Over the past two years, "AI-powered" has found its way into the pitch for almost every sales product. In the demo, the assistant forecasts next quarter's revenue, names the customers you are about to lose and whispers to the rep what to say next. At a real distribution company, the assistant usually meets a different question first: "Why is this customer's name in the system spelled three different ways?"
When we added an AI assistant to BernSFA and BernCRM, we gave ourselves one simple rule: the assistant should shorten a job that a person already does but finds dull or slow. This piece is about where that rule led us, and where we had to stop.
Where it works, part 1: Visit and meeting summaries
Writing visit notes is the job reps like least. They walk out of a customer's shop with the next address already waiting, so the note is either never written or shrinks to "visited". Three months later, a different rep calls on the same customer with no idea what was promised last time.
What AI does here is modest but worth having: it turns a few hurried, scattered sentences from the rep into an orderly summary. What was discussed, which product came up, what was promised, what to check on the next visit. The same applies to office-based salespeople, whose notes land on the customer record as a readable summary after each meeting.
One detail matters a lot: no summary should be saved until the rep has approved it. AI can misread what happened, so the last word belongs to the person who was in the room.
Where it works, part 2: Next-step and order suggestions
A good rep knows their customers: "This grocery buys detergent at the start of every month, and this month it hasn't." But nobody with a hundred customers can hold that kind of detail in their head for all of them.
Here AI can look through past orders and visits and offer suggestions:
- This customer hasn't bought a product it usually takes in this period; worth asking about.
- A quote went out two weeks ago and there has been no reply; it may be time to follow up.
- An item the customer asked for on the last order was out of stock then and is back now.
These are reminders, not decisions. The rep accepts a suggestion, edits it or skips it. The assistant is there to steer the rep's attention to the right place, not to do the selling for them.
Where it works, part 3: Managers asking for reports in plain language
Plenty of managers think in questions rather than report screens: "Compared with last month, which rep on the Asian side of Istanbul made fewer visits?" or "Which customers did we send quotes to this quarter without winning an order?"
The answers are usually sitting in the system already, but someone has to set the filters, export to Excel and build a pivot table. Asking in plain language removes most of those steps. The assistant interprets the question, pulls the relevant records and replies with a table or a short explanation.
A rule applies here as well: the assistant has to show which data its answer rests on. A manager should be able to ask "where did this number come from?" and open the list underneath it.
The overhyped part
Some promises are worth treating with caution:
- "AI will forecast your sales." Forecasting models learn from history. At a company whose data covers only a few months, or has gaps and inconsistencies, the forecast may be no more reliable than the manager's gut feeling.
- "The assistant will talk to customers for your reps." In B2B distribution, the relationship is the rep. Flooding customers with automated messages tends to damage it.
- "Install it and the AI takes care of the rest." An assistant is only useful when the system holds consistent records, which brings us to the real point.
Data first, AI second
An AI assistant can be no smarter than the data underneath it. If one customer exists under three different names, if visits go unrecorded, if some orders come through the system and others over the phone, every answer the assistant gives will be partial.
That is why the order of work matters:
- Make recording a habit. Are visits, orders and quotes actually being entered into the system?
- Give master data a single owner. Customer accounts, products and prices should come from the ERP rather than being copied by hand on the sales side.
- Clean up the data. Merge duplicate customers and weed out stale records.
- Only then switch the assistant on. And start small, for instance with visit summaries alone.
In the field, a recording habit rests on making data entry easy for the rep. When the visit note and the order are entered during the visit instead of afterwards, they are more complete and they attach to the right customer and the right date. That is also how the data an AI assistant needs gets built.
KVKK: which data goes where?
Visit notes, meeting summaries and customer records all hold personal data: the contact person's name, their phone number, sometimes a personal remark. Before any of it is sent to an AI service, you need answers to these questions:
- Where does the data go? What infrastructure does the model run on, and does the data leave the country?
- Is the data used to train models? The contract should exclude this explicitly.
- What can the assistant see? A rep's assistant should see only the customers that rep is authorised to see; the role and permission model has to cover the assistant too.
- Who decides what gets sent? The safest path is never passing unnecessary personal fields to the assistant at all.
- Is there an audit trail? You should be able to trace who asked which question and which data was accessed.
In BernSFA and BernCRM, data is held on the customer's own server or in a cloud of their choosing, and the role and permission model, audit logs and encrypted communication are designed in from the start.
In short
In sales software, AI earns its keep on work people already do but tend to let slide: summarising notes, flagging what has been forgotten, turning a report into a question. Forecasting and automation promises only mean something once the data is consistent. Recording habits and clean data come first, AI after.
The AI assistant in BernSFA and BernCRM works inside these limits: visit and meeting summaries, next-step and order suggestions, and plain-language report queries for managers.
Planning a project like this?
Plan it in 10 steps