AI for a sales team: scripts, emails and call review from transcripts

A neural network for a sales team will not replace a manager at the negotiating table, but it takes the text routine off their hands: email drafts, call review, filling CRM cards, script editing. These are four tasks with a clear input (text) and a clear output (text or a table), so they are easy to automate and verify. Below, in order: how to set up each one, a code example for call review and a cost calculation for a team of ten managers.

Task 1. Scripts and objections

A sales script is text with branches. A model helps in three places:

• A script draft for the product. You provide a product description, a customer profile and the call goal, and get a structure: greeting, questions, presentation, close.
• Answers to objections. A list of your real objections ("too expensive", "we do not need it", "we already work with someone else") becomes answer options in your tone.
• Practice. The model plays a customer with a given character and the manager rehearses the dialogue in chat.

An important limit: give the model only facts you are ready to confirm. Insert prices, deadlines and terms into the script from verified sources, and do not ask it to "come up with benefits". An invented guarantee in a script becomes a customer complaint.

Task 2. Follow-up emails

After a meeting or call a manager needs an email that repeats the agreements and proposes the next step. It is nearly template work, and a model does it quickly if you give it three things: a summary of the conversation, a tone (formal or not) and mandatory points.

Write a follow-up email to the customer after the call.
Tone: business-like, no bureaucratese, up to 120 words.
Required: thank them for their time, repeat two agreements, propose a call on Thursday or Friday.
Do not promise discounts or deadlines that are not in the summary.
Summary: ...

The line "do not promise anything that is not in the summary" insures against the most unpleasant mistake. The manager reads the draft, edits and sends it personally: automatic sending of emails to customers on behalf of the company without review is not worth enabling.

Task 3. Call review from a transcript

Language models work with text, so you first need a transcript of the conversation. It comes from your telephony or a separate speech recognition service, and the model analyses the finished text. What you can get from one transcript:

• a short summary of 3–5 lines;
• the customer's identified needs and objections;
• agreements and the date of the next step;
• script compliance: did the manager greet, did they ask the key questions;
• a draft CRM record.

A Python example with the official openai SDK. The address is https://api.zukko.pro/v1, parameters are described in the chat completions documentation↗:

import json
from openai import OpenAI

client = OpenAI(base_url="https://api.zukko.pro/v1", api_key="fetch_your_key")

SYSTEM = """You are a sales department analyst. You get a call transcript.
Return only JSON with fields:
summary (string, up to 300 characters),
needs (list of strings),
objections (list of strings),
agreements (list of strings),
next_step ({"what": string, "when": string or null}),
script_followed (object with boolean fields greeting, needs_questions, closing),
confidence ("high" | "medium" | "low").

The rule "if it is not in the text, use null" exists so the model does not invent agreements. Validate the JSON in code and retry on failure; techniques are collected in the article on strictly structured JSON↗.

Task 4. CRM hints

From the call review result you can build a CRM record: subject, funnel stage (if you have rules for moving between stages), a next-step task. A practical scheme: the model proposes, the manager confirms with one click, and only then does the record go into the system. Automatically creating deals and changing stages from model text without a human is risky. For the link with a specific CRM read the article on amoCRM integration↗.

Personal data

Transcripts contain names, phone numbers and sometimes bank details. Before sending text to an external service, check personal data legislation and your company policy. A practical minimum: mask phone numbers and document numbers before sending, do not write whole transcripts into application logs, keep the key out of the repository. Techniques are described in the article on logging without leaking personal data↗.

What it costs per month

Calculated for 18 October 2026. Assumptions for the example: 10 managers, 15 calls a day each, 22 working days: 3,300 transcripts a month. Each transcript is 6,000 tokens plus an 800-token prompt, so 6,800 input tokens and 500 output. Additionally 2,200 follow-up emails a month: 1,500 input tokens and 300 output. Read real numbers from usage: Russian text often takes more tokens than it seems.

Formula: number of requests × (input × input price + output × output price).

For ten managers that is from tens of cents to fifteen dollars a month in tokens. For comparison, an hour of manager time spent on manual call review costs noticeably more. But the point is not the price, it is verification. Take 30–50 calls you have already reviewed by hand, run two or three models and compare the conclusions. Which model understands your product and customers better nobody can tell you in advance. Current prices and the model list: GET https://api.zukko.pro/v1/models.

So that a script bug does not cost much, create a separate key for the sales team in the dashboard↗ with a daily and monthly dollar limit. Top-up is available via SBP, MasterCard or crypto.

FAQ

Can a neural network listen to calls itself?

The models in our catalogue work with text (some also accept images). Audio must first be turned into text in telephony or a separate speech recognition service, and then the transcript is passed to the model.

Can I send emails to customers automatically?

Technically yes, but it is wiser to leave draft review to the manager. A model can distort an agreement or add an extra promise, and the email goes out on behalf of the company.

Which model suits call analysis?

Start with an inexpensive one and test on 30–50 calls already reviewed by hand. If it misses nuances, try a stronger model and compare the result and cost.

How do I assess review quality?

Compare the model's conclusions with manual review of the same calls: do the summaries, objections and next step match. Separately check whether the model invents something that was not in the conversation.