Most sales follow-up fails quietly. A lead asks for a proposal, the salesperson is busy, and ten days later nobody has written back. An AI agent is a natural fix: it does not forget, it does not get busy, and it can answer at 7 a.m. on a Sunday.
The same strengths cause the risks. An agent that never gets tired can also send the wrong promise to forty people before anyone notices. This article describes how a small or mid-sized company can set up a follow-up agent so that it helps salespeople instead of replacing their judgement. It is a practical format, not legal advice.
Start with the job, not the tool
"Follow-up" covers very different tasks. Write down which of them you want the agent to do. A typical split looks like this:
- Safe to automate: confirming that a form or enquiry was received, sending a reminder to someone who has not answered, proposing meeting times, sharing a published brochure or case study, answering factual questions from approved material (opening hours, supported integrations, how onboarding works).
- Draft only, a person sends: a proposal cover note, a reply to a detailed technical question, a message to a large or strategic account, anything that mentions a discount.
- Never by the agent: quoting a price outside the published list, promising a delivery date, agreeing to contract terms, commenting on a competitor, replying to a complaint, or contacting someone who asked not to be contacted.
The middle category is where most companies should begin. A draft in the salesperson's inbox, one click from sending, already saves most of the time and lets you see how good the drafts are before you let any of them go out unchecked.
Write the guardrails as a list
A prompt that says "be polite and professional" does not protect you. A list of concrete rules does, because you can test it. For a follow-up agent, we would expect at least these:
- Sources. The agent may state only facts found in an approved set of documents: price list, product pages, onboarding guide, FAQ. If the answer is not there, it says so and hands over.
- No commitments. No prices, discounts, deadlines, guarantees or "we can definitely do that". It may say "I will ask a colleague to confirm" and then really do it.
- Identity. The agent does not pretend to be a named human. A short line such as "I am the assistant of the sales team at..." is enough. Our article on telling customers when an AI agent is involved covers wording and timing.
- Frequency. A fixed maximum number of messages per contact, a minimum gap between them, and a final stop. Three or four touches over a few weeks is plenty for most B2B leads. After that, the agent writes a CRM note and stops.
- Opt-out. Any request to stop, in any wording, ends all automated messages to that person immediately and is recorded. Marketing and email rules differ by country and by whether the recipient is a business or a private person, so check what applies to you. The US CAN-SPAM guide and the UK ICO direct marketing guidance are two examples of what regulators expect.
- Personal data. The agent uses only the contact data it needs for the message. See data handling and retention for AI agents for what to keep and for how long.
Keep the list to one page. If it takes ten pages, nobody will review it, and the agent will not follow it reliably either.
Decide the handover triggers before launch
The most important design decision is when the agent stops talking. Our general piece on when to stop and pass the case to a person applies here, with sales-specific triggers:
- The buyer asks about price, terms, legal matters or a custom scope. These are negotiation, not follow-up.
- A buying signal. A reply such as "yes, send the contract" or "can we meet the whole team on Thursday?" is exactly the moment when a human should take over, quickly.
- A senior or unusual contact. A reply from a managing director, a procurement officer or a legal department deserves a person.
- Negative emotion. Irritation, a complaint, sarcasm, or "why do you keep writing to me?"
- Confusion. The agent cannot tell what the customer wants, or the customer asks the same question twice because the first answer did not help.
- Anything unusual in the message, such as attachments, a different language than expected, or instructions that look like an attempt to steer the agent.
Handover must be fast and visible. Set a response time, for example one working hour for a buying signal, and decide who is on duty when the owner is away. A handover to an inbox nobody reads is worse than no agent, because the customer has already been told someone will reply.
When the agent hands over, it should tell the customer in one sentence what will happen: "I have passed your question about pricing to Anna from our sales team; you can expect a reply by tomorrow noon." Then it must not send anything else on that thread unless the salesperson releases it.
What to write into the CRM
The CRM is where the agent's work becomes accountable. If the salesperson cannot see what the agent did, the customer will eventually tell them, usually at a bad moment. After every action the agent should write a short note with the same fields:
- What was sent or received, with a link to the full message.
- Why: which rule or step in the sequence triggered it.
- What the customer said, in one line, including any stated deadline, budget or objection, quoted rather than interpreted.
- Status: continuing automated follow-up, waiting for the customer, handed over to a named person, or stopped, with the reason.
- Next step and date: what happens next and who does it.
Two further points matter in practice. First, the agent should write only to the fields it needs. It does not need permission to change the deal value, the deal stage or the owner, and it should not have it. Our article on least-privilege access explains why. Second, separate facts from guesses. "Customer wrote that budget is approved for Q1" is a fact. "Customer seems very interested" is the agent's opinion, and it should be labelled that way or left out. A CRM that fills up with confident guesses becomes useless for forecasting.
Review and measure
Do not judge the agent by the number of messages sent. Look at what a salesperson would look at:
- How many leads got a reply within the agreed time, compared with before?
- How many meetings were booked, and how many of them were real opportunities?
- How many messages were edited or rejected by a person, and why?
- How many opt-outs, complaints or "please stop" replies came in?
- How often was handover late, or missed?
For the first month, have a salesperson read a sample of messages every week. Keep a log of what the agent sent so that you can answer a question later; see what to record in an agent activity log. If the edit rate stays high, tighten the rules or the source documents rather than adding more prompt text. When a message goes wrong, treat it like any other incident: a post-mortem of a customer communication mistake shows a simple way to do it.
A sensible rollout is gradual: drafts only for two to four weeks, then automatic sending for the safe category, then more, as the numbers justify. Having a fast way to pause the whole sequence, available to more than one person, is not optional.
A short checklist
- The allowed, draft-only and forbidden actions are written down on one page.
- The agent answers only from approved documents and makes no commitments.
- Message limits and a final stop are set, and opt-outs work immediately.
- Handover triggers, response times and a backup person are defined.
- Each action leaves a CRM note with what, why, status and next step.
- A person reviews a sample every week, and anyone on the team can pause the agent.
Conclusion
A follow-up agent is worth having when it makes sure no lead is forgotten and every customer gets a quick, accurate and honest reply. It is not worth having if it negotiates, improvises or hides what it did. Give it the repetitive part, write down where it must stop, and make sure that when a real buyer appears, a real person is already on their way.
