An AI assistant for a sales team is software that answers a customer's first questions from company documents, fills in the CRM record and drafts emails, while decisions on price, discount and deal terms stay with the sales rep. Rollout starts with a single operation and with a written list of what the assistant must never do: answer from data it does not have, quote a price of its own, or message a customer without consent to advertising.
Key takeaways
The assistant does not close the deal — it takes over the work around it: sorting inbound enquiries, collecting customer data, keeping records current and preparing text. The rep owns the deal: the assistant has neither the right to concede terms nor responsibility for the contract.
Reviewing calls and chats is a separate job for the same assistant: transcription, a short field-by-field summary and marks on which script stages were covered are described in the article on AI analysis of sales calls and messages.
Work around selling eats most of a rep's week — and that is exactly what the assistant takes on.
Do not trust it with anything where a mistake costs money or customer rights: price and discount, commitments on deadlines, turning a customer down, and sending messages without consent. The boundary is written down before launch and enforced by permissions, not by reminding the team.
| Sales operation | Who decides | What is configured in the system |
|---|---|---|
| Answering a standard product question | Assistant | Answers only from current documents and shows the source of the answer |
| Qualifying an enquiry, filling the deal record | Assistant | A closed list of fields and values; free text goes into a separate field |
| Price, discount, deferred payment | Sales rep | The assistant takes the price from the price list but never calculates a concession or names a range |
| Sending an email or message to a customer | Sales rep | The assistant prepares a draft; a person approves the send |
| Campaigns and calls to a collected list | Head of sales | Consent is checked before launch, opt-outs are kept in a separate list |
| Refusals, complaints, returns | Head of sales | The assistant gathers the facts and hands the case to a person |
| Editing and deleting CRM data | Sales rep | The assistant may create and complete a record; deletion is off limits |
The technical side of this rule is spelled out in agent documentation. OpenAI's guide to guardrails and human review tells you to pause the agent before actions with side effects — cancellations, edits, shell commands — and wait for a person to approve or reject the call; the check itself sits next to the tool that creates the side effect.
Important Law No. ZRU-1115 of 21 January 2026 added an article to Uzbekistan's law on informatisation: when taking legally significant decisions that affect human rights and freedoms, it is not allowed to rely solely on conclusions produced by information resources and systems built on artificial intelligence technologies. A refusal of instalments or of a shipment based on a model's score is signed by a person.
An invented answer is a property of language models, not a defect of one vendor, which is why numeric answers are pulled from the system of record at the moment of the conversation rather than from the model.
The authors of “Why Language Models Hallucinate” (September 2025) name the cause plainly: models are trained and graded so that a guess scores higher than admitting uncertainty, so on an unfamiliar question the system will produce a plausible but wrong answer rather than stay silent. That leads to one simple configuration rule: “I don’t know” has to be a permitted and expected answer.
How search over company documents works, and why it needs no model of your own, is covered in the article on AI assistants and chatbots for business.
The assistant gets a narrow set of operations on specific objects rather than a full employee role — plus its own account, so every action it takes is visible.
How roles are granted, what belongs in the log and how to limit database exports is covered in the article on access rights in CRM.
A customer's chat, phone number and call transcript are personal data, so the processing needs a lawful basis and the assistant needs a defined role in that processing.
The law «On Personal Data» No. ZRU-547 lists in article 18 the conditions under which processing is allowed:
Article 21 allows consent in any form that makes it possible to confirm it was obtained. So the notice at the start of the conversation and the consent flag are stored in the system, while purposes and retention periods are described in the data processing policy.
Campaigns and outbound calls are a separate matter. Article 30 of the law «On Advertising» No. ZRU-776 allows advertising to be distributed through telecommunications services only with the subscriber's prior consent. A list of warm contacts that the assistant assembled from conversations does not amount to that consent.
The cost of getting it wrong has gone up: the same law No. ZRU-1115 extended article 46-2 of the Code of Administrative Liability — unlawful processing of personal data using artificial intelligence technologies carries a fine of fifty to one hundred base calculation units with confiscation of the items involved.
The order of the steps matters more than the choice of model: each step shows a result, and the work can be stopped without losing what is already done.
The fifth step is the one most often skipped, yet it is the only one that shows how the assistant behaves with your wording, mixed Uzbek-Russian messages and abbreviated product names.
The mistakes repeat from company to company and almost always concern how the task was framed, not the technology.
The gain is measured not by the number of answers but by what changed in the team's work, so the metrics are chosen before launch and taken on the same sample.
The last metric only reads together with the funnel: how stages are defined and where deals are lost is covered in the article on the sales funnel in CRM.
Tip Take the metrics a month before launch. Without baseline figures, any discussion of the gain turns into an argument about impressions: “it got faster” can be neither confirmed nor disproved.
The price depends on which operation the assistant covers and which systems it is connected to. Benchmarks from the service page:
Three things move the price from there: the number of systems in the chain, the volume and condition of the documents, and the depth of testing before launch. The full list is on the service page Telegram bots and AI assistants.
We start with the process, not with the choice of model: we look at where the team loses time and measure the volume of manual work on that stretch. Then, together with the head of sales, we assemble the list of prohibitions — what the assistant decides on its own, what it prepares for approval and what it never does.
We connect the assistant to the systems you already run: CRM, 1C, telephony, Telegram. We test it on your past conversations, bilingual ones included, and launch it with a human in the loop — at first the rep approves outgoing messages. Under the contract, the rights to the code pass to you after full payment. How a model is connected to CRM, ERP and 1C is covered in the article on AI integration into corporate systems.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
No, it takes over the work around selling: the first reply, qualification, filling in the record and drafting emails. Price, discount and deal terms stay with a person, and legally significant decisions may not be taken by relying solely on the conclusion of an artificial intelligence system.
With the first reply to inbound enquiries and qualification: that stretch has a clear volume and a clear quality test — did the assistant answer the way a rep would have. Other operations are connected once that first stretch has run without errors for weeks.
Only with the subscriber's prior consent: the advertising law allows distributing advertising through telecommunications services on that condition alone. A list of warm contacts the assistant assembled from conversations is no substitute for consent, so opt-outs are kept separately and checked before every send.
Find out where it took the figure from. If the price came from model text rather than the price list or the system of record, the setup changes: numeric answers have to be fetched at the moment of the conversation. If two versions of the price list were in the base, the old one is removed from the index and the document gets an owner.
Not necessarily. The assistant is connected to the system you already run, provided it has an API and allows a separate account with restricted rights. Replacing the CRM makes sense when it lacks the deal fields you need and a change log.
By the metrics taken before launch: time to first reply, the share of enquiries left unanswered for more than an hour, the share of records with required fields filled in, and the share of answers the rep rewrote before sending. Without baseline figures there is nothing to confirm the gain with.
Yes, provided it has been tested on your conversations. The difficulty is not the two languages separately but mixed messages and product names in a third language, so real enquiries from the past month are run through before launch and compared with the rep's answers.
A process review and an automation plan with priorities costs 400 $. After that: a bot for requests and bookings from 1 200 $, automation of internal routine from 2 500 $, an AI assistant answering from company documents from 3 000 $, and team training from 300 $.
Sources
Cover photo: Mahdi Bafande, Pexels