What your customers actually said.
Surveys reach the 3% who answer them. QBRs quote the two calls an executive happened to hear. Nivākya reads every conversation your agents and your people have, groups it into themes you can act on, and ties each theme back to the revenue sitting behind it.
Forty-six themes, ranked by how much of your phone life they consume.
Themes are emergent — they are discovered, not configured — and each one carries a share of call volume, a trend against the previous quarter, and a randomised verbatim you can play. Six of the top themes from a national retail corpus this quarter:
Delivery dates slipping
Callers quote a promised date, then a revised one, then ask for a reason. 61% of the calls are resolved without a human; the rest become a policy argument about compensation.
Trend vs last quarter: +4.1 pts · Top region: Maharashtra · First seen 11 Jul
Unexplained charges
Small-amount disputes on statements, usually a convenience fee the caller was never told about at checkout. High emotional intensity despite low value — the cheapest theme to eliminate.
Trend: +0.8 pts · Recommended owner: checkout product team
Product fit questions
“Will this fit my kitchen / my fleet / my plot?” Genuine buying intent that gets answered on the phone instead of on the page. Every one of these calls is a content gap with a price tag.
Trend: −1.2 pts — the new spec table is working
Account and login lockouts
Password resets, OTPs that never arrive and numbers attached to an old handset. Predictable, high-volume and entirely automatable — the theme with the best cost per resolved call.
Resolution without a human: 89%
Rival quotes and comparisons
Named competitors, named prices, often read aloud from a competitor's own document. The highest-value theme per call in the corpus and the one most often mishandled by an unprepared agent.
Win rate when handled with the comparison card: 64% vs 38% without
Technician and visit quality
Lateness, no-shows and repeat visits for the same fault. The theme with the strongest correlation to churn in the next 90 days — a 2.1× lift for callers who mention it twice.
Trend: +2.3 pts · Escalated to operations weekly
Who is beating you, and on what sentence.
Every competitor name and price objection is extracted with the surrounding turns, then framed as wins and losses. Sales leadership gets a league table instead of a feeling.
| Competitor mentioned | Mentions | Where they win | Where you win | Head-to-head outcome |
|---|---|---|---|---|
| Northwind | 1,142 | Perceived durability; longer warranty stated in their literature | Delivery inside 6 days; local service network in 84 cities | Won 58% · Lost 31% · Unclear 11% |
| Bluepine | 876 | Price — 7 to 12% lower on the entry SKU | Bundle value and instalment plans; no deposit on the mid SKU | Won 64% · Lost 27% · Unclear 9% |
| Meridian | 603 | Brand familiarity with older buyers; dealer-only support | Self-serve support and a 90-day return window | Won 44% · Lost 46% · Unclear 10% |
| Vantage | 418 | Aggressive outbound sales; quoted discounts on the call | Transparent pricing and no surprise fees at checkout | Won 61% · Lost 33% · Unclear 6% |
| Orbit | 291 | Faster installation booking during peak weeks | Post-sale service quality; 2.1× fewer repeat visits | Won 47% · Lost 41% · Unclear 12% |
Price, timing, trust, fit, authority and inertia — 214 objection phrases mapped to six roots, so coaching targets a root cause, not a sentence.
Every number in the table opens thirty seconds of audio around the moment the claim was made. Nobody has to take a dashboard's word for it.
Lost-call transcripts are summarised weekly with the turn where the conversation turned, and circulated to the sales enablement queue automatically.
A transcript has a pipeline number, and it can be defended.
Attribution here is deliberately conservative: we only claim revenue where a conversation produced a durable identifier — a deal id, an order number, a signed quote — and the chain from call to cash can be replayed months later.
- Write-time linkage
- The wrap-up node writes the CRM record with the call id, so the deal and the conversation are joined at the moment of the call, never stitched afterwards.
- Chain of custody
- Deal id → call id → transcript hash → tool calls → order id. Each hop is stored; a finance analyst can follow it in the console without asking an engineer.
- Deal-stage influence
- For long cycles, we report influence rather than creation: which conversations touched the deal, in what order, and what the agenda was on each one.
- Channel honesty
- A call that merely confirmed an order the customer had already placed online is tagged confirm-only and excluded from created-pipeline totals by default.
