An AI assistant for sales follow-up that prepares the next move.

An AI assistant for sales should prepare follow-up from the whole conversation. It needs the recipient, company, current opportunity, last meaningful exchange, unanswered question, promised materials, approved offer, pricing source, opt-out state, and the action that would genuinely move the work forward. It should return one draft with the reason for sending, evidence for every factual claim, the approval status, and the planned next step.

Keep external sends behind a clear operating rule. Routine messages can move quickly when the recipient, content, timing, channel, and authority have already been approved. New outreach, discounts, legal claims, sensitive data, unusual commitments, and uncertain recipients need human review. NIST's AI Risk Management Framework calls for documented roles and human oversight appropriate to the context. In practice, that means the approval boundary belongs in the workflow, not only in the prompt.

Do not tell the system to follow up when it feels appropriate. Tie the job to evidence in the workflow: a recipient asked for pricing and did not receive it, a proposal reached its agreed review date, a meeting ended with a promised document, an invoice needs a status check, or a qualified opportunity has been waiting longer than the approved service level. The trigger should identify the record, the expected next action, and the condition that cancels it.

Cancellation conditions prevent embarrassing automation. Stop when the recipient replies, books, opts out, declines, changes the date, or the opportunity closes. Stop when another person on the team sends the answer. Deduplicate by recipient, channel, opportunity, and intended outcome immediately before sending. The final gate should read the live thread again, because the state may have changed after the draft was prepared.

Timing should come from the relationship, not a generic sequence. A promised Tuesday document belongs on Tuesday. A proposal review date belongs after that date. A buyer who said next quarter should not receive a three-day nudge. Record the source of the date so a reviewer can see whether it came from the recipient, a contract, the CRM, or an internal guess.

The sender remains responsible for the message. In the United States, the FTC says CAN-SPAM applies to commercial email, including business-to-business email. It requires accurate header information, non-deceptive subject lines, clear sender identification and address information, an opt-out method, prompt opt-out handling, and monitoring of vendors acting on the sender's behalf. An AI assistant does not transfer that responsibility.

Rules differ by recipient, channel, and location. The UK Information Commissioner's Office explains that electronic marketing can involve both PECR and data-protection law, and that business-to-business treatment differs between corporate subscribers, sole traders, and some partnerships. Its guidance also requires a lawful basis where personal data is used and makes clear that people can object to direct marketing. Determine the rule before the assistant prepares or sends the message.

For personal data covered by the GDPR, Article 5 requires lawful, fair, transparent processing, purpose limitation, data minimization, accuracy, and storage limits. The workflow should expose only the information needed for this follow-up, keep it current, and preserve the suppression record when someone opts out. Do not paste an entire contact database into a prompt because one message needs a job title.

Reply rate alone rewards noise. Track valid replies, qualified next steps, kept meetings, time to answer, and opportunities moved with buyer agreement. Put those beside approval rate, correction rate, opt-outs, complaints, duplicate attempts, wrong-recipient blocks, unsupported claims, and messages without complete receipts. A higher reply count is not a win if it creates more complaints or consumes more review time.

Review failed and edited drafts weekly. Label the cause: missing context, stale commercial fact, wrong tone, unclear trigger, weak next step, compliance risk, identity mismatch, or tool failure. Fix the shared source or rule before changing the prompt. If a new price repeatedly causes corrections, update the price source once instead of teaching every message around stale data.

NIST recommends ongoing monitoring and periodic review with clear roles. Use that principle as a practical operating loop: keep examples of accepted and rejected drafts, rerun them after a model or policy change, and tighten the autonomous lane only when the measured error pattern supports it.

A buyer asked after a product review whether the workflow could preserve approval before customer messages. The account owner promised a written diagram on Tuesday. The trigger is not three days without a reply. It is Tuesday arriving with a promised item still unsent. The thread record shows the buyer's exact concern, the approved architecture note, no opt-out, and no later human response.

The assistant prepares one short message that reconnects to the review, attaches the approved diagram, explains in one line where the approval gate sits, and asks whether the buyer wants the same rule mapped onto their current channel. It does not add pricing, a meeting link, a case study, or a second ask. Because an attachment and product-control claim are involved, the named owner approves the final version.

After the provider accepts the send, the workflow stores the immutable message identifier and reads the thread back once. The CRM records the promised item as delivered and waits. If the buyer replies, the scheduled follow-up is cancelled. If the provider call fails before acceptance, the system reports the error and does not guess that the message arrived.