How do I set up AI agents for a small B2B sales team without sending spam?
Start narrow and keep a person on the send. Define one ideal customer profile, authenticate your sending domain, let agents research accounts and draft emails, require a person to approve every first-touch email, cap daily volume to what that person can read carefully, and make opt-outs automatic and permanent. Loosen any of this only after a month of clean results. Spam comes from volume without review, not from AI itself.
This guide is for a founder or head of sales with a team of two to twenty. It covers the setup, the rules you cannot break, a 30-day rollout and how the main platforms handle review. We make Nagent; how our AI Sales Chief of Staff, SERA, and her team handle it is in its own section.
Why do AI sales agents end up sending spam?
Rarely because the writing is bad. Usually because of one of these:
The target list is too wide. An agent told to "find SaaS companies" will find thousands. Most of them never needed you.
Sending is switched to automatic on day one. The first 200 emails go out before anyone reads one.
The agent writes from the open web. It invents a pain point, a mutual connection or a product claim, and the reply is a complaint.
Opt-outs live in someone's inbox. A prospect says "remove me", nobody records it, and the next sequence step goes out anyway.
Volume grows faster than reputation. A new domain sends hundreds of emails a day, complaint rates rise and mailbox providers start filtering everything you send, including replies to real customers.
Every one of these is a setup choice. That is good news: you can make different ones.
What should you set up before any agent writes an email?
One ideal customer profile. Industry, company size, region, the buying role and one signal that says "now" (hiring for a role, a funding round, a new product line). Write it in a few sentences. Agents score against it.
Knowledge the agent may write from. What you sell, who you have sold it to (only customers who agreed to be named), your positioning and the claims you must never make. Without this, drafts guess.
An authenticated sending domain. SPF, DKIM and DMARC set up for the domain you send from. Many teams send outbound from a separate subdomain so a problem there does not touch their main domain.
A suppression list. Import every address that has ever asked not to be contacted, plus current customers and open opportunities, before the first send.
A clean CRM. Agents read your records. Duplicates and dead contacts become duplicate and dead emails.
A named approver. One person, with a deputy, who reads drafts before they go.
What are the email rules you cannot break?
These are external rules, not preferences.
Gmail's bulk sender rules (Google, 3 October 2023): senders of more than 5,000 messages a day to Gmail must strongly authenticate their email, offer one-click unsubscribe, process unsubscribes within two days and stay under a spam rate threshold.
Yahoo's bulk sender rules (Yahoo Sender Hub): SPF and DKIM, a DMARC policy of at least p=none, one-click unsubscribe, a spam complaint rate below 0.3% and unsubscribes honoured within two days.
CAN-SPAM in the US (15 U.S.C. 7704): accurate headers and subject lines, a valid physical postal address, a clear way to opt out, and no further commercial email more than 10 business days after an opt-out. The FTC's compliance guide adds that the law makes no exception for business-to-business email.
Other regions have their own rules, often stricter about consent for cold email. If you sell into the EU, the UK or India, check with counsel which basis you rely on before you scale.
An AI agent does not change any of this. It just makes it possible to break the rules faster.
Which jobs should AI agents do first, and which should wait?
Order matters. Give agents the jobs where a mistake stays inside your workspace first.
Start now: account research, finding the right contact, summarising why the account fits, drafting the first email for a person to approve, logging replies.
After two to four weeks of clean drafts: follow-up drafts, booking meetings from positive replies, nurture emails to people who already know you.
Only with evidence: sending without a person reading each email. Many small teams never need this; the time saved is small next to the risk.
How much volume is safe?
There is no universal number. Use review capacity as the ceiling: if your approver can read 25 drafts carefully in a 30-minute session, that is your daily cap until the data says otherwise.
Watch three numbers weekly: bounce rate, spam complaints and the share of replies that are negative or "remove me". If any rises, cut volume and look at targeting before you look at copy. Raise volume in steps, one step a week at most.
How should approval work so it does not slow the team?
Approve in batches, once or twice a day, not as notifications all day.
Approve the draft, not the idea. The person should see the exact email, recipient and why-now reason.
Reject with a reason. "Wrong persona" or "claim not supported" teaches the agent and the knowledge base.
Check suppression at the moment of sending, not at approval. A draft approved on Monday and sent on Friday must respect an opt-out that arrived on Wednesday.
How do the main platforms handle review before send?
