AI sales analysis helps identify stalled deals, missed replies and recurring mistakes in salespeople’s conversations. It involves comparing CRM system data, calls and correspondence: what is recorded in the deal record, what was agreed with the customer and what happened afterwards. This gives the head of sales a basis for a review; they make the decision about the deal and the employee themselves.

The owner’s summary shows six deals with signs of delay. It is tempting to demand an explanation from the head of sales straight away: why is the team leaving customers without replies? But the head of sales brings correspondence in which the buyer set the date for resuming the conversation themselves. Another record contains an unrelated call. In a third, the salesperson really did miss the promised deadline. This is an illustrative example with fictional data; below, we will work through the numbers.

Who should you believe — the head of sales or the machine? I think the question itself draws the owner into an unnecessary argument. First, we need to find out which part of the sales team’s work the summary is based on and what exactly supports each alert. Otherwise, AI will speed up the review of mistakes the salespeople never made, while the real loss will remain in correspondence that the analysis never even reached.

Using AI to analyse sales

Start with a specific question about sales. For example: after sending a proposal, does the team continue the conversation with the customer or simply wait? The deal value and stage are not enough to answer this. You need the date the proposal was sent, what was agreed about continuing the conversation, the last contact and what happened afterwards. Ask the head of sales to collect these details for one part of the sales funnel — this will show what is missing before you can draw a conclusion.

In this kind of review, each source has its own role:

  • CRM shows who is handling the deal, its value and its current stage.
  • The conversation reveals the condition for continuing: the customer is waiting for a calculation, seeking approval for the purchase or has already declined.
  • Correspondence shows whether the salesperson sent what they promised and what the buyer replied after the call.
  • The activity history links the agreement to a deadline: when contact was due and whether it happened.

Useful discrepancies emerge where these sources meet. The record says “too expensive”, but in the conversation the buyer asks for a calculation for a different specification. The head of sales then needs to find out why a request became a refusal. Sending the salesperson to learn how to handle price objections before this review makes no sense: the customer is still waiting for a proposal that suits them.

Six alerts on 1.2 million roubles: how much has been lost?

Let us continue the illustrative example. There are twenty open deals in CRM after a proposal has been sent, each worth 200 thousand roubles. The total is 4 million roubles. Call recordings for the chosen period are available for twelve records; eight have no such recordings. So calls are available for 12 out of 20 deals — that is 60%. It is too soon to judge the whole sales team by them.

The initial analysis flagged six records: no follow-up after the proposal was found in the data reviewed. These deals total 6 × 200 thousand = 1.2 million roubles. The head of sales checked the alerts against the original conversations and correspondence:

  • In two deals, the salesperson promised to come back with an answer, the deadline has passed and there has been no follow-up. Contact needs to be restored; these deals total 400 thousand roubles.
  • In three, the buyers agreed a later date for the conversation. It has not arrived yet. Demanding an immediate call would break the agreement with the customer.
  • In one, a conversation about a different purchase is attached to the record. First, the link between the records needs to be corrected, then the deal reviewed again.

Of the six alerts, two were confirmed as missed actions. This gives grounds for assigning specific tasks to the salespeople. But 400 thousand roubles is still the value of open deals. It cannot be called lost revenue: the sales outcomes are unknown. Nor can it be called a future gain after taking action.

The eight deals without recordings are a separate matter. They total 1.6 million roubles, and the absence of alerts about the conversations says nothing about the quality of work with these customers. The owner needs two figures side by side: how many deals have been reviewed and how many remain without sufficient data. Otherwise, the quietest part of the report will be the part you know least about.

There is another trap when adding up money. A single deal can appear in the “no reply”, “overdue step” and “reason not recorded” lists at the same time. You cannot add up the amounts from the three lists: one purchase would be counted three times. Ask the head of sales to count each deal record once and separately show the signs that brought each one into the review.

A neural network for sales analysis

Choose a tool based on what materials it will receive and whether the head of sales can trace the basis for a conclusion. A summary table can also give you a polished explanation of falling sales. But that table cannot tell you what the buyer asked to change in the proposal or whether the salesperson did it.

For a one-off task with a prepared data export, a review of the table and attached texts is suitable. Here, the head of sales is responsible for what the export contains: the period, stages, dates and links between conversations and deals. For daily work, you need a way to get fresh data regularly from the relevant sources. When choosing a service, find out exactly this: which events it reads, for what period and where gaps remain.

Not every task needs a neural network. A standard CRM filter can count deals with no deadline recorded. AI is needed where meaning must be analysed: whether the buyer set a date, the salesperson suggested it themselves or no date was discussed at all. This distinction changes what the head of sales does — which is why it is worth making.

For the first review, give the head of sales this sample analysis task:

For the chosen period, review deals where a proposal has been sent. For each, state what the salesperson promised, what the customer agreed to, what deadline was given and what happened afterwards. Include the passage from the conversation or correspondence on which the conclusion is based, and identify the source. List missing materials separately. If no deadline was given, say so. Do not invent one from the usual deal duration. Separate missed actions from waiting as agreed with the buyer.

