, simple interactions suitable for self-service, conversations scoring below 60 on quality, and negative-sentiment clusters on one topic). Step 2: annualise — the pattern's share of your analysed conversations is projected out to a full year at your current contact volume. Step 3: value it — the annualised count is multiplied by a per-conversation cost figure for that pattern type
- •The per-pattern cost model — each conversation in a detected pattern is valued at a standard modelling figure: R3,200 per avoidable failure-demand contact (a full handling cost wasted), R2,800 per repeat contact (a near-doubled handling cost), R1,200 per conversation deflectable to self-service, R1,500 per bot conversation that escalated to a human, and R4,500 per low-quality or negative-sentiment conversation (reflecting churn risk and rework). Worked example: if 30 of 2,000 analysed conversations are failure demand about billing, that is a 1.5% rate; annualised at your volume and multiplied by R3,200, a 100,000-conversation year models out at R4.8m of avoidable cost
- •Why these figures and not your cost-per-minute — the pattern costs are standard industry-style modelling assumptions built into the platform, deliberately not derived from your cost-per-minute setting. Your cost-per-minute setting (used for per-conversation operating cost) measures what handling already costs; Recovery Identified measures the avoidable cost of patterns on top of that — wasted handling, churn risk, and deflection opportunity. Keeping the two separate prevents double-counting and keeps the recovery figure comparable across clients
- •What it is NOT — Recovery Identified is a modelled estimate, not an invoice, not a realised saving, and not money that has been recovered. It answers 'what is this pattern costing us annually if nothing changes?'. It stays an estimate even after you mark an insight 'resolved' — validating the actual saving requires your own before-and-after measurement
- •How to use it well — treat it as a prioritisation and conversation tool: rank insights by impact to decide where to invest first, use the figure to size the business case for a fix, and track the tile over time — as patterns are fixed and stop appearing in new conversations, the identified recovery should fall. Because it is annualised from your current volume, it grows and shrinks with contact volume too, so compare periods at similar volumes
- •Currency and display — the figure is shown in your tenant currency (rand by default) in compact form on the Dashboard tile (e.g. R27.9m), with exact amounts on the Process & Cost Insights page; both cover the same date range, so the two always reconcile
30.Data Retention (Process-and-Purge)
ClarityCX supports data retention modes that control what happens to the raw material behind each conversation — the audio link, the full transcript, and the per-segment transcript breakdown. Full Retention (the default) keeps all raw data indefinitely so you can replay and re-read any conversation. Process-and-Purge is designed to remove the raw data from the conversation record after a configurable age while keeping every extracted insight — so reporting, dashboards, and briefings stay complete even after the source material is gone. Process-and-Purge is not currently operational. It is available on request and will be configured and activated by your ClarityCX provider for a live client after the retention requirements have been agreed. Selecting a retention setting does not activate automated deletion.
- •Full Retention (default) — raw audio link, transcript, and transcript_segments are kept indefinitely; conversations can be replayed and re-read at any time
- •Process-and-Purge is not currently operational — it is available on request and will be configured and activated by your ClarityCX provider for a live client after the retention requirements have been agreed; selecting a retention setting does not activate automated deletion
- •Intended design once activated — the raw-data fields (audio_url, transcript, transcript_segments) would be cleared on the conversation record once a conversation is older than the configured purge age (in hours), with a purged_at timestamp stamped for audit; the conversation record itself stays
- •Derived insights would be preserved once activated — sentiment, sentiment score, topics, predicted/actual CSAT, AHT and its variance/drivers, compliance score, quality score, FCR, root cause attribution and drivers, estimated cost, and the analysis summary would remain on the conversation record and continue to appear in every report, dashboard, trend chart, and briefing
- •Purge age is a configurable setting in hours — 0 would mean raw data is removed once analysis is complete, or any positive number of hours to retain it for that window; setting this value alone does not start any deletion
- •Retention access is provider-controlled — the Data Retention settings card is hidden by default and only appears once your ClarityCX provider enables retention access for your organisation; contact your provider to request it. Purge scheduling for specific scopes (e.g. after a managed report delivery) is also arranged with your provider
- •Once activated, every purge would be recorded in an audit log with the scope, counts, and timestamp, and a purge certificate could be issued for your compliance records — certificate handling will be agreed with your provider as part of activation; no certificates are being issued today because Process-and-Purge is not yet operational
31.Improvement Area Taxonomy
Every analysed conversation includes an improvement area classification. When the agent's performance was the primary driver of a negative or mixed outcome, the AI tags the single most impactful improvement area — helping you focus coaching on the right skill rather than a vague 'do better'. When the call was positive, neutral, or the root cause was company-side, the area is set to n/a.
