AI cold email deliverability, step by step: the Gmail, Yahoo and Outlook rules, safe sends per inbox, and the stop rules an AI SDR must enforce itself.
Yes, AI-sent cold email reaches the inbox when the sending setup is healthy. Gmail, Yahoo and Outlook judge authentication, volume and complaints, not who wrote the message. An AI SDR needs separate authenticated domains, slow warm-up, a daily cap per inbox (roughly 30 to 50 cold sends) and automatic pauses well before Google's 0.3% spam-rate limit.
An AI agent can research, write and send in an hour what a sales rep sends in a month. That speed is the reason to use one. It's also the quickest way we know to burn a sending domain. Mailbox providers don't care who wrote the email. They care whether the sender is authenticated, sends at a sane pace and gets few complaints. Here's the setup we'd put in place before an agent sends anything, with the official numbers and the maths behind them.
No. Gmail, Yahoo and Outlook publish what they check, and it's authentication, DNS, sending volume, unsubscribe handling and spam complaints. None of those rules mention AI. An AI-written email from a healthy, well-paced setup lands like any careful sender's. A human-written blast to a scraped list lands in spam.
Deliverability testers say the same. GlockApps puts it plainly: AI-generated content doesn't hurt deliverability on its own. Where AI hurts is indirect. It makes it cheap to send thousands of near-identical, poorly targeted emails, and recipients respond by ignoring or reporting them. Those reports are what filters act on.
All three want authenticated mail and low complaint rates. Google and Yahoo apply basic rules to every sender and stricter ones above roughly 5,000 messages a day. Outlook's 2025 rules apply to domains sending over 5,000 a day. Here is what the official pages say, checked on 17 October 2026.
| Provider | Every sender | High-volume senders | Spam complaint limit |
|---|---|---|---|
| Gmail | SPF or DKIM, valid forward and reverse DNS, TLS, RFC 5322 formatting | Over 5,000 a day to personal Gmail: SPF and DKIM, DMARC (p=none is fine), From domain aligned, one-click unsubscribe for marketing mail | Below 0.3%; Google's FAQ says aim below 0.1% |
| Yahoo | SPF or DKIM, valid DNS for sending IPs, RFC 5321 and 5322 | SPF and DKIM, DMARC at least p=none, one-click unsubscribe, honour unsubscribes within 2 days | Below 0.3%, measured on mail delivered to the inbox |
| Outlook.com | Best practice: valid reply address, clear unsubscribe, list hygiene | Over 5,000 a day: SPF, DKIM and DMARC must pass; non-compliant mail goes to Junk first and may later be rejected | No published number |
Two details from Google's sender FAQ matter for agents. The 5,000 count is summed across a whole primary domain, so mail from sales.northwind.com and northwind.com adds up. And once a domain crosses it, bulk sender status is permanent. An agent that spikes volume on one domain for a single day can lock that domain into the stricter rules for good.
You need four things in place: sending domains that aren't your main domain, DNS records that pass authentication, a verified contact list, and someone who owns the go/no-go decision. Skip any of them and the agent's speed works against you from the first hour.
That last item is the one teams forget. An AI SDR should draft and queue, but a person should approve each campaign and every restart after a pause. We've written about how a human approval loop fits an AI SDR in more detail, and our AI SDR overview shows where approvals sit in the workflow.
Send from a few lookalike domains, never your main one. Your main domain carries invoices, password resets and customer support, and a reputation hit there costs far more than a lost campaign. Register domains such as getnorthwind.com or trynorthwind.co, point them at your real site, and authenticate each one.
p=none, which both Google and Yahoo accept.A missing or broken record is the most common reason a brand-new campaign goes to spam on day one. Check all three with a test send to a Gmail address and read the "Show original" headers before the agent touches a real prospect.
No faster than a careful person would. Start with a handful of emails a day mixed with warm-up traffic that gets opened and replied to, and raise volume over three to four weeks. UserGems' help centre suggests ramping to 40 to 50 a day and waiting for inbox placement above 90% before a mailbox joins outbound.
The agent-specific risk is impatience in code. A rep warming one inbox gets bored and waits. An agent given "send 1,000 today" will try to find capacity. So the ramp has to be a schedule the sender can't override, not a suggestion in a prompt. Outbound Labs plans come with warmed mailboxes (5 on Growth, 25 on Scale); see how mailbox warm-up works.
Plan on 30 to 50 cold sends per warmed inbox per day and grow by adding inboxes. Published vendor numbers disagree wildly, though, and it's worth knowing why. Instantly's AI SDR guide sets a recommended cap of 30 emails per inbox per day. UserGems says to cap AI-sent email at 200 per SDR mailbox per day, while enrolling 50 new prospects a day.
Those numbers are closer than they look, because they count different things. Run the maths on a four-step sequence:
So "200 a day" and "50 new a day" describe the same mailbox. The trap for an AI SDR is that follow-ups stack quietly. An agent that enrols 40 new contacts a day into a five-step sequence will be sending 200 emails a day from that inbox by the fifth step. A safe setup caps both numbers: new contacts per inbox per day and total sends per inbox per day.
