One of six directions
AI does the work there are never enough people for.
Not «let us add a neural network», but a specific job: today it is either done by hand for hours on end, or not done at all, because a person physically cannot keep up. Below are two systems already running at clients every single day.
· 
· 
· 
AI head of sales · running in a car dealership
A head of sales who has listened to 100 % of the calls
A human head of department manages to sample a handful of conversations a week and judges by feel. The system listens to all forty to sixty calls a day, reads the WhatsApp threads and judges every one of them by the same ruler.
A customer calls — the service pulls the recording from the phone system, transcribes it with speakers separated, hands the text to the model together with the written sales standard and gets back a review: a score across eight blocks, a summary of the conversation, the extracted facts (which car, what budget, what timing, is there a trade-in, what was agreed), red flags. A short review for the salesperson lands in the customer’s card. If something serious came up in the call, the manager gets a Telegram message immediately. From the end of the call to the comment in the card — no more than ten minutes.
The Telegram bot here is not a set of commands but a counterpart: the manager types «show me who is worst at booking visits» — and gets an answer from the data, not a report that still has to be read.
What is under the hood
- Volume
- 40–60 calls a day, 3–5 hours of audio
- Time to feedback
- no more than 10 minutes from the end of the call
- The recording
- deleted right after transcription — only the text remains
- Scoring
- 8 blocks, 100 points, one written standard
- What else it reads
- the customer’s contact history, the ad source, the deal stages
- Where it writes
- a comment in the CRM, an alert to the manager, an evening digest for the salesperson, a weekly report
- What else it reviews
- WhatsApp threads and compliance with the response-time standards
- Economics
- Google Ads and Meta spend is matched to deal stages: cost per lead, per visit and per sale by channel
- A second client
- is connected through settings, not by rewriting the code
The audio is deliberately not kept. That means less sensitive data on your hands and fewer questions under Law 195/2024, where a call recording counts as personal data.
The ruler
What the system scores a conversation against
The weights are not made up. In the dealership’s funnel, 408 enquiries in a month produced 233 visits, and 199 of those visits ended in a purchase — that is, 85 % of booked visits turn into a deal. So the salesperson’s job on the call is not to sell a car over the phone but to get the person into the showroom. Hence the points.
- 20 Uncovering the need make, budget, timing, who it is for, trade-in, method of payment
- 20 Selling the visit invited them in and gave reasons: see it in person, test drive, trade-in appraisal
- 15 A pitch that fits the need plus alternatives from what is actually in stock
- 15 Specifics a date and time for the visit, or a next step with a date
- 10 Opening the call named the dealership, their own name, the customer’s name
- 10 Initiative asks questions back, leads the conversation
- 5 Recap of what was agreed said at the end what the two of them had settled on
- 5 Tone and the customer’s language speaks the language the customer came in with
And the response-time standards — the system counts those separately from the points
- 5 minutesto respond to an enquiry from the website
- 10 minutesto call back after a missed call
- 15 minutesto the first reply on WhatsApp during working hours
- 3 attemptsto reach them on the first day — morning, midday and evening, then a message
- 0 customerswithout a next step with a date: «the customer is thinking it over» with no date does not exist
Uncovering the need and selling the visit together are 40 points out of 100, almost half. Never offered a visit at all — the whole block goes to zero. Did not ask about a trade-in — minus ten. Every weight and standard lives in the settings, not in the code: after the first weeks of work the numbers almost always change, and that must not cost a rewrite of the program.
Who gets what
One system, three different recipients
The salesperson, the manager and we ourselves need different answers out of the very same conversations. So each gets their own, rather than one shared report nobody reads to the end.
The salesperson
A short review of their own call in the customer’s card within minutes: what worked, what to fix, which wording was worth using. In the evening — a personal daily summary with points and the trend. Effectively a personal coach who heard every one of their conversations.
The manager
A morning summary of yesterday, instant signals on red flags and broken standards, a weekly report with the trend for each salesperson. No more sampling recordings to work out what is going on in the department.
