AI Sales Agents: Research, Follow-Ups and CRM Hygiene

An AI sales agent earns its place on four jobs that surround the sale rather than make it: researching an account before first contact, drafting follow-ups after a call, preparing a one-page brief before a meeting, and keeping the CRM accurate. In each, the agent gathers and drafts and a person decides. Anything sent to a customer, any price and any commitment waits for a person’s approval. In the US, email the agent drafts is still covered by the CAN-SPAM Act, and calls made with an AI-generated voice fall under the Telephone Consumer Protection Act, as the FCC confirmed in February 2024. Measure the agent by the quality of its drafts and the accuracy of the CRM, not by how many messages it sends.

An “AI SDR” usually means an agent that prospects and sends outreach on its own. That is the version with the most risk and the least room for a person to catch mistakes. The jobs below work whether or not you ever go that far, and they fit a small team as well as a large one; for the wider list of jobs in a small business, see AI agents for small businesses.

Four jobs worth handing over

1. Account research

Before a first email or a discovery call, the agent reads the company’s website, recent news and filings, and your own CRM history, and writes a short brief: what they do, who is likely involved, what they have bought from you before, and one or two reasons they might care now. Ask it to cite each fact, because a confident wrong fact about a prospect is worse than none.

2. Follow-up drafts

After a call, the agent turns the notes into a follow-up email that recaps what was agreed, lists next steps and attaches what was promised. It drafts; the rep edits and sends. The draft also shows what the agent thought was agreed, which is a useful check on the notes.

3. Meeting prep

An hour before a meeting, the agent pulls together the last conversations, open deals and tickets, and anything that changed since the last contact. HubSpot’s own guide to using its MCP server↗ (opens in a new tab) lists this kind of work, from pipeline summaries to turning meeting notes into record updates, among the jobs it is meant for.

4. CRM hygiene

The agent finds contacts with no company, deals with no next step, duplicate records and stages that do not match the activity. It proposes fixes in a table; a person approves the batch. This is the job most teams underrate, because every other job depends on the CRM being right.

Connecting an assistant to the CRM is covered per system: HubSpot MCP for HubSpot and Salesforce MCP server for Salesforce. Both act as the signed-in user, so the agent can see and change what that user can.

What a human must approve

• Anything sent to a customer or prospect: emails, messages, sequence enrollments, calls and voicemails.
• Prices, discounts, quotes and payment terms.
• Commitments: delivery dates, features, service levels, contract terms, and anything that begins “we will”.
• Merges and deletions in the CRM, and bulk updates beyond a handful of records.
• Adding a new contact to outreach, and whether that person has agreed to be contacted the way you plan to.
• Anything the agent would say about a competitor.

Start with every item on that list approved one at a time. Loosen only where the drafts go out unchanged for weeks, and never for prices or commitments. The patterns for building that approval step are in AI agent approval workflows.

US rules in plain words

This is a plain-language summary, not legal advice. Rules differ by state and by industry; talk to counsel before an agent sends or calls anyone on its own.

Email: CAN-SPAM

The FTC’s CAN-SPAM compliance guide↗ (opens in a new tab) says the law covers all commercial email, not just bulk mail, and makes no exception for business-to-business email. Its main requirements, which apply whether a person or an agent wrote the message:

• No false or misleading header information, and no deceptive subject lines.
• Identify the message as an ad, clearly and conspicuously.
• Include your valid physical postal address.
• Explain clearly how to opt out, and honor opt-out requests within 10 business days.
• Monitor what others do on your behalf: hiring a vendor, or an agent, does not move the responsibility off you.

For an agent, the practical rules are: it drafts from a template that already has the address and the opt-out, it never writes subject lines that overpromise, and it checks the opt-out list before any send. The FTC’s rule text is in 16 CFR Part 316↗ (opens in a new tab).

Calls: the TCPA and AI voices

The FCC’s February 2024 declaratory ruling↗ (opens in a new tab) confirmed that the TCPA’s restrictions on “artificial or prerecorded voice” cover AI technologies that generate human voices, including voice cloning. In plain words:

• A call that uses an AI-generated voice needs the called party’s prior express consent, absent an emergency purpose or an exemption.
• If the call carries an advertisement or telemarketing, the FCC’s rules require prior express written consent.
• Artificial voice messages must identify the business responsible for the call, and telemarketing messages must offer a way to opt out.

The statute itself is 47 U.S.C. 227↗ (opens in a new tab). For most teams the simple answer is that an agent prepares call lists and talking points, and a person places the call.

Measuring an AI sales agent

• Draft acceptance: how many follow-ups go out with light edits, and what the heavy edits have in common.
• Research accuracy: a weekly spot check of facts in the briefs against their sources.
• CRM accuracy: a sample of records checked by a person each week, before and after the agent’s cleanup.
• Rep time: hours a week spent on prep and admin, measured before launch and after.
• Opt-outs and complaints: if these rise, the agent is sending the wrong things to the wrong people, whatever the reply rate says.
• Pipeline outcomes, such as meetings held and deals progressed, compared over a quarter rather than a week.

Reply rate alone is a poor measure. An agent can raise it with subject lines a person would never approve.

A board for the follow-ups that need a person

Once an agent is drafting, the bottleneck is the approvals. Keep dated follow-ups where they already live, as tasks in the CRM against the contact or deal, because fenbs has no due dates and no assignee you can set. Use a fenbs board for the work that needs a decision rather than a date: “approve the discount for the renewal”, “review 40 proposed contact merges”, “confirm the new outreach template”. The assistant files each as a task with the details and the CRM link in the note, a priority from 1 to 10, and a plan; a person moves it through To Do, Next Up, In Progress and Completed. Each change is recorded in History under the assistant’s name, and a standing rule such as “no emails sent without a person” goes on the Decisions and rules page, which every connected assistant reads before it starts.