/ 03.AINTENT
Intent detection
Why is the customer here? Classifies each interaction into a stable taxonomy — billing, refund, defect, delay, technical, access, cancellation.
bert-base42 classesmultilingual
A mid-sized consumer services company handles thousands of support interactions a month — calls, chats, emails, surveys. We replaced random QA sampling and spreadsheet reporting with an AI system that reads every interaction, scores it, and routes the signal where it matters.
Call transcripts
Speech-to-text from contact center recordings.
~9.2K / dayChat logs
Live agent + bot conversation threads.
~5.1K / dayEmail tickets
Support inbox messages and threads.
~2.4K / dayCRM data
Customer history, plan tier, ticket metadata.
All accountsSurveys
CSAT, NPS, post-resolution feedback forms.
~640 / dayAgent notes
Internal resolution notes and wrap-up codes.
~1.1K / daySocial mentions
Public posts referencing the brand.
~310 / daySupport generates more data than any other function in the company. Almost none of it makes it back into the operating loop. Random call sampling and manual tagging cover less than 2% of interactions — the other 98% is invisible.
Conversations stream in from seven channels, get cleaned and normalised, pass through six AI analysis components, surface in dashboards, and trigger downstream actions — escalations, routing, coaching cues.
Every interaction passes through the same six analyses. The outputs land as structured labels in the warehouse — the dashboard and the routing engine read from there.
Why is the customer here? Classifies each interaction into a stable taxonomy — billing, refund, defect, delay, technical, access, cancellation.
Tracks emotional tone turn-by-turn. Detects positive, neutral, negative, frustrated, angry — and flags the moment a conversation tilts.
Multi-label classifier over service quality, product, staff behaviour, refund delays, technical bugs, delivery. Topic modelling surfaces emergent themes.
Predicts customer satisfaction even when a survey is never submitted. Inputs: tone, resolution speed, repeat contact, escalation level, agent quality.
LLM-judged eval suite scoring empathy, resolution efficiency, compliance adherence, response time, escalation frequency, sentiment-improvement delta.
Beyond the words: interruptions, long silences, frustration spikes after holds, compliance violations, escalation triggers.
A real-world shape of inbound interaction. The transcript is shown left; the AI layer's output for that conversation is on the right.
Hi. My internet has been down for 2 days and nobody is helping. I've called three times already.
I'm really sorry to hear that. Let me pull up your account — can you confirm the address on file?
Sure. But I want to know why I keep getting passed around. I work from home. This is costing me money.
Completely understand. I can see two earlier tickets — I'm escalating this to the field-ops team now so it stays with one owner. Expect a call within the hour.
The dashboard is built around the three questions ops leaders actually ask: what's happening right now, what are customers complaining about, and which agents need coaching.
| Agent | Tier | Empathy | Resolve | CSAT | Flags |
|---|---|---|---|---|---|
| Priya R. | A | 0.91 | 0.88 | 4.7 | — |
| Marco D. | A | 0.85 | 0.92 | 4.6 | — |
| Sara N. | B | 0.79 | 0.74 | 4.1 | 2 |
| James K. | B | 0.72 | 0.81 | 4.0 | 1 |
| Lina P. | C | 0.58 | 0.62 | 3.2 | 5 |
| Ahmed Y. | C | 0.54 | 0.66 | 3.0 | 6 |
Measured against the 90-day pre-launch baseline. All deltas computed on the population, not the sample.
Off-the-shelf where it earns its keep. Custom only where it actually matters — the eval rubrics, the topic-clustering, the agent scoring.
The interesting failure modes show up in production, not the notebook.
Data quality
Noisy transcripts, mixed languages mid-call, incomplete CRM joins. We invest in normalisation more than in models.
Privacy & compliance
GDPR, HIPAA where in-scope, PCI DSS for payment talk. PII masking happens before anything reaches a model.
Model bias
Emotion and accent are where misclassification concentrates. We audit by language and demographic monthly.
Integration surface
CRM, telephony, ticketing, BI — five systems, three vendors. Most of the engineering work lives here.
Currently in design or pilot. Shipping through 2025.
Real-time agent assist
Recommended responses, next-best-action, contextual knowledge-base snippets while the call is live.
Generated summaries & coaching
Auto-written call summaries; per-shift coaching memos with the three things to work on tomorrow.
Voice-emotion detection
Prosody-aware stress and frustration scoring, layered on top of the text signal.
Autonomous resolution agents
Closed-loop agents for the long tail of simple, high-volume intents — password resets, plan changes, status checks.
If you're sitting on a year of call recordings, chat logs, and tickets you've never read, we'll help you turn them into a working system. Build & operate, or build-and-hand-off — your call.