AI SDRs and cold email: why autopilot outbound burns your domain

In this article I will explain why fully automated AI SDRs tend to damage the domain they send from, what Google and Yahoo actually require from senders, and the setup I use instead so that AI writes the drafts and a person decides what goes out.

To begin with, the problem. In 2026 anyone can connect an AI agent to a mailbox, give it a list of 5,000 contacts and let it write and send a "personalised" email to each one. The software makes writing free, so volume goes up. However, the cost of outbound was never really the writing. The cost is what happens to your domain when enough of those people press "report spam", and that cost does not go down when the writing gets cheaper.

How many spam complaints can a domain take?

Very few. Yahoo's Sender Hub tells bulk senders: "Keep your spam rate below 0.3%." That is three complaints for every thousand messages that reach the inbox. Yahoo also requires bulk senders to "honor unsubscribes within 2 days."

Google set out its own rules in an October 2023 Gmail post by Neil Kumaran, Group Product Manager for Gmail Security and Trust. From February 2024, anyone sending "more than 5,000 messages to Gmail addresses in one day" must authenticate their email, offer an unsubscribe that "should take one click", process unsubscription requests "within two days", and stay under a spam rate threshold that Gmail enforces. Google does not give the number in that post. In the same post it says Gmail already blocks "nearly 15 billion unwanted emails every day", which tells you how little patience the filter has.

Now put that next to how buyers feel about cold outreach. Gartner surveyed 632 B2B buyers and found that 73% actively avoid suppliers who send irrelevant outreach. If even a small fraction of that 73% reach for the spam button instead of the delete key, an autopilot campaign of a few thousand emails will cross three in a thousand without much effort. Thus, the limit on outbound is set by the recipients, not by the sender.

What is an AI SDR?

An AI SDR is software that does the job of a sales development representative on its own: it finds prospects, writes cold emails, sends them on a schedule, and often answers the replies, with little or no human review of each message. In simple words, it is outbound on autopilot. You set the audience and the goal, and the software decides what each person receives.

It is important to note the contrast with AI-assisted outbound. There, the AI does the research and writes the draft, but a person reads every message and decides whether it goes. The writing is just as fast. The difference is that nothing reaches a stranger's inbox without someone who understands the business having looked at it first.

Why autopilot outbound burns domains

1. Free writing removes the natural brake on volume

When a founder writes each email by hand, they send perhaps twenty a day, because that is all the time allows. The time cost was acting as a quality filter: nobody spends ten minutes writing to a company they do not believe in. Remove the cost and the filter goes with it. The list grows from the fifty companies you would take a meeting with to the five thousand that match a filter in a database.

2. The maths of reply rates works against volume

Instantly's 2026 cold email benchmark, built from its own platform data between 1 January and 18 December 2025, puts the average reply rate at 3.43%. In other words, the average campaign gets no reply from roughly 96 in every 100 people it emails. Every one of those 96 is a person who could report the message. Sending ten times more email at the same quality does not produce ten times more meetings without also producing ten times more chances to be reported.

3. One mistake is copied to everyone at once

When an AI agent misreads a company's website, gets a job title wrong or invents a reason to write, a human SDR would make that mistake once. An autopilot makes it in every message built from the same bad assumption, and sends them all before anyone notices. Moreover, the reader cannot tell the difference between a careless person and a careless machine. Both read as spam.

4. Your domain is shared with everything else you send

The domain that sends cold email is usually the same one that sends invoices, password resets and replies to your customers. If its reputation drops, it is not only the cold emails that land in spam. The usual workaround is to buy lookalike domains for outbound, which protects the main domain but adds a second problem: the prospect now receives email from a domain they have never seen, which looks exactly like spam.

The constraint has moved

Outbound is no longer limited by how many emails you can write. It is limited by how many complaints your domain can absorb, and a person approving each message is the cheapest deliverability control there is.

Everything below follows from that. The goal is not to stop using AI for outbound. It is to use AI for the part that was slow, the research and the first draft, and keep a person on the part that protects the domain, the decision to send.

How to set up AI-assisted outbound that keeps your domain safe

These are the steps I follow, in order. The DNS examples use a placeholder domain, example.com; replace it with your own.

1. Authenticate the domain before the first send. Publish an SPF record listing the services allowed to send for you, turn on DKIM signing in Google Workspace or Microsoft 365, and publish a DMARC record. A sensible starting DMARC record is a TXT record at _dmarc.example.com with the value v=DMARC1; p=none; rua=mailto:dmarc@example.com, which reports failures without blocking anything while you check that everything legitimate passes. Move to p=quarantine once the reports are clean.
2. Watch your reputation where the mailbox providers report it. Google's Postmaster Tools shows the spam rate Gmail measures for your domain. Check it weekly. If it moves towards Yahoo's 0.3% line, stop sending and find out why before you send anything else.
3. Send from a real mailbox, at a human pace. Send each message from the founder's or salesperson's own Gmail or Outlook, one at a time, at the volume a person could plausibly write. A considered email that one person sends twenty times a day is nowhere near Google's 5,000-a-day bulk line, and it reads as what it is: a person writing to a person.
4. Give every message a reason that can be checked. The first line should cite something real and dated: a job post, a new leader, a funding round, a launch or a pricing change. I explain which of these to trust in B2B buying signals: which ones are worth acting on. If there is no such reason, do not send the email.
5. Make opting out easy, even below the bulk threshold. For marketing mail sent in bulk, the one-click standard is RFC 8058, which requires the header List-Unsubscribe-Post: List-Unsubscribe=One-Click. For 1-to-1 cold email, a plain line such as "If this is not relevant, reply 'no' and I will not write again" does the same job, and honouring it within two days matches what Google and Yahoo ask of bulk senders.
6. Approve every message before it sends. Read the draft as the recipient would. Check the fact, the name, the company and the reason. If any one of them is wrong or weak, edit it or delete it.

