Sinimex presented an AI coach for enterprise sales with document-based verification

Synimex has unveiled a prototype of an AI trainer for B2B sales that teaches managers to conduct negotiations, handle objections, and avoid overstating product features. The system is already undergoing testing: after an architecture change, it became noticeably more stable, and the team expects that regular training will increase requests for commercial proposals by at least 15%.

From Failure to Simulator

The idea grew out of a typical B2B sales problem. A novice can confidently conduct the first part of a meeting, but gets lost when the client starts asking uncomfortable questions about the product, documentation, and the benefits of switching. In consumer sales, such a mistake can still be survived, but in the corporate segment, each failed meeting is a lost lead, team time, and money.

For companies with long sales cycles, such a failure is particularly painful and costly. That's why Synimex decided to build not just a chatbot, but a negotiation simulator. The trainer should guide the manager through the entire cycle of conversation from greeting to fixing next steps, raise realistic objections, and provide feedback after the conversation.

For training, they used two main types of data: visit scenarios with goals and typical objections, as well as product documentation in PDF, so the agent could rely not on "feelings" but on facts.

Architecture Under Control

The team tried three approaches. They rejected a single-agent scheme due to context overflow: in a long conversation, the model would start losing the thread and follow instructions worse. A classical multi-agent approach also didn't work: it's flexible, but too much like a black box, where it's hard to understand why one agent handed the task to another and at what stage the error occurred. In the end, they chose sequential prompting — a chain of narrow steps with controlled JSON output and separate prompts for each stage of conversation.

• prompt selection by conversation phase
• generation of multiple reply variants
• selection of the best answer by a separate agent
• sending the final reply to the manager

The first version of the system still failed: only 2 out of 8 dialogues were considered successful. After that, they added a second "editor" agent to the scheme, which selects the best answer option from several generated ones. This two-stage pipeline unloaded the main prompt and gave the model more variability without loss of control. In the second iteration, the result grew to 7 successful dialogues out of 8, and the answers themselves sounded more natural. This became the turning point of the project.

Checking for Fabrications

The next step was to complicate the simulation. The trainer now has difficulty levels: at easy level, the virtual client is more accommodating, at medium level requires arguments, and at hard level starts checking each manager's statement carefully. This is where RAG was added to the system so the agent could cross-check with documentation and catch unsupported promises.

Otherwise, the trainer would either believe any promise the manager makes or argue without relying on facts. The scenario looks like this: the manager makes a statement about the product, the system highlights the specific thesis, through vector search finds a relevant fragment of documentation, and then passes the statement and the found text to the model for a verdict. If the manager, for example, says that the cloud version of the product can handle 500 simultaneous users, but the documentation only contains data from a test with two, the agent doesn't automatically agree and gently points out the mismatch.

According to the team, each prompt was rewritten on average about 12 times to achieve such behavior.

What This Means

Synimex's project shows that corporate AI agents are increasingly shifting from "smart chatbots" to applied trainers with verifiable logic. If such a format confirms the forecasts for growth in offers and conversion to pilots, AI could take on the role of a permanent trainer in B2B sales, one that is available at any time and relies on documents rather than improvisation. For internal sales departments, this is no longer an experiment for experiment's sake, but a potential tool for scaling training.