From each vendor's own site, checked 24 October 2026:
HubSpot Prospecting Agent: two settings: review and approve each message before it goes out, or send without review based on your outreach settings. You set a maximum number of emails, minimum days between them and a send window (HubSpot knowledge base). It costs 100 HubSpot Credits per lead, which is $1.00 at $0.010 per credit (HubSpot Credits). Best for teams already on HubSpot.
Salesforce Agentforce Engagement (formerly Lead Nurturing): works new or updated leads and contacts automatically, sends intro emails, answers replies and connects qualified leads to reps; Salesforce help lists articles on reviewing and approving agent emails and on queuing them for manual review with a flow (Salesforce Help). Agentforce actions are 20 Flex Credits each, with Flex Credits at $500 per 100,000 (Agentforce pricing). Best for Salesforce-native teams with an admin.
Outreach: agents for research, personalisation and revenue work; admins set how autonomously agents operate across segments (Outreach AI agents). Plans are custom quotes with AI credits (Outreach pricing). Best for established sales teams running sequences at scale.
Gong Engage: Gong says it prioritises and personalises every sales engagement; pricing is per user plus a platform fee, by proposal (Gong pricing).
Salesloft, which now hosts Clari: clari.com redirects to salesloft.com, where Clari Forecast is listed beside Cadence and AI agents; no pricing is published.
The pattern: every platform can review before send, and most can also send without review. Which one you switch on is your decision, not theirs.
What does a 30-day rollout look like?
Week 1: profile, knowledge, domain authentication, suppression import. Agents research 50 accounts. No sending.
Week 2: agents draft first emails for the best 25. A person approves, edits or rejects each. Send only approved drafts.
Week 3: review the reasons for rejections. Fix the profile and knowledge, not just the drafts. Keep volume flat.
Week 4: compare replies, complaints and bounces with week 2. Raise volume one step only if all three are healthy.
Our guide to finding ICP-fit accounts and researching them with AI goes deeper on week 1.
How does Nagent's Sales team handle this?
Nagent is multiplayer AI for growth: people and AI agents in one shared context, delivering the number. On the Sales team, SERA is the AI Sales Chief of Staff, NORA runs outbound research and first outreach, VANI (Beta) writes NORA's emails and LIPI (Beta) writes to leads already in the CRM.
How the setup above maps to it:
Profile first. Nothing is discovered without an active customer profile; NORA scores companies against it and records why.
Writes from your material. VANI and LIPI write from your knowledge and offerings only; every product a draft names must come from your list of offerings.
Approval before send, whatever the level. NORA refuses to send a draft nobody approved and checks the approval again just before it leaves. Agents start at L2 Assist; the ladder runs L1 Observe to L5 Lead, and plan ceilings run from L3 Operate on Solo and Team to L5 Lead on Enterprise.
Unsubscribe built in. Every prospect email carries an unsubscribe link and the two one-click headers mail clients read. If the link cannot be made, the send is refused.
Suppression checked at send. Opt-outs, spam complaints and hard bounces go on the workspace's do-not-contact list, and the list is checked at the moment of sending. No agent can clear an opt-out.
Volume you can review. Campaigns set a maximum number of profiles per run, and the docs advise starting with a few a day.
What it does not do: it does not close deals or take meetings, which stay with your people. Deal execution (DEXA) and revenue operations (RIVA) are coming soon. Without a sending account your domain authorises, drafts approve and then fail when sent. Work is metered in credits; see pricing.
Frequently Asked Questions
Should AI agents send sales emails without human approval?
Not at the start, and for many small teams not ever. Approval costs a few minutes a day; one wrong email to a key account costs far more. Earn automation with a month or more of clean results.
How many cold emails a day can a small team send safely?
As many as one person can review carefully, then raise volume in small weekly steps while bounces, complaints and negative replies stay low. Yahoo, for one, asks bulk senders to keep complaints below 0.3%.
Does CAN-SPAM apply to B2B cold email?
Yes. The FTC's compliance guide says the law makes no exception for business-to-business email. Include a physical address and a working opt-out, and stop mailing anyone who opts out within 10 business days.
What are the best AI SDR agents for a small team?
Ones that research before they write, write only from your material and wait for approval. We compared options in top AI SDR agents for outbound prospecting.
See SERA's Sales team draft and hold outreach for approval: book a demo.