What to check before approving an AI-written email

Approval only works if it is more than a click. This is the checklist I use, and it takes under a minute per message:

1. Is the first line a fact about them, with a date? Not a compliment, not a question about their goals, not a line about you.
2. Is it the right person? The person who owns the problem now, not the most senior name in the database.
3. Would the message make sense if they forwarded it to a colleague? If it only works because it is vague, it is not specific enough.
4. Is there exactly one question, and is it about their work?
5. Is there anything the AI could have invented? Numbers, customer names, claims about their stack. If you cannot verify it in ten seconds, cut it.

The first email matters most. Instantly found that 58% of all replies come from step one of a sequence, so the follow-ups cannot rescue a weak opening. InstinctGTM has a free cold email grader that scores a draft against checks like these before you send it, along with an AI slop checker for the phrases that make a message read as machine-written.

Where InstinctGTM fits

I built InstinctGTM on the rule in bold above. Its cold email capability drafts one message at a time, to a named person, citing a real change at their company with its date and source, and sends it from your own Gmail or Outlook once you approve it. It never sends on its own, on any plan, and that rule is enforced in the database rather than only in the interface. It also never drives your logged-in session on LinkedIn or any other platform.

InstinctGTM is live and in the revenue stage, and its paying customers include a $151 million company. Plans start at $99 a month for one venture, and the free plan writes your go-to-market plan before you pay anything. The full list is on the pricing page.

Habits that get a domain flagged

Sending to the whole database. If you would not take a meeting with the company, do not email it. Every irrelevant send is a chance to be reported.

Ramping volume on a new domain. A domain with no sending history that suddenly sends hundreds of messages a day looks exactly like a spammer, because that is what spammers do.

"Personalisation" that is only a merge field. Inserting a first name and a company name into a template is not personalisation. Readers have seen it a thousand times.

Letting the AI reply on your behalf. A prospect who replies with a real question and receives an automated answer that misses the point is more likely to report you than one who never replied at all.

Hiding the opt-out. A reader who cannot find a way to stop your emails will use the one button they can always find, which is "report spam".

What approval does not solve

Approval is slow by design. A founder approving every message will send dozens a day, not thousands, and some teams genuinely need more volume than that. For them, this approach caps growth, and the honest trade-off is a separate sending domain with its own reputation risk.

Approval also does not fix a bad list. A perfectly written email to a company that will never buy from you is still irrelevant outreach. The list and the reason to write matter more than the wording.

It is important to note what I have not verified myself. Google's 2023 post does not state Gmail's spam rate threshold as a number, so the 0.3% figure in this article is Yahoo's, from Yahoo's own page. Instantly's benchmark is a vendor's data about its own users, and it does not separate AI-written email from human-written email. Lastly, I have not published a measured comparison of complaint rates between autopilot and approved outbound, and I have not found an independent study that runs one. The argument here rests on the providers' published rules, on Gartner's survey and on arithmetic.

Frequently asked questions

What is an AI SDR? An AI SDR is software that does the work of a sales development representative on its own, finding prospects, writing cold emails, sending them on a schedule and often answering replies, with little or no human review of each message. AI-assisted outbound is different, because a person reads and approves every message before it sends.

Do AI SDRs hurt email deliverability? They can, because they make it cheap to send a large volume of messages that many recipients find irrelevant, and irrelevant messages get reported as spam. Yahoo asks bulk senders to keep their spam rate below 0.3%, and Google enforces a spam rate threshold for anyone sending more than 5,000 messages a day to Gmail addresses.

What spam complaint rate is too high for cold email? Yahoo's Sender Hub tells bulk senders to keep their spam rate below 0.3%, which is three complaints for every thousand messages delivered to the inbox. Google's 2023 announcement says Gmail enforces a clear spam rate threshold but does not give the number in that post, so treat 0.3% as a ceiling, not a target.

What is a good cold email reply rate? Instantly's 2026 benchmark, built on its own platform data from 2025, puts the average reply rate at 3.43%, the top quarter of senders at 5.5% or more, and the top 10% at 10.7% or more. It also found that 58% of replies come from the first email in a sequence.

Do I need SPF, DKIM and DMARC for cold email? Yes. Google's 2023 Gmail rules require bulk senders to authenticate their email, and SPF, DKIM and DMARC are the three standards that do it. Even below the bulk threshold, authentication is what lets a mailbox provider confirm the mail is really from you, so set all three up before you send a single cold email.

Is it better to have a human approve AI-written cold emails? In my view, yes. A person reading each message before it sends catches wrong facts, weak reasons to write and the wrong recipient, which are the things that produce spam complaints. It caps volume, which is the point, and it keeps a single mistake from reaching hundreds of people at once.

Check your domain before your next send

Open Google Postmaster Tools, add your sending domain, and look at the spam rate it reports. If you do not have a DMARC record yet, publish the p=none record above today. Then write tomorrow's cold emails the slow way, or let InstinctGTM draft them and approve each one yourself.