article · November 10, 2026 · Gregory Shevchenko

How we built analytics around a car dealer's business: AI call analysis and end-to-end reporting for GAC

How we built AI call analysis and end-to-end analytics for the GAC car dealership in two months, and why an off-the-shelf analytics tool did not fit.


Cited across

  • ChatGPT
  • Gemini
  • Google AIO
  • Google AI Mode
  • Perplexity
  • Copilot

Cover for the article Analytics that listens to every call

The head of sales at GAC, a car dealership in Saint Petersburg, could not see what happened on the sales team's phone calls. Which callers were serious, whether managers followed the script and where leads got lost was judged by listening to a few recordings at random. Then a second request arrived from the marketing director. She needed to know which ad channels paid back, and she wanted amoCRM cleaned up so that the agency, the analyst, the manager, the marketer and the owner could all see the source of every deal.

In two stages of about a month each, we built a system that analyzes every call and shows the path from an ad channel to a sale. Below we explain how it works, why an off-the-shelf analytics tool did not fit and what the client got.

In this article

Section 01

Where did the project start?

The client already had a ready-made end-to-end analytics module from its phone provider, UIS. Almost nobody used it. The data did not add up to one picture, and the setup had never been finished.

The tool also had limits of its own.

  • UIS does not keep the history of deal stage changes, so a sales funnel cannot be built from it.
  • Source grouping was awkward. The same Avito marketplace arrived through several integrations with different tags and multiplied in the reports.
  • Deals were read from amoCRM fields, while ad spend came from contractors' reports. The two views never matched.

Section 02

What did we build in the first stage?

The first stage took about a month. We built a speech analytics system that we call "AI Head of Sales." It processes every call from the sales team.

Here is how it works.

  1. Each call is transcribed with the Whisper model.
  2. The text goes through four separate analyses with language models. The first decides whether the call is a target call. The second scores the manager on 13 criteria across five sections of the script. The third extracts the facts from the conversation. The fourth collects marketing insights.
  3. Each lead gets a grade of AAA, BBB or CCC based on purchase timing and budget.
  4. The system fills in 12 deal fields in amoCRM and adds the call score to the card.

On top of this we set up alerts and reports.

  • Telegram and Max alerts fire when nobody answered a lead within 5 or 15 minutes, a deal has no task, the script score is 2 or lower, or a new lead arrives. Alerts are split between the two showrooms.
  • A digest of marketing insights arrives at 8:00 every morning.
  • The head of sales receives PDF reports for the GAC and JAC brands.

Section 03

What did we build in the second stage?

The second stage took another month. We set up a PostgreSQL database in Yandex Cloud and began loading the deal stage history from amoCRM, tasks, ad spend and call recordings into it.

From this data we built a view that goes from channel to leads, prepayments, sales and cost per lead. It refreshes every 20 minutes and feeds a Yandex DataLens dashboard.

Call attribution has its own story. At first, sources came from UIS by session. Then the client moved to the CallTouch call tracking service, and we switched the sources to it. Some static numbers stayed in UIS, so there is a double safety net. If a call is missing in CallTouch, we take the source from UIS.

Section 04

Why did the off-the-shelf tool not fit?

We built our own system for three reasons.

  1. Our own source grouping. The system records every deal transition with dates from day one and folds sources the way the business needs. Avito no longer multiplies in the reports.
  2. A phone provider change without stopping the reports. Moving from UIS to CallTouch broke no report, because sources are a replaceable layer.
  3. Funnel specifics. The client has two types of refusal, one from the manager and one from the head of sales. Sales are counted by the date of the stage change, not the date the deal was created. And there are two funnels, one for GAC and one for JAC.

Section 05

What changed for the client?

  • Call control. Analysis covers about 100% of target calls. Before, calls were checked selectively.
  • Response speed. The 5 and 15 minute response targets are now under control, and the head of sales gets a finished report every morning instead of building one by hand.
  • Money by channel. Cost per lead and sales for each channel are visible live, not once a month from a contractor's report.
  • Time. By the team's estimate, the head of sales and the marketer save about 8 hours a week on data analysis and reports.

The client keeps ordering improvements. Recent requests include a breakdown by manager, a list of deals without tasks and full quotes from calls for review. That is a good sign: the system became a working tool and not a report for the sake of a report.

Section 06

What can you take from this?

  • If an off-the-shelf analytics tool does not keep the history of deal stages, it cannot build you a sales funnel. Check this before you buy.
  • Keep every deal transition with dates in your own database. Then a change of phone provider or CRM does not break reporting.
  • Start with call control. It gives a quick result and addresses the manager's main pain before end-to-end analytics is ready.
  • Group sources by the logic of your business, not by the tags in ad accounts.

Section 07

Want the same analytics for your business?

We will show how to calculate the path from channel to sale for your business. Contact us and we will review your funnel.

Section 08

FAQ

How long does the implementation take?

Speech analytics went live in about a month, end-to-end analytics in another. The full setup took about two months, followed by support and improvements.

What data does the system collect?

Call recordings, the deal stage history in amoCRM, tasks, ad spend and call sources from call tracking systems.

What happens if the client changes phone provider?

Reporting does not stop. Sources are a replaceable layer, which is how we handled the move from UIS to CallTouch.

Can this be used outside car dealerships?

The logic is the same for any business where the path from a call or lead to a deal matters. Only the funnel and the sources change.

For your team

Stop hiring agencies and freelancers

Hire not agencies and freelancers — but Marketing AI Agents for the AI Search.

  • Per-engine citation map across 6 AI engines
  • Content + schema work that earns the citation
  • Honest 30-min strategy call before you commit

Cited across

  • ChatGPT
  • Gemini
  • Google AIO
  • Google AI Mode
  • Perplexity
  • Copilot


Want to talk?

Book the strategy call. Thirty minutes, free.

An engineer from the team runs your brand through Hermes before the call.

You arrive to a per-engine citation map of your category, the closeable gaps, and an honest read on whether any tier fits.