Before a meeting, a good salesperson does their homework: they reread the account history in the CRM, browse the prospect’s website, check the company’s latest news, find previous exchanges, prepare a few questions. It is valuable work, but it takes time, and it is often rushed when the calendar fills up.

It is an excellent use case for an AI agent: the task requires combining several sources, the result is easy to review, and the salesperson stays in control of everything that matters. Here is how we design this kind of agent, what it does, and above all what it does not do.

What the agent does

  • It gathers internal context: account and contact records in the CRM, open or lost opportunities, recent email exchanges, quotes sent, any support tickets if the prospect is already a customer.
  • It adds public information: the company’s website, recent press releases, published job openings, annual reports where they exist. Only professional, public sources about the company.
  • It writes a meeting brief: one page, always structured the same way. Who the company is, where the relationship stands, what has changed recently, the likely topics of the meeting, and a few questions to ask.
  • It suggests qualification elements: identified need, people involved, timeline mentioned, points still unknown. These are suggestions, presented as such, with the source of each piece of information.
The agent prepares the ground. The salesperson remains the one who decides, writes and talks to the customer.

What the agent does not do

This is the most important part of the design. A poorly scoped sales agent can do a lot of damage, both to the customer relationship and to compliance.

  • It sends no messages: no email, no follow-up, no invitation. It can suggest a draft, which the salesperson reviews, edits and sends themselves.
  • It does not change the CRM without approval: after the meeting, it can suggest an update (meeting notes, next step, qualification fields), but nothing is saved until the salesperson has approved it.
  • It does not look for personal information: no research into contacts’ private lives, no collecting personal addresses, no profile aggregation. Its scope is the company and the professional relationship.
  • It does not decide the final score: it suggests, the salesperson decides. A lead must not be dismissed by a machine without human review.

The personal data question

A prospect is a company, but above all it is people. Their names, job titles and business contact details are personal data under the GDPR. This calls for a few simple rules: processing generally based on legitimate interest and limited to the business relationship, minimizing the information collected, informing people in line with your privacy policy, and a defined retention period.

On the technical side, data sent to the model goes through contracted business access, and meeting briefs are stored in the CRM, not in an uncontrolled side tool. The agent’s account can only access the accounts in the requesting salesperson’s portfolio.

Sources and reliability

A wrong meeting brief is worse than no brief: it builds confidence in the wrong place. Two rules therefore always apply. Every piece of information comes with its source, internal or public, so the salesperson can check it in one click. And the agent clearly separates what it found from what it infers: “the company is hiring three data developers (openings posted last month)” does not have the same status as “the company seems to be building a data team”.

When information cannot be found, the brief says so. That is often the most useful part: it tells the salesperson which questions to ask in the meeting.

The flow, from meeting to CRM

  • The day before the meeting: the agent is triggered automatically from the calendar, or on the salesperson’s request. The brief lands in the CRM and a notification is sent.
  • Before the meeting: the salesperson reads the brief in a few minutes, corrects it if needed, and adds their own notes.
  • After the meeting: the salesperson dictates or types a few raw notes. The agent turns them into structured meeting notes and a proposed CRM update.
  • Approval: the salesperson approves, edits or rejects each proposal. Only approved changes are saved.

A concrete example

An SME selling software solutions to industrial companies books around twenty meetings a week for its sales team. Before the project, preparation depended on each person: very thorough for some, almost nonexistent for others, especially at the end of the week.

The deployed agent reads the CRM and the mailbox in read-only mode, checks public sources about the company, and produces a standard brief the day before each meeting. After the meeting, it suggests meeting notes and an update to the qualification fields, which the salesperson approves in a few clicks. The team arrives better prepared, CRM records are finally up to date, and the sales manager has reliable information for pipeline reviews. No message has ever gone out without a salesperson having written or reviewed it.

Our recommendation

Start with meeting preparation rather than automated prospecting: the gain is immediate, the risk is low, and the result is easy to assess. Define from the start what the agent will never do, in particular sending messages and changing the CRM without approval. Then measure preparation time saved and CRM data quality. If both improve, you will have a solid foundation to go further.