AI sales follow-up: connect your CRM before automating messages

Choose one follow-up moment

Start with a specific situation: an inbound inquiry that needs a first response, a quote awaiting an agreed follow-up, or a meeting that needs a recap. These should not all use the same rules. Decide which event starts the process, who owns the lead and what makes the task complete.

For a first version, I would choose follow-up after a sales meeting. There is usually a known contact, a conversation to summarize and a next step to confirm. Keep sending under human control while checking whether the system understands the context. Do not start with an agent free to message every old lead.

Match the person and gather current context

Before drafting, link the conversation to the correct contact and opportunity. A shared company name is not enough when several people work on different deals. If the match is uncertain, ask the account owner to choose the record. Attaching a good message to the wrong deal is still a failure.

The agent should receive the recent conversation, the current opportunity stage, the promised next step and any relevant approved quote. Include dates and links to the supporting records. If the meeting notes conflict with the CRM, flag the conflict instead of quietly choosing whichever source fits the draft.

Anthropic’s engineering guidance on tools for agents gives a useful example: assemble relevant customer context rather than make the agent sift through unrelated records. For a sales workflow, I would apply that idea to a short, clearly sourced brief—not a dump of the entire CRM.

• Who is the recipient, and which opportunity does this conversation concern?
• What was actually agreed, and when was the record last updated?
• Has someone already replied, closed the deal or decided to stop following up?

Draft the next step without inventing commitments

Consider this hypothetical case: a customer asks after a meeting whether delivery can happen next Thursday. The CRM contains an older estimate, but availability has not been confirmed. A useful draft acknowledges the question and asks the responsible person to confirm. It does not turn an old estimate into a new promise.

Ask the agent to produce three things: a short context summary, a proposed message and the questions still unresolved. Keep prices, discounts and delivery commitments tied to approved records. The salesperson should be able to see why each factual claim belongs in the message.

Human review needs to be convenient. Put the draft next to the supporting information, with options to edit, approve or return it. If checking a draft takes longer than writing it, the workflow has not earned its place yet.

Check again before sending

A draft can become stale while waiting for approval. Before it is sent, check whether the customer has replied, another teammate has contacted them or the opportunity has changed. Keep a record of the approved version and the send result. A failed delivery should not silently mark follow-up as completed.

Use a task identifier so a retry cannot create a second identical send. Make cancellation and stop conditions explicit. I would start with a person approving the recipient, channel and message; any later automatic sending needs a separate decision about permissions and the business’s communication rules.

Test the cases that make the team nervous

Run through representative examples before using live work. Include a missing customer match, two open deals for the same contact, an outdated quote, a new reply arriving during review and a delivery failure. The system should pause or escalate when appropriate, not manufacture a smooth-looking answer.

These are proposed test cases, not a report of a client implementation. Ask the salespeople what would make them distrust the system and add those cases too. The goal is useful support in the messy version of the workflow, not just a convincing demo on clean data.

Measure the time returned to salespeople

Count time spent finding context, preparing the draft, reviewing it, correcting errors and handling exceptions. Also track missed follow-ups, duplicate attempts and whether the promised next step was recorded correctly. More generated messages do not automatically mean better sales work.

My ambition is to remove at least half of suitable repetitive office work. In this workflow, that is a target to test against the full preparation and review burden, not a guarantee of higher conversion. Sales outcomes also depend on the offer, timing and relationship.

If you want to build this, start with one real follow-up example and the systems your team opens to handle it. Keep the relationship with the salesperson. Build the software around the information gathering and handoffs that keep taking their time.