When you build an AI assistant for a sales team, the temptation is to give it a chat box: “ask anything about your clients”. It looks good in a presentation. On a company’s real data, things look different.
Below is what we tested for a distribution company in September 2026, and what we delivered.
What an Area Sales Manager needs before a visit
The Area Sales Manager walks in to see a client after a few minutes of preparation, usually in the car. They do not need a conversation. They need the answers to a few fixed questions:
- does the client have overdue payments? how much and since when?
- how is this month going compared with the same period last month?
- which product groups are falling, which are growing?
- what did they buy before and stopped ordering?
- what do similar clients buy that they do not?
This list did not come from a questionnaire. We built it starting from what the person in the field actually does in the first five minutes of a visit. We even removed things that seemed obvious. Stock, for example, is not vital before a visit: the client places an order, and if a product is missing, the warehouse restocks it.
This list produced a single command: /vizita <client>. The Area Sales Manager types the client name and gets a one-page brief.
What we learned from testing open conversations
In parallel we also tested the “ask anything” version: natural language questions put directly to the AI assistant. We tried both a local model and a larger model in an EU data centre.
In testing, we saw three kinds of problems:
- language: answers in English to questions asked in Romanian;
- consistency: the same question, asked several times, received different answers;
- the numbers: for a question of the type “top N clients by overdue payments”, the model received the correct rows from the CRM but displayed more than existed, filling in the list on its own.
The last one was the conclusion that mattered: even with the correct data in front of it, a model that rewrites a table on its own cannot be trusted with numbers.
What we delivered
Two decisions defined the product.
1. The commands do not use the model. /vizita and /briefing read the data from the CRM and display it in a fixed format. The same numbers the sales director sees in the CRM, without interpretation. There is nothing to invent.
2. When the model is used, the text with numbers is written by the server, not by the model. The CRM tools return the final text directly, in Romanian, and the AI assistant passes it on unchanged. The model picks the right tool and phrases the rest; it does not touch the amounts.
The result for users:
- Area Sales Managers use the commands: fast, deterministic, on the phone, including by voice dictation;
- management also has access to open conversations, on individual clients, with the limits stated explicitly. For a question where there is no suitable tool, the AI assistant says so instead of guessing.
In short
- Decide yourself what the user needs to see. A good brief is a short, fixed list, not an open conversation.
- Numbers do not pass through the model. The model can choose and phrase; tables and amounts are calculated and written on the server.
- Refusal is a feature. “I do not have the tool for that” is a good answer. A wrong number, stated with confidence, is the most expensive answer possible.
- Test on real data, with repeated questions. A correct answer on the first try says nothing about the third.
Open conversations remain useful for people who analyse data. For the person who walks into a client in five minutes, a command that always answers the same way is more useful than a model that answers elegantly.
Product details: Quiet Core and the full case.