- •soft_skills — empathy, wording and courtesy, de-escalation, active listening, and rapport. Example: the transcript shows the agent dismissing the customer's stated frustration with curt wording, escalating the exchange instead of calming it
- •product_knowledge — gave wrong information, didn't know the answer, or couldn't explain a product or policy. Example: the agent told the customer the wrong returns window, causing a follow-up call to correct it
- •process_knowledge — didn't follow the correct workflow, missed an escalation step, or applied the wrong process. Example: the agent should have escalated to a supervisor after two failed resolution attempts but continued retrying instead
- •communication — unclear explanations, talking over the customer, using jargon, or poor conversational structure. Example: the agent gave a technically correct answer but in jargon the customer didn't understand, leading to confusion and a repeat contact
- •system_navigation — struggled with tools or systems, slow system handling, or couldn't find the information needed. Example: the agent spent several minutes clicking through multiple screens to find the customer's order status, stretching the handle time
- •n/a — used when the call was positive or neutral, OR when the root cause was company-side (policy restriction, product defect, system outage) rather than the agent's performance. Company-side root causes are not coaching opportunities and are never tagged with an improvement area
- •Only one area per conversation — the AI picks the single most impactful factor, not a list, so coaching priorities stay focused
- •Each non-n/a classification includes a one-sentence improvement_detail explaining exactly what the agent could improve, grounded in the conversation transcript
32.How Agent Performance Works
The Agent Performance page turns the per-conversation analysis into a single-agent trend and coaching view. Understanding the data flow helps you trust the charts and know exactly what each visualisation is built from.
- •Data source — every chart on the Agent Performance page is built from existing Conversation records filtered by the selected agent, the signed-in user's tenant, the chosen date range, and a status of 'complete' (or conversations that have an analysis summary). No data is fabricated or estimated beyond what the analysis pipeline already produced
- •Sentiment trend — the daily line chart averages the sentiment_score of every completed conversation for that agent on each day in the range. Green segments are days where the daily average was above zero; red segments are days below zero. A shaded neutral band marks the -0.1 to +0.1 zone where sentiment is effectively flat
- •Call distribution donut — counts each completed conversation by its overall_sentiment label (positive, neutral, negative) and shows the total in the centre. This is a raw count, not a weighted score
- •Category heatmap — groups completed conversations by primary_topic and computes total calls, average sentiment score, average quality score, and the positive and negative percentage for each topic. The average sentiment cell is colour-coded green (>0.3), yellow (-0.1 to 0.3), or red (< -0.1) for instant scanning
- •Improvement area breakdown — counts only conversations where improvement_area is not n/a (i.e. the agent's performance was the primary driver of a negative or mixed outcome), grouped by the five improvement areas. The percentage is relative to the total number of negative-sentiment conversations for that agent in the range
- •Coaching priorities — the auto-generated card cross-references the primary_topic with the improvement_area for every agent-driven negative/mixed call, counts each combination, and ranks the top three by count. Each priority shows the improvement area, the topic, the number of negative calls, and the average sentiment across those calls
- •Plan gating — Starter shows the sentiment trend and call distribution donut, with any date window up to 7 days; Growth adds the category heatmap, improvement area breakdown, coaching priorities, and the daily volume bar (with unlimited date range); Enterprise adds PDF and CSV export of the report, side-by-side agent comparison, and group filtering by team, tenure, or BPO partner
- •Real-time — the data is fetched fresh each time you change the agent, date range, or group filter, so the charts always reflect the latest analysed conversations
33.Coaching Plans
The Coaching module turns coaching flags surfaced by conversation analysis into structured coaching plans, tracked from issue through signatures to a measured outcome. Plans can be generated from the Coaching page or from an agent profile.
- •Generate a plan from the Coaching page — select any agent from the agent dropdown, choose the coaching framework (GROW or SMART) and the trigger flags that apply, and generate the plan
- •Plans can also be created from the existing agent-profile entry point — both routes produce the same plan type
- •Plans are editable until they are signed; once both parties have signed, the plan is locked and its signed document is tamper-evident
- •Progress is measured only after the coaching period has elapsed and enough post-coaching conversations exist, so the outcome is not measured prematurely