Conversation intelligence for sales meetings

Executive summary

Client: an enterprise client whose sales team runs discovery and other meetings with customers.
Problem: customer insight sat in notes that were hard to count, compare across accounts, or trace to a quote.
Meeting records: every recorded meeting filed under its account and deal.
Vendor and tool data: the vendors and tools each customer uses, confirmed by the rep.
Quote-linked evidence: customer statements labeled by type and tied to the exact quote and speaker.
Executive view: pain points and vendor mentions ranked across accounts.
Main lesson: whether meeting participants agree to be recorded limits a conversation platform more than any model does. Design for the meetings that are not recorded, and weight every figure by who said it.

The problem

Customer meetings reveal which vendors and tools a customer uses, where things break down for them, and what that costs. Reps hear these signals every day. The gap is getting them to leadership in a usable form. The client captured them in notes and documents in varying formats, which cannot easily be counted across accounts or traced back to what the customer actually said.

What the client can see from its meetings

A record of every recorded meeting

Each recorded meeting produces a live transcript, AI key points and a summary, action items, and confirmed attendees. The record is filed under the customer account and the deal, so managers can see which meetings happened on each deal and who owns each follow-up. It covers any customer meeting and supports multiple meeting templates, discovery among them.

Vendor and tool data the rep confirms

During the meeting, the AI picks up which vendors and tools the customer uses today. The rep accepts or rejects each item, and a rejection needs a short reason. Accepted items are meant to fill in the account record, so sales leadership can see verified, current account data without relying on manual entry.

A labeled dataset of what customers said

Transcripts are labeled as they arrive: workflow steps, pain points with severity, tools and vendors, stated metrics, claims, and decision points. Each item keeps its exact quote, speaker, and confidence score, and the rep approves it. Pain points map onto a governed taxonomy of recurring patterns, so different phrasings of one problem become one countable pattern with several supporting quotes. The aim is for sales and product leaders to see how many accounts report the same problem. Because every label stays tied to its source, past conversations can be relabeled for new questions without asking customers again, and what was learned in each deal stays available to the whole team.

Discovery coverage

The AI ticks off which questions from the company's existing discovery question list were asked. Unasked ones stay visible, so a manager can see which areas still need a follow-up conversation. Coverage views would also show, for each deal, how many conversations and stakeholders have been captured, so gaps in discovery are easy to spot.

Cost figures that show where they came from

The cost view is designed to tag every figure by source: stated by the customer, estimated by the rep, drawn from a benchmark, or an AI-estimated range. Sales leaders reviewing a proposal would see what each problem is worth to the customer and how much weight each number deserves.

Drafts the rep finishes

These give the customer's leadership a clear picture of the current process and its problems, and the rep checks each one before it is shared.

Workflow map generated from the conversation, for the rep to edit, confirm, and export in an editable format.
Summary for the customer's leadership in the customer's own words, with several AI-ranked candidate quotes for the rep to choose from.

An executive view across accounts

Designed for executives, this view ranks pain points by how many accounts mention them and shows vendor mentions as a rate per 100 conversations, so leaders can see which problems are most widespread and where customer pressure is rising. Comparisons stay hidden until enough conversations stand behind them, and every figure links back to the verbatim quote, speaker role, and account behind it.

Also in the design: account picture, audit trail, demo prep

One picture of each account, stitched from every conversation with that customer.
Audit trails: a record of what the AI suggested against what the rep approved.
Pain points carried forward as reminders of what to show in the product demo.

What we learned

Consent limits the product more than any model does

In business meetings, not every participant will agree to be recorded, and a conversation platform has to work well for those meetings too. The design treats an unrecorded meeting as a normal case, not a failure. The specification lets reps add details by hand, such as figures gathered after the conversation. Consent should come from each participant directly. The design calls for scripted verbal consent at the start of each meeting, a go or no-go decision on recording, and a consent record for each person. As a general practice, involve legal review of consent wording early, since recording rules vary by location.

Give the customer a reason to say yes

Put the benefit in the product, not only in the rep's pitch: a shorter meeting, fewer follow-up meetings, and notes the customer can check. The design calls for emailing those notes right after the meeting, with a chance to correct them. The consent step confirms the customer's email address, so the notes reach the right person.

Capture context up front

Context added at capture, such as account, speaker role, and deal stage, decides what can be counted later.

Weight each figure by who said it

Rank evidence the way buyers already weigh claims: figures from the customer or a reference account count most, a benchmark less, and an internal estimate least. Keep cost and return apart too: the cost of a problem exists whatever the customer buys, while the return depends on a purchase that may fix only part of it.

Put a human between the AI and the system of record

Treat the rep's accept or reject as a gate, not a formality: nothing the AI captures should reach the CRM without it. Require a reason for every rejection, and you get labeled feedback for tuning the model as a by-product of normal work.

Build for more than one meeting type

Conversation tools tend to be useful beyond one meeting type, so build in templates from the start. Make in-meeting guidance optional so each user can choose how much to see. Recordings and summaries deliver value either way.

Tell users early suggestions will be basic

Early suggestions will be basic until the system is tuned on the company's own conversations. Tell users before they start.

Filter more than personal identifiers

Removing names and contact details is not enough. The specification also requires keeping only what relates to the business discussion, and dropping personal and unrelated content at the point of capture.

Get transcription and capture right early

General speech models often mishear product names and internal acronyms, so plan a custom vocabulary from the start, as this project's specification does. Capture had to work for any meeting format with nothing to install. In the build, capture starts from a one-time secure browser link and runs on common mobile devices. Keep raw audio only as long as needed to check transcript accuracy, and avoid playback features unless they are required.

Want this view of your own customer conversations? Talk to us.

Turning sales conversations into evidence that leadership can count and trace