- Lag reporting
- Attribution matures over 30, 60 and 90 days. The report shows maturity alongside value, so nobody quotes a ninety-day number on day nine.
For the humans — reviewed at four times the sample.
A QA lead with a headset can review maybe twelve calls a week. The same rubric applied to every call means a person gets a scorecard built from a thousand conversations in the period, with the worst twenty queued for coaching and the best twenty queued for the playbook.
Ravi Nair · 412 calls
Strength: keeps the caller talking. Coaching queue: two calls where a discount was implied before the check.
Meera Kulkarni · 388 calls
Top decile for de-escalation. Two above-average holds length, flagged for the process team, not the person.
Team median · 6 agents · 2,104 calls
Night shift beats day shift on retention calls by 4 points; day shift wins on upsell by 11. The roster conversation just got data.
Automated scoring reduces the manual review burden by roughly 70% while increasing coverage from 12 calls per reviewer per week to the full population. Reviewers spend their time on the twenty calls that matter.
Monthly calibration sessions compare human reviewer scores against the model on a shared sample. Disagreements above the threshold are folded back into the rubric instead of being argued in a meeting.
Scores are checked quarterly for systematic bias by language, region and shift, and the audit output is retained. Scorecards are performance information, not surveillance: agents see their own data first.
The sentence people say eleven weeks before they leave.
Churn is rarely a surprise in the transcript; it is a surprise in the dashboard. Nivākya scores every conversation for signals that precede cancellation, groups them by account, and hands the list to the team that can still do something about it.
Signals are weighted per industry. In subscription retail, effort complaints outrank price talk; in lending, a single mention of a competing rate is the stronger predictor. The weights are yours to tune and are versioned like everything else.
Your analysts work where they already work.
Structured conversation data lands in your warehouse on a schedule you control, with a stable schema, a documented grain and the same hashes the audit spine holds. We do not ask you to learn our BI tool to trust our numbers.
Native Snowflake share
Six tables — conversations, turns, tools, themes, scores, risk — refreshed every fifteen minutes into a share you subscribe to, or a direct load into your own database. No extract files to babysit.
Typical volume: 1.8M conversations ≈ 74 GB compressed per quarter
BigQuery dataset
Partitioned by day, clustered by workspace and journey, with the transcript stored as a nested array so a single query can join a theme to the exact sentence that produced it.
Supports streaming inserts for teams that build on the last hour
Object storage drops
Parquet and JSONL exports to your own bucket with customer-managed keys, plus the audio archive if you want it. Lifecycle rules are yours; we never hold a second copy longer than your retention policy allows.
Includes recording, redacted transcript and the redaction receipt
Events as they happen
Signed webhooks for call-ended, theme-assigned, risk-flagged and score-published events, with at-least-once delivery, replay from any offset in the last 30 days, and a dead-letter queue you can inspect.
HMAC-SHA256 signatures · replay API · 3 retry ladder
| Table | Grain | What it answers |
|---|---|---|
| conversations | One row per call | Outcome, duration, agent version, disposition, deal id |
| turns | One row per utterance | Speaker, text, redaction state, latency, policy decision |
| tools | One row per tool call | Tool name, latency, retries, terminal status, idempotency key |
| themes | One row per call-theme pair | Assigned theme, confidence, verbatim offset |
| scores | One row per rubric per call | Rubric, value, human or model reviewer, calibration batch |
| risk | One row per account per week | Signal weights, churn score, save-queue membership |
- ▪ Redaction runs before persistence — card numbers, CVVs and PINs never reach storage in the clear.
- ▪ PII fields carry a classification tag, so a warehouse query can exclude them without a legal review each time.
- ▪ Retention is per data class, per region, and enforced by the platform rather than by policy documents.
- ▪ Every export is written to the audit spine too: who pulled what, when, and which tables were touched.
- ▪ Deletion requests propagate to your warehouse on the next sync through a tombstone table you can join against.
Send us a month of calls. We will tell you what they said.
Twenty thousand transcripts, analysed into themes, objections, competitor pressure and churn signals, in five working days. You keep the output whether or not you deploy a single agent.
Analysis runs in a single-tenant workspace. Recordings are deleted at the end of the engagement on your instruction, and we send you the deletion receipt.