What to Check Before an AI SDR Sends Email for You

Should you let an AI SDR send email for you without checking first?

No. An AI SDR can research accounts and draft outreach faster than any person, but sending is where the risk lives. Before it sends a single email from your domain, check the list it can touch, the facts it can use, the rules your inbox provider enforces, and who reads the replies.

AI SDR tools are everywhere this year. Some are features inside platforms you already pay for, and some are standalone products that promise a full outbound motion with no humans involved. The pitch is always the same: more meetings, less work. The reality depends almost entirely on what you set up before the first send.

I build outbound and enrichment systems for B2B teams, and the failures I see are rarely about the model. They are about a bad list, a missing rule, or nobody watching the inbox. This is the checklist I would run before letting any AI SDR send on your behalf.

What is an AI SDR actually doing?

An AI SDR is software that does the work of a sales development rep: finding accounts, researching contacts, writing outreach, sending sequences, and handling early replies. Most combine a contact database, an AI writing step, and a sending tool. The quality of each part, and the rules between them, decide whether it books meetings or burns your domain.

It helps to split the job into three layers. The data layer decides who gets contacted. The writing layer decides what they read. The sending layer decides when and from where. A human SDR carries judgment across all three without thinking about it. Software needs that judgment written down as rules.

Most problems happen where the layers meet. The writing step invents a detail because the data step gave it too little. The sending step keeps going because nobody told it to stop when replies turn angry. When you evaluate a tool, look at those seams first.

Which accounts should it be allowed to contact?

Only accounts that pass your written ideal customer profile and your disqualification rules. An AI SDR will happily email anyone in its database. If your list is loose, it scales your mistakes. Lock the list before you touch the copy, and give the tool a short exclusion list it can never ignore.

Your exclusion list should include current customers, open deals, partners, past churned accounts you are not ready to win back, and competitors. Most CRMs can export these in minutes. The cost of skipping this step is an automated cold email landing in the inbox of a customer you signed last month.

I covered the logic in detail in the disqualification rules I write before any outbound starts. The short version: say who you will not contact before you decide who you will. It is the cheapest filter in your whole stack.

What should it know about an account before it writes?

At least one specific, verifiable reason this account might care right now, pulled from a source you can check. A hiring post, a product change, a new market, or a public statement all work. If the tool cannot point to the source of its personalization, treat that line as invented and remove it.

The biggest trust risk with AI-written outbound is confident fiction. A line like "congrats on the new funding round" reads well and costs you the account if it is wrong. Ask your tool to store the source URL next to each personal detail. If it cannot, limit personalization to facts from your own data, like industry and role.

I would rather send a short, honest, role-based email than a long one with a made-up compliment. Buyers can smell generic, but they punish false far harder.

What sending rules does your inbox provider enforce?

Google publishes clear rules. All senders to Gmail need SPF or DKIM authentication. Senders of more than 5,000 messages a day to Gmail need SPF, DKIM, and DMARC, plus one-click unsubscribe on marketing messages. Google also says to keep the spam rate shown in Postmaster Tools below 0.3 percent, and recommends staying under 0.10 percent.

These requirements have been in place since early 2024, according to Google's own sender guidelines, and they apply whether a person or a model wrote the email. An AI SDR that sends at volume can cross the bulk threshold faster than you expect, especially if it runs several sequences in parallel.

The spam rate number is the one I would watch most closely. A rate that sits near 0.3 percent is already a warning sign, not a target. If your tool cannot show you complaint rates per sending domain, you are flying blind. Set up your domains properly first; I wrote a full walkthrough on cold email domain setup before the first send.

How should you review the first drafts?

Read the first fifty drafts yourself before anything goes out, and mark each one as send, edit, or reject. Track the reasons. If more than a handful need edits for the same reason, fix the prompt or the data rather than editing one by one. Only turn on sending once most drafts are ones you would send unchanged.

Fifty is not a magic number. It is enough to see patterns without taking a week. You are looking for repeated problems: the same awkward opener, a claim about your product that is too strong, a call to action that asks for too much. Each pattern points to a fix upstream.

Keep the review log. When results dip in a month, that log tells you what the tool used to get wrong, so you know where to look first. It also gives you something concrete to show a co-founder who asks whether the AI is any good.

What should an AI SDR never say?

It should never invent facts about the prospect, never promise pricing or outcomes you have not approved, never claim a relationship that does not exist, and never pretend to be a named human who does not exist. Write these as hard rules in the tool's instructions and check for them in your draft review.

Pricing and outcomes are the quiet risk. A model trying to be persuasive will reach for numbers. If your instructions do not forbid it, you may find an email promising results your team cannot deliver. Give the tool an approved list of claims, with wording, and tell it to use nothing else.

On identity, I take a firm line. If the email comes from a named sender, that person should exist and should be reading replies. Buyers forgive automation. They do not forgive finding out the person they replied to was never there.

Who reads the replies, and how fast?

A named person on your team should own every reply, with a clear rule for which replies the tool may answer and which it must hand over. Positive replies, pricing questions, and anything angry should always go to a human. The tool can handle simple out-of-office and unsubscribe replies on its own.

This is where many setups break. The tool books a meeting, but nobody updates the CRM, or the rep does not see the thread until the next day. Map the handoff in writing: where the reply lands, who gets notified, and what gets logged in HubSpot or whatever CRM you use.

If you are also testing always-on agents like the ones I covered in my look at OpenAI's dots, the same principle holds. Let software draft and sort. Keep a person on anything that commits your company to something.

How do you know if the AI SDR is working?

Measure positive replies and qualified meetings per hundred contacts, not emails sent or open rates. Also track complaints and unsubscribes per domain. A tool that sends a lot and books little is not saving you work. It is spending your reputation on activity that looks good in a dashboard.

Compare against a baseline. If you had a person doing outbound before, use their numbers for the same segment. If you did not, run a small manual batch yourself for two weeks. Without a baseline, any number the tool reports will look fine.

After six years and more than 100 projects, the pattern I trust is simple. The teams that win with automated outbound treat it like a junior hire on probation. They review the work, give feedback, widen the scope slowly, and pull it back fast when something goes wrong.

What should you do next?

Write your exclusion list, confirm your domain authentication, and set a spam rate ceiling below Google's limit. Then run fifty drafts through manual review before turning sending on. If you do those four things, most AI SDR tools become safe enough to test, and you will know quickly whether yours is worth keeping.

If you want a second pair of eyes on your outbound setup before you switch an AI SDR on, reach out. I am always happy to look at how the pieces connect and where they might break.