For reference, Outbound Labs sends up to 40 emails per inbox per day. On the Growth plan that's 5 warmed mailboxes, or at most 200 sends on a working day; see plans and volumes. If you're comparing sending tools on this point, our Outbound Labs vs Instantly page covers the differences.
An AI SDR should pause itself when bounces, complaints, replies or placement go wrong, and wait for a person before restarting. The thresholds should sit well inside the providers' limits, because providers measure differently from your dashboard and small daily volumes produce noisy rates.
| Signal | Healthy | Automatic action | Why that threshold |
|---|---|---|---|
| Hard bounce rate | Under 2% | Above 2%: pause the campaign and re-verify the list | Bounces mean stale data, and stale lists also hold spam traps |
| Spam complaints (7-day, whole domain) | Under 0.1% | Above 0.1%: pause new enrolments; never let it reach 0.3% | Google says aim below 0.1%; both Google and Yahoo cap at 0.3% |
| Reply rate | Above 2% | Under 1% for a week: flag list and copy for review | Low replies usually mean poor targeting before they mean spam |
| Placement test | Inbox | Spam or promotions: slow the ramp and recheck DNS | Catches problems before prospects' data does |
| Authentication | SPF, DKIM and DMARC pass | Any failure: stop sending from that domain | Required by all three providers at volume |
Two calculations explain the strict complaint line. First, Yahoo calculates spam rate on mail delivered to the inbox, not on mail sent. Say you send 1,000 emails, 600 reach the inbox and 3 people click "report spam". Your sending tool shows 3 ÷ 1,000 = 0.3%. Yahoo's maths gives 3 ÷ 600 = 0.5%, already over its limit. The worse your placement, the bigger the gap.
Second, small numbers are noisy. At 40 sends per inbox per day, a single complaint is 2.5% of that inbox's mail. A per-inbox daily trigger would pause everything on one grumpy recipient. Measure complaints over a rolling week across the whole sending domain instead. Google uses a similar window: a bulk sender regains eligibility for mitigation after its spam rate stays below 0.3% for 7 consecutive days.
Also know what your data can't see. Google's own help pages say Postmaster Tools data only covers personal Gmail accounts, and most B2B prospects use company domains. A low bounce rate isn't proof of health either. In one Hacker News thread on "verified" B2B data, a founder describes a list with 0% bounces and near-zero engagement, and argues that enterprise gateways accept every address at catch-all domains and then quietly quarantine mail from unknown senders. That's why the reply-rate row is in the table. Outbound Labs pauses a campaign automatically when bounces go above 2%.
Write short, plain-text emails that differ for every recipient because they're built on a real signal, not because a spinner swapped synonyms. Use one link at most, no images or attachments in the first touch, and a human way to opt out that's honoured within two days. Here is the kind of first email we'd approve:
Example (Northwind and the sender are made up):
Subject: Northwind's SOC 2 renewal
Hi Priya, saw Northwind posted two security engineer roles this month, which usually means the audit queue is growing faster than the team. We help security teams cut evidence collection for SOC 2 from weeks to days. Worth a 15-minute look before your renewal?
Sam
If this isn't relevant, just reply "no" and I won't follow up.
Nothing in it needs a tracking pixel, a link or an image. The signal does the personalisation. For how an agent finds signals like that, see how AI SDRs personalise emails.
Run this list before an agent's first campaign and after every pause. Any row that fails blocks the send.
| Check | Pass when |
|---|---|
| Sending domains | Separate lookalike domains, main domain untouched |
| SPF, DKIM, DMARC | All pass in a test message's headers; DMARC at least p=none |
| Warm-up | Three to four weeks done; placement test lands in the inbox |
| Daily caps | New contacts per inbox and total sends per inbox both capped in software |
| List | Verified within the last few weeks; role and catch-all addresses reviewed |
| Unsubscribe | Plain opt-out line; honoured within two days |
| Stop rules | Bounce, complaint, reply and authentication triggers set; restart needs a person |
| Approval | A named person approved the campaign, list and copy |
The fastest ways to burn a domain are about volume and data, not wording. Every item below is something an agent will do unless the software stops it:
Our view is that an AI SDR is safe when its limits are stricter than any provider's, enforced by the system and visible to the person who approves campaigns. In Outbound Labs the Sender agent sends up to 40 emails per inbox per day, pauses above 2% bounce, and nothing goes out until you approve the campaign. If you're still choosing tools, our guide to AI cold email software covers what to look for.
Belkins reviews from Clutch, Trustpilot and TrustRadius, its four published prices, cost per held meeting maths, and who should skip it. Checked Nov 2026.
Eathan Brown · 2 Nov 2026 · 8 min readSDR agency vs in house SDR compared on speed to meetings, control, management load, contracts, meeting quality and attrition, with a switch-over checklist.
Eathan Brown · 2 Nov 2026 · 8 min readInstantly AI pricing from Instantly's own pages: outreach, credits, CRM and bundles, cost per 1,000 emails, and the contradictions to check before you pay.
Eathan Brown · 1 Nov 2026 · 9 min readA Lindy AI alternative for every job: Zapier Agents, n8n, Gumloop, Relevance AI, Apollo, Salesforge, AiSDR and Outbound Labs, with a cost for 500 prospects.
Eathan Brown · 1 Nov 2026 · 9 min read