Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Build an asynchronous pipeline: capture audio, transcribe with speaker labels, summarise and extract structured fields (next steps, deal stage signals, competitors mentioned), then propose CRM updates for the rep to confirm rather than writing them silently. Keep facts linked to transcript timestamps, respect consent and recording laws, and measure extraction accuracy and how often reps accept proposed updates.
How to approach it
Clarify scale (calls per day, average length), latency needs (minutes after the call is fine), where calls happen (video platforms, phone), which CRM fields matter, and consent requirements across regions. Then design the pipeline stages and decide carefully what is automatic.
A strong answer
Scale. 5,000 reps, about 20,000 calls a day, average 30 minutes. Summaries are needed within a few minutes of the call ending, not in real time.
Pipeline.
Follow the flow, one stage at a time.
Illustrated sequence; animation speed does not represent processing time.Read diagram description
Record the call only with consent, queue it, transcribe with speaker labels, summarise and extract fields with schema-constrained output, validate the fields against CRM records, propose updates, let the rep confirm or edit each one, then write to the CRM with an audit trail. Rep edits feed back into the extraction evaluation.
1. Capture and consent. Recording laws differ by country and, in the US, by state; some require all parties to consent. The system announces recording and records consent status per call. Calls without consent are not processed.
The platform defaults will not do this for you. Microsoft Teams has a participant agreement policy that, once recording or transcription starts, mutes everyone else until they answer Yes or No, and gives anyone who says No a view-only meeting. Its documentation (checked 6 November 2026) lists the policy as off by default, so a tenant that never changed it records every participant without asking. The legal exposure of the other design is visible in Brewer v. Otter.ai, a proposed class action filed in the Northern District of California on 15 August 2025 (5:25-cv-06911). The complaint alleges that Otter's notetaker could join a meeting through a linked calendar without the affirmative consent of any participant, and that recordings were used to train its transcription models without telling attendees. No court has ruled on them. They still map each decision you own: who has to consent (the rep, or every person on the call, including the customer), whether a non-user can stop the recording, and whether customer audio may train anything.
2. Transcription. Speech-to-text with speaker diarisation (who said what), tuned for the languages and accents reps and customers use. Domain vocabulary (product names, competitor names) improves accuracy and can be supplied as hints.
3. Summary and extraction. A model produces a short narrative summary and structured fields: next steps with owners and dates, objections raised, competitors mentioned, budget and timeline signals, sentiment. Use schema-constrained output. Every extracted fact links to the transcript timestamp it came from.
4. Validation. Cross-check fields against the CRM: does the account exist, is the contact on the opportunity, is the mentioned product one we sell? Flag mismatches.
5. Proposed updates, not silent writes. The CRM drives forecasts and commissions. A model that silently moves a deal to "commit" because a customer sounded positive distorts the revenue forecast. So the system proposes updates ("Add next step: send pricing by Friday", "Competitor mentioned: X") and the rep confirms with one tap. Low-risk fields (call logged, summary attached) can write automatically.
6. Cost and throughput. 20,000 calls a day of 30 minutes each is 10,000 hours of audio a day for transcription, plus a summarisation call per transcript (a 30-minute call is very roughly 5,000 to 8,000 tokens of transcript). This is a batch-friendly workload: queue it, process in parallel, and use batch pricing where results within minutes rather than seconds are acceptable.
Evaluation. A labelled set of calls with human-written summaries and fields. Measure field-level extraction accuracy (especially next steps and dates), hallucinated commitments (things the summary says were agreed that were not), and transcription accuracy per language. In production, track acceptance rate of proposed updates and edit patterns: a field reps always edit is a field the model gets wrong.
What interviewers probe next
"Sales leadership wants deal stages updated automatically for forecast accuracy." Show them that model-inferred stages from call tone would make the forecast less accurate. Offer stage suggestions with evidence that reps confirm, and measure whether suggested stages predict outcomes.
"Customers in the EU object to having calls analysed." Respect consent, offer opt-out, process in-region where required, and minimise retention of audio and transcripts.
"Can we use the call recordings to improve our models?" Only with a consent the customer actually gave for that purpose, recorded per call. Default to no training on customer audio, keep it in the contract, and show it in the consent prompt. The Otter complaint is a reminder that the training question gets asked by the people on the call who never signed up for the product.
How it's scored
| Signal | Weak | Hire | Strong Hire |
|---|---|---|---|
| Pipeline | Transcribe and summarise | Adds extraction and CRM integration | Async queue, diarisation, schema-constrained extraction, validation against CRM, timestamp links |
| Automation judgment | Writes everything to the CRM | Human review option | Proposed updates for forecast-relevant fields, automatic only for low-risk fields |
| Constraints | Ignored | Mentions privacy | Recording consent by jurisdiction, regional processing, batch economics, acceptance and edit metrics |
Common mistakes
Letting the model change deal stages that drive forecasts and commissions.
Ignoring recording consent laws.
Building it as real-time when batch processing is cheaper and sufficient.