AI head of sales: call scoring, lost-deal detection, CRM control
An AI head of sales that scores every call and thread, flags stalling deals and shows the owner exactly where revenue leaks.
Context
Sales oversight in most companies relies on a manager listening to a fraction of calls and trusting that the CRM reflects reality. Both assumptions fail at scale: the vast majority of calls are never reviewed, and the CRM decays into wishful thinking. The owner learns about a lost deal only after it is lost — and never learns why.
The system inverts this model. Instead of sampling, it scores every conversation and every thread, and continuously reconciles what was said against what the CRM claims.
Approach
Every call is transcribed with STT/Whisper and scored against a structured rubric; correspondence threads go through the same pipeline. Managers are ranked on the full scored corpus, with the supporting evidence attached to each score rather than a bare number.
Deals that went quiet are flagged with a probable reason extracted from the conversation itself — an unanswered objection, a competitor mention, a promise nobody followed up on. The owner sees the cause, not just the count.
Architecture and data flow
Four sources feed one analysis loop: the CRM for pipeline state, telephony for call audio, 1C for financial ground truth, and the company knowledge base for product and pricing facts. Transcripts flow into an LLM reasoning layer whose conclusions are grounded in that knowledge base, which keeps scoring consistent with the company's actual playbook rather than generic sales folklore.
Results are delivered on two surfaces: a daily Telegram digest for the owner and a management dashboard for drill-down — from a revenue-leak summary to the specific call that triggered the flag.
Key engineering details
The central mechanism is reconciliation between two states of every deal: the conversational state (what was actually discussed and agreed) and the declared state (what the pipeline shows). A deal advancing in conversation but frozen in the CRM is a hygiene violation; a deal advancing in the CRM with no supporting conversation is a fabrication signal. Both are raised automatically.
Grounding every LLM conclusion in the client's own knowledge base bounds hallucination: a score must cite the rubric criterion and the transcript fragment it rests on, so a disputed evaluation can always be traced back to the recording.
Outcome and what shipped
The result is a product line ready for pilot deployment: an AI audit of the sales department, full-coverage call-quality control, automatic CRM-fill verification, and early warning on revenue loss. Four modules, four integrated data sources, two delivery channels — designed to plug into an existing sales stack without replacing it.
The pilot path is defined: run the pipeline on a live team's call archive, calibrate the rubric to the client's playbook, then switch on continuous scoring.
What we built
Call & thread scoring
Every call and message thread transcribed and scored against a rubric — full coverage, not a sample.
Manager evaluation
Managers ranked on the scored conversations, with the evidence attached.
Lost-deal detection
Stalled deals surfaced with a probable cause drawn from the conversation.
CRM-fill control
Automatic checks that the pipeline reflects what was actually said and agreed.