Coldread

Conversation analytics for sales and booking teams

Every call reviewed. Every loss named. Every ruble counted.

Coldread listens to calls from Bitrix24 and answers the questions a sales leader asks on Monday morning: why the customer did not buy, what it cost, who was never called back, what the team should say differently. With a verbatim quote and a timestamp under every number.

See the workspace on real calls What we heard

One-minute sign-up, code by e-mail instead of a password. Inside: 776 calls of a health resort, reviewed in a single night.

Calloutbound · 11:11 · 5:06reviewing…

0:06ManagerHello, this is the premium resort. You left a request on the website for a quote, is that right?

0:23ClientYes. Honestly any time works, I am retired. October, with treatment.

2:06ManagerTen nights with treatment on this rate come to 261,000, and I can give you a personal 15% discount.

4:40ClientThanks, I will write down the prices. I will think it over and call you back.

4:52ManagerSure, we will be waiting. All the best.

Loss reasonNeeds to consult family no next step
“I will think it over and call you back” PriceQuoted before the programme was presented price before value DynamicsManager talks 85% of the time · 13 questions vs 4 from the client AdviceOffer to send the quote on WhatsApp and agree on a day to call, instead of waiting for the client to call back.

776calls of one department reviewed overnight, no buttons pressed

₽15.6Mat risk: 81 sales conversations with no result × median booking value

46%of promised call-backs never happened (66 of 142)

12customer questions the managers could not answer, across 220 calls

8.9%of conversations end in a booking; booked calls have twice as many questions

Numbers from the demo workspace: booking department of a health resort, September, 570 conversations longer than 15 seconds. Money and call-backs are computed from CRM data without a model; every number opens down to the recording.

What we heard

Customers do not leave because of price. They get sent away.

The top loss reason in real calls turned out to be an action of the manager, not an objection: a transfer instead of an answer. The price objection came fourth. Here are verbatim phrases from the reviews and what the system did with each.

Manager · the client asked about a stay

“So look through a travel agency, contact a tour operator”

sent elsewhere90 of 570 conversations ended with someone else’s phone number. Action in the workspace: a single point of entry, the manager books or transfers the call instead of dictating a number.

Manager · the client wanted to book procedures

“I can give you the medical desk’s number. Please call them and they will advise you”

no informationThe most frequent unanswered question: spa, pool and mineral baths without a stay, 18 calls. The system drafted the missing memo.

Client · called back after the quote

“That is a bit expensive”

price before valueIn 19 calls the price was quoted before the client heard what it includes. The checklist item “programme presented” averages 0.45 out of 1.

Manager · the client compared with the website

“If the website shows that price, we can book at it. It is a promotional offer, let me lock it in for you right now”

worked · bookedReplies after which the objection went away are collected into a library: 26 verbatim client → manager pairs for onboarding new staff.

Workspace

Seven screens, each with one question and one conclusion

A screen title is a conclusion with a number, not the name of a chart. Below are the real titles of the demo workspace at the time of publishing.

Losses

Sent to another department: 36% of all losses. ₽15.6M at risk

Reasons with quotes, a reason × call type matrix, objection handling, who was never called back.

Open →

Demand

Peak demand: October, 99 requests; 30 product gaps named out loud

Arrival calendar, booking lead time, party size, rooms, budgets, what was missing.

Open →

Conversations

In booked calls the manager asks 11.8 questions, in lost ones 7

Talk share, monologues, pauses, interruptions, pace: from timestamps, no model. Heatmaps by hour.

Open →

Knowledge

12 gaps across 220 calls; 5 are closed by a single memo

Unanswered questions, a playbook from what already worked, answer cards for objections.

Open →

Managers

Weakest checklist item: programme presentation, 0.45

Checklist by item, conversation manner, coaching themes per person with links to calls.

Open →

Clients

50 clients called three or more times and never booked

Customer journey by phone number: “we will call you back → we will call you back” chains nobody owns.

Open →

How it works

From dial tone to conclusion without a single button

An event from Bitrix24 starts the pipeline. Every step can be replayed on its own: a new version of the ontology recomputes the reviews without touching the transcripts.

The call ends

Bitrix24 sends the event, Coldread fetches the recording itself. Short recordings and dial tones are counted but not reviewed.

Transcript with roles

Speech to text with timestamps, manager and client separated, conversation dynamics computed from the timestamps.

Review against the ontology

The model returns structure, not prose: call type, loss reason with a quote, outcome, entities, checklist, confidence.

Doubtful cases checked

Low-confidence reviews go to a review queue. Your “correct / fix” marks calibrate the threshold.

Workspace, Bitrix24, brief

The call in the workspace, a comment on the deal card, a morning brief in Telegram. The knowledge base and playbook refresh overnight.

≈10 min from the end of a call to its review776 calls in one night100% of calls instead of sampled listening

Results

Pilot: booking conversion rose from 31% to 43%

The technology Coldread grew out of ran for three months at a health resort on every line: 6,000 calls, 847 objection-handling patterns, 156 phrases correlated with a successful outcome.

Automatic analysis of every call replaced sampled listening by supervisors. The outcome of a conversation was confirmed by a booking in the CRM, not by an annotator’s opinion.

Pilot January–March 2025, before/after comparison with the same period. Published in the proceedings of the 68th MIPT scientific conference (M. Kadochnikov, MIAS). Results of a single deployment need confirmation on a larger sample.

+38%booking conversion: from 31% to 43%

+53%revenue over the pilot period

+13.7%average booking value

6,000calls processed, 1,284 classified as successful

Security

Customer data never leaves your perimeter

Recordings, transcripts and reviews stay on your server. What goes to an external model, if anything, is decided by policy, not by a developer in the moment.

Local model by default

Reviews run on a model on your GPU or in a dedicated environment. An external provider is used only to double-check doubtful reviews and can be switched off entirely.

Masking before sending

Phone numbers, e-mails and card numbers are replaced with placeholders before text leaves the server. The client’s number is stored as a hash and a mask.

Roles and an access log

The leader sees the department, a manager only their own calls, an analyst works without audio. Every playback of a recording is logged.

Retention and data residency

Audio 90 days, transcripts a year, aggregates indefinitely, all configurable. Personal data stays in-country; the “local only” mode works end to end.

Questions

What people ask before connecting

What do we need to connect?

Bitrix24 with telephony and access to recordings. The outgoing webhook takes an hour to set up; after that calls are reviewed on their own. No change of telephony or CRM.

Does it work outside health resorts?

The ontology (topics, loss reasons, objections, checklist) is edited in the workspace. Hotels, clinics, education, any phone sales with call recording.

How accurate is the model?

Every review carries a confidence score; doubtful ones go to the review queue, and your “correct / fix” marks calibrate the threshold. Money, call-backs and speech dynamics are computed without a model.

Can we try it on our own calls?

Yes. Send ten recordings and within a day we will show the reviews, checklists and the brief for them. Connecting Bitrix24 takes one working day.

Demo workspace

See how it looks on real calls, not on slides.

776 calls of a resort booking department: losses with quotes, money at risk, call-backs, knowledge base, coaching themes per manager. One-minute sign-up, code by e-mail.

See the workspace Send us your recordings