The agency
Google Ads and Meta spend is matched to deal stages — the cost of a lead, a visit and a sale is visible per channel. The advertising learns from visits and purchases, not from enquiries.
Second example · runs every morning
The owner does not open the CRM. So the CRM comes to him
The owner of two companies has over a thousand tasks in Bitrix24 and zero appetite for opening the interface. We did not try to talk him into using the system — we moved the system to where he already spends his day.
At eight in the morning a team review arrives in Telegram: who is buried, who is carrying someone else’s work, where a task has not moved for three days. The model reads not only the tasks but the comments and the history of deadline pushes — that is where you see what is really happening. In the evening his own tasks arrive, each with buttons for «close», «push by a day, three, a week», «drop the deadline». On Saturdays — a weekly report with scoring and a comparison against the week before.
What is under the hood
- Tasks processed
- around 1050 in a single run
- Building the report
- 2 minutes, entirely without a human
- Schedule
- the team one at 08:00 Mon–Fri, personal tasks in the evening, the weekly one on Saturday
- Where it lives
- the client’s own server, no database and no cloud subscriptions
- After a failure
- it comes back up by itself, a minute of downtime at most
What this changes
The four things it was all built for
The manager stops judging by feel
A human head of department manages a few conversations a week and remembers the vivid ones. The system listens to all of them and measures with one ruler — which is how «Petrov is doing badly» turns into «Petrov’s median response time is 40 minutes against a standard of 5».
The salesperson gets a review, not a dressing-down
Within minutes of the call, two or three lines appear in the card: what worked, what to fix, which phrase was worth saying. There are deliberately no points there — a score in plain view of colleagues reads as a public flogging. The points arrive privately.
A problem surfaces the same day
The customer asked to come in and no visit was booked; «I will call you back» with no date and no call back; rudeness on the call. That reaches the manager immediately instead of coming out at the end of the month through fallen revenue.
It becomes visible where the lead is lost
The advertising already knows how to learn from showroom visits. Reviewing the calls adds a layer between the enquiry and the visit — and the argument «bad traffic or bad salesperson» is settled by facts for the first time, not by seniority.
How we take it on
Four steps, and the first one is not about technology
We look for the job, not the technology
The conversation starts not with «let us implement AI» but with the question of what gets done by hand in the company every day, or does not get done at all because there are not enough hands.
We write the standard down
The model cannot judge by its own taste — it needs a benchmark. For a car dealership that is a document of several pages: what counts as a good call, how many minutes to respond, what counts as a failure.
Into the log first, into service after
For the first weeks the system writes its conclusions only for itself, and we check them against reality. We turn on comments to people once we are satisfied it is not systematically wrong.
We leave it running without us
Auto-start, backups, a switch on every circuit. A system you have to prod every morning does not solve the job, it adds a new one.
Honestly
Where AI will not help
- The system scores only what it heard. Conversations on a personal mobile, bypassing the phone system, never reach it — and that creates a temptation. It is closed by agreements inside the department, not by software.
- AI will not put your processes in order. It will show them — and that is often an unpleasant sight the company turns out not to be ready for.
- If there is no data, or the data is junk, the model will write a tidy text about nothing. A report on an empty CRM looks convincing and means nothing.
- Jobs where the cost of a mistake is high and there is nobody to check the result, we do not automate. The model is wrong confidently, and you cannot see it in the text.
So the first conversation is about the job, the data and the standard, not about the model. If it turns out that fixing the process or hiring a person is cheaper, we will say so.
One step
Tell us what gets done by hand at your company
We will look at how much time it eats and whether it can be handed to a model. If the job will not pay back the development, we will say so plainly — that is faster and cheaper for both of us.
What happens next
- We call you back during working hoursUsually within an hour. If it is awkward to talk — we write instead.
- We ask a few questionsWhat you have now, what does not work, which numbers you count as a result.
- We tell you straightWhat is worth doing first, how long it takes and whether you need us at all.
Would rather not leave your number — call us: +373 76 090 111 We answer during working hours.