This is a task for a review, not a promise that every model will give the same answer. Before using the result, the head of sales checks the quoted passages against the original. A link to the call has a practical purpose: the head of sales must be able to find the relevant words quickly, without listening to everything again from the start and trying to guess where the conclusion came from.

The owner decides the limits of use here. If only records are available, we start by putting CRM in order. If records and conversations are linked, we review discrepancies between promises and actions. We do not ask the tool to assess an employee on the basis of materials that are absent from the analysis.

AI analysis of salespeople’s calls

An individual call is easy to misread if its purpose is unknown. The salesperson did not establish the customer’s needs — but the customer called to check the payment details for an order already agreed. No meeting was arranged in the conversation — but it was arranged yesterday in correspondence. Applying the same requirement to every call creates criticisms where the salesperson was dealing with a completely different issue.

So you review each conversation as part of its deal. Before assessing it, the head of sales needs to know the sales stage, the previous agreement and the purpose of the current contact. Requirements come from the company’s established working procedures. If a rule was never set, you cannot later declare that breaking it was the salesperson’s mistake.

Here is another illustrative example with invented dialogue. The buyer says: “I’ll agree it with my partner and get back to you.” The salesperson replies: “All right, I’ll wait.” A payment date appears in the record: Friday. AI is useful here as an assistant that compares the words with the entry: in the conversation, the customer did not promise to pay on Friday.

The head of sales needs to decide how to clarify the approval process: what the partner needs to make a decision, who will pass on the materials and when it is appropriate to get in touch. Simply telling the salesperson to call more often means increasing the number of contacts while leaving the condition for the purchase unknown. The date in the record also needs to match what was actually agreed.

One such recording gives material for a conversation with the salesperson. To draw a conclusion about their work, the head of sales looks for the same action repeated in comparable deals: for example, the employee records their own expectation as the customer’s promise every time. After explaining the requirement, the head of sales looks at new conversations at the same stage. If the behaviour changes, the salesperson has learnt to meet the requirement. If the mistake repeats, you need to find out whether it is a matter of skill or of following instructions.

A decision about an employee cannot be replaced by a general description such as “poor at negotiating”. The head of sales must name a specific action and show what they taught the employee. How to divide the work after finding a discrepancy is covered separately: AI has spotted that sales are stalling. Who will deal with it now?

How to tell whether the analysis is fit for use

Judging the tool only by the mistakes it finds inflates trust in it. If you look only at deals with alerts, you will not know what it missed. So include records where the analysis flagged no issues in the trial review. The head of sales selects them without prompts from the system; first, they review the original materials themselves, then compare conclusions.

For a small trial, you can take five deals with alerts and five without. This is a suggested number for manual checking, not an accuracy standard. This set is intended to uncover major discrepancies before daily use. It does not justify claiming a reliability percentage for the whole sales team.

The owner ultimately only needs answers to three questions. Do the conclusions have a source that can be opened? Are the limits of the available data visible? Does the head of sales have enough time to review what has been found? If every alert requires the entire history to be assembled again, the sales team will buy itself an extra backlog of investigations. Then you need to fix the links between records or narrow the analysis task first.

A task for the head of sales over five working days

You can pass the text below to the head of sales in full. The deadline is for getting a basis for a decision, not for promising sales growth.

By the end of the third working day, prepare a trial review of deals after a proposal has been sent. State the period and the total number of records. Separately, count how many have calls and correspondence available, where materials are missing and why.

>

Check five alerts and five deals with no issues flagged against the original materials; if a group has fewer than five, take all of them. For each discrepancy, state the cause: a missed action, a misunderstood agreement, an error in the data or missing materials. Add up the amounts without counting any deal twice.

>

By the fifth working day, bring me a decision: which analysis we will use daily, which source needs to be put in order first and which conclusions we will not use. State how long the manual check took. For confirmed missed actions, assign tasks to salespeople within your authority; bring me only matters that require my decision.

This task limits the time spent on the trial. Based on the result, the owner decides whether to expand the use of AI or restore the data first. Listening and comparing remain the head of sales’ job, while the owner gets an explanation of the choice with examples they can open.

When deals and conversations have to be gathered manually

If the head of sales is constantly looking for one agreement in CRM, another in a call and a third in correspondence, a one-off review quickly becomes outdated. RENTROP has AI ROP specifically for working with these sources — AI for the sales team: it analyses calls and correspondence alongside the deal record and shows every day which deals to bring through to payment, who to reply to and which deals to bring back. AI ROP only reads data; decisions and tasks remain with the head of sales.

The head of sales should still have the right to challenge a summary with six alerts — and the duty to show what the challenge is based on. When the argument is settled by a call recording, the owner no longer has to choose whom to believe based on their position or the confidence of their answer. Now they can demand an explanation for a specific missed step.