Send from a separate, authenticated domain (SPF, DKIM and DMARC set up), warm it up for two to four weeks, and keep each inbox to a few dozen cold emails a day. Have AI write short plain-text emails personalised from real research, with no images, at most one link and an easy opt-out. Keep spam complaints under 0.1%.
Most spam problems with AI-written cold email have little to do with "spam words". They come from two places: a sending set-up mailbox providers don't trust, and emails so generic that recipients report them. Gmail's sender guidelines ask all senders to keep the user-reported spam rate below 0.1% and never let it reach 0.3%. On a small campaign that's a very thin margin: one complaint in 300 emails already puts you at 0.33%. So the rules below are about being trusted and being relevant, in that order.
Why AI-written cold emails land in spam
Mailbox providers such as Gmail and Outlook decide where an email goes mainly from the sender's reputation and how recipients react. AI makes it easy to damage both at once:
- Volume jumps. AI can write 500 emails in a minute, so people send 500 from a mailbox that sent 20 last week. Sudden volume from a new or quiet domain looks like a compromised account.
- Sameness. A template with "[First name]" and "[Company]" swapped in produces hundreds of near-identical messages, a pattern filters recognise.
- Generic, long, hyped copy. "I hope this email finds you well" followed by "innovative solutions" and four paragraphs. Recipients don't reply; some press "report spam".
- Links, images and tracking. Logos, tracked links and open-tracking pixels make a one-to-one email look like a marketing blast.
- Poor lists. Addresses from old or bought lists bounce, and bounces damage reputation quickly.
The same generic-copy problem affects every kind of AI marketing writing; why AI marketing copy sounds generic explains the causes. For cold email, the cure starts before any writing.
Set up a separate sending domain first
Never send cold email from the domain your invoices, quotes and customer emails come from. If the cold campaign damages its reputation, your everyday mail starts landing in junk too. Here's how that goes wrong, in an illustrative case: a wholesaler sent 400 AI-written introductions in one afternoon from its main domain. Within a week, two regular customers mentioned that invoices had gone to their junk folders, and it took over a month of careful sending to recover.
Instead, register a second domain that's clearly yours, for example get[yourbrand].com or [yourbrand]storage.com, point its website address at your main site, and create one or two mailboxes on it with a business email service. Google Workspace Business Starter or Microsoft 365 Business Basic cost about $7 a user a month on annual plans. Use real names for the mailboxes (the salesperson who'll answer replies), not "sales@" or "info@".
The three DNS records, filled in
SPF, DKIM and DMARC are short text records you add in your domain's DNS settings (the control panel where you bought the domain). They tell receiving servers which services may send for your domain and let them check your messages haven't been forged. Your email provider's admin console gives you the exact values; here's what they look like for a domain sending through Google Workspace:
SPF (TXT record on the domain itself)
v=spf1 include:_spf.google.com ~all
DKIM (TXT record; the admin console generates the key)
Name: google._domainkey
Value: v=DKIM1; k=rsa; p=MIIBIjANBgkqh...(long key)...
DMARC (TXT record)
Name: _dmarc
Value: v=DMARC1; p=none; adkim=r; aspf=rA few notes on those. If a cold-email tool sends on your behalf through its own servers, it will ask you to add its details to SPF and set up its own DKIM; follow its set-up page exactly, and keep a single SPF record (two SPF records on one domain break both). Starting DMARC at p=none means "monitor, don't block", which is what Gmail's rules require as a minimum. Add a reporting address (the rua tag) pointing to a mailbox you check or to a free DMARC report service, so you can see failures. Once reports show everything passing for a few weeks, you can tighten the policy.
Why this matters even for small senders: Gmail requires everyone to have SPF or DKIM, and requires senders of more than 5,000 messages a day to personal Gmail accounts to have SPF, DKIM and DMARC, with the From domain aligned to one of them (see Google's email sender guidelines). Microsoft began applying similar rules to high-volume senders to Outlook.com addresses from 5 May 2025; non-compliant mail is rejected with the error "550; 5.7.515", as Microsoft's announcement explains. You won't be anywhere near 5,000 a day, but meeting the full set anyway removes one reason to distrust you.
Warm up and pace the sending
A new mailbox has no reputation. Build one slowly with real, two-way email before any cold sending, then ramp up gradually. A pace I'd use for a new domain with two mailboxes:
| Week | Per mailbox, per day | What to send |
|---|---|---|
| 1-2 | 5-10 | Normal emails to colleagues, suppliers and friendly customers who'll reply; or the warm-up feature of your sending tool |
| 3 | 10 cold + warm-up | Your best-researched prospects first |
| 4 | 20 cold + warm-up | Watch bounces and replies daily |
| 5 onwards | 30-40 cold | Hold here unless you add mailboxes; don't push one mailbox harder |
Spread the sends through working hours rather than firing them at 9am, and send on weekdays. If you need more volume, add a mailbox and warm it the same way, rather than raising the daily number. Cold-email tools such as Instantly (outreach from $47 a month) and lemlist (from $69 a month, or $55 billed annually) include warm-up and spread sending across mailboxes; the comparison of lead generation tools covers what else they do.
Prompting AI for emails that read like one person wrote them
The single biggest improvement is giving the AI real research about each recipient and forbidding the habits that make cold email look mass-produced. Here's a prompt used by an illustrative storage facility offering business storage to online sellers:
Write a first cold email from [name], manager at [storage
facility], to the person below. We offer business storage units
with a loading bay, pallet racking, parcel acceptance on weekdays
and month-to-month contracts.
