I built a system that organizes calls to companies. It pulls in a list of businesses, leads me through the successive stages of contact, attaches the recording to the right company and helps prepare a summary of the conversation.
It also had two layers of artificial intelligence. Before the call, the first one read the company’s website and prepared a short cheat sheet. After the conversation, the second one analyzed the transcript and arranged the material for a quote.
Both worked. Only one solved a real problem.
What was supposed to happen
The process was organized from start to finish: a new company, an attempt at contact, a scheduled conversation, a date sent, analysis, a quote, a decision. For every entry I knew what had already happened and what the next step should be.

Every company has a clear stage, so no promised contact gets lost in the notes.
On paper, the greatest value was supposed to be AI’s help before the call. The system visited the company’s website, summarized it and suggested a way to open the conversation. It sounded reasonable: less preparation, more calls on target.
What happened in practice
First, you have to make the call. It is obvious, but this is exactly the part that is easiest to hide behind building a tool. You can spend a weekend polishing the screen, the order of the stages and the automatic notes, and on Monday still not have talked to a single company.
Second, a public number often leads to the front desk. Reaching the decision-maker takes a calm conversation, tact and the ability to explain why you are calling. No language model does that work for the salesperson.
Third, company websites are often sparse, out of date or written so generally that no useful cheat sheet can be drawn from them. The automatic summary looked neat, but did not give me the confidence I needed during the conversation. I opened the website myself anyway.
Technically, this part of the system worked. In business terms, it did not save me time.
Where AI really helped
After the conversation the situation was the reverse. The recording contained real material: the client’s words, their problem, the exceptions, what was agreed and the next steps. The model did not have to guess on the basis of a general website. It could organize what a human had drawn out earlier.
This part turned out to be useful. After putting down the phone I got:
- a transcript of the conversation,
- a list of the most important needs and constraints,
- the agreed next steps,
- a draft of a message with a summary or a quote.
I did not hand the model the decision about the price or the sense of the project. I got well-arranged material on the basis of which it was easier to make my own decision.
The most boring part had the greatest value
The second winner was an ordinary register of contacts. With no artificial intelligence at all: save the conversation under the right company, mark the stage, set the date of the next call and not lose a promise made a week earlier.
It does not sound spectacular. That is exactly why it is easy to underrate. In a company, a simple action performed reliably often has more value than an impressive feature that does not change everyday work.
Recording calls means obligations
Analyzing a conversation requires recording and processing it. A voice, a phone number and the content of a conversation can be personal data. That is why the record button stays locked until consent has been noted.

First the consent, only then the recording.
This does not close the subject of data protection. You still have to settle the purpose, the retention period, access and the model provider. But it shows an important rule: the legal and organizational cost should go into the assessment of a feature just like the cost of writing it.
The streetlight
There is a parable about a man looking for his keys under a streetlight. He lost them a few meters away, in the dark, but he searches where the light is better.
In AI projects it is easy to do the same. The light is brightest where you can build models, connections and beautiful screens. The keys, however, may lie in a less convenient place: in the conversation with the client, in reaching the decision-maker, in changing an employee’s habit or in putting the data in order.
Artificial intelligence is good at organizing material that already exists. It cannot always obtain it. In this project a human brought the truth back from the conversation, and technology helped not to lose it.
It is a less spectacular conclusion than “AI runs sales”. But you can base a tool on it that really stays in everyday work.