RECIPIENT RESEARCH (use at least one specific fact):
[company, what they sell, evidence of growth, where they
currently ship from if known, any source notes]
Rules:
- Under 90 words. Plain text. No links, no images.
- First line is about them, not us. No "I hope this finds you well".
- One benefit that matches the research. One question at the end.
- No hype words: innovative, solution, leading, synergy, game.
- No false urgency, no discounts.
- Last line: "If this isn't relevant, just say and I won't email again."
- Write like a busy manager, not a marketer.With research notes saying the recipient sells handmade candles online, recently added a wholesale range, and posted that "the spare room is officially full", an illustrative first draft came back as:
Subject: Your spare room
Hi [first name], I saw your post about the spare room being officially full, which usually happens right after a wholesale range takes off! Congratulations. We rent business storage units with a loading bay and racking, and we can accept your parcel deliveries on weekdays, so you can pack and ship from one place. Contracts are month to month. Would one of our smaller units cover the next six months of growth?
If this isn't relevant, just say and I won't email again.
That's close. Three fixes: the exclamation mark and "Congratulations" add a salesy note, so cut them; "we can accept your parcel deliveries" should say incoming deliveries, since she ships outgoing parcels too and the service only covers incoming ones (the kind of factual slip you must check); and the size question assumes too much, so ask about the problem instead. Revised: "Hi [first name], I saw your post about the spare room being officially full, which tends to happen right after a wholesale range takes off. We rent business storage units with a loading bay and racking, and we accept incoming stock deliveries on weekdays, so you can store, pack and ship from one place, month to month. Is space the bottleneck right now, or is it packing time?"
Where do the research notes come from? From the kind of prospect research described in finding B2B prospects with AI research tools. Without real notes, the AI invents plausible-sounding "personalisation", which is worse than none.
The same email, before and after
Before (a typical unguided AI draft):
Subject: Transform Your Storage Today! 🚀
Hi [First Name],
I hope this email finds you well! I'm reaching out because [Company] is clearly a leader in the e-commerce space, and I believe our innovative storage solutions could be a total breakthrough for your business. [Logo image]
At [Storage Facility], we offer state-of-the-art units, flexible terms and exceptional customer service. See our amazing units: [tracked link]. Book a call here: [tracked link]. Follow us on social media: [three icons]
Best regards,
[Name] | [Title] | [Phone] | [Website] | [Banner image]
Count the problems: an emoji and hype in the subject, a placeholder left unfilled ("[First Name]" is a real risk when merge fields fail), no fact about the recipient, a string of tracked links, a logo, social icons and a banner, and nothing to reply to except a booking link.
After: the revised version above: 70 words, no links, no images, one fact about them, one question, a plain opt-out, and a text signature of name, job title, business name and phone number.
Links, images, tracking and attachments
- Links. None in the first email if you can manage it; at most one, to your own main website, written out plainly. Never use link shorteners.
- Images and logos. Leave them out, including in the signature. A plain-text email looks like one person writing to another.
- Open tracking. Switch it off for cold email. It works by adding a hidden image, which makes the message look like marketing, and open counts are unreliable anyway, because many email apps now load images automatically before anyone reads the message.
- Click tracking. Tools rewrite links through a tracking address shared with other senders. If you must track clicks, set up a custom tracking domain in your tool; better, don't include links.
- Attachments. Never in a first email. Offer the price list or brochure and send it when they reply.
A storage facility's first 200 cold emails
Here's how the approach can play out for the illustrative storage facility. The manager registered a second domain, set up two mailboxes on Google Workspace (about $14 a month), added the three DNS records, and warmed up by hand for two weeks by emailing existing business customers about their accounts. She researched prospects in batches of 30 (online sellers and small trades businesses within a short drive), and sent at the pace in the table above for four weeks, writing with the prompt above and editing each draft.
| Four weeks of sending (illustrative) | Figure |
|---|---|
| Emails sent (first touch plus one follow-up) | 212 to 140 prospects |
| Bounces | 3 (2.1% of prospects) |
| Replies | 19 |
| Positive replies | 7 |
| Site visits booked | 5 |
| Business units let | 2 |
| Spam complaints reported | 0 visible |
| Manager's time | About 40 minutes a day researching and editing |
The quick sum: two business units at an illustrative $260 a month each, likely to stay a year or more, is over $6,000 of annual revenue from a month of 40-minute sessions and $14 of mailboxes. A 13% reply rate (19 of 140) is what well-researched, low-volume email can achieve; a blast of AI templates to a bought list typically gets a fraction of that, plus complaints.
One caution on measuring: at this volume, Google Postmaster Tools may show little or no data, because most of its dashboards only show data once a domain sends a sizeable daily volume to Gmail users (Google puts it in the hundreds). Small senders should watch bounces, replies and a weekly test email to their own Gmail and Outlook addresses instead.
Six checks before every batch goes out
Run through these each time you load a new batch into your sending tool. It takes five minutes and catches most of the errors that cause complaints:
- Merge fields filled. Preview five random emails; look for "[First name]" or an empty greeting.
- Suppression list loaded. Existing customers, past opt-outs and anyone who said no are excluded.
- Every address verified. No "accept-all" or guessed addresses in the first touch.
- One fact per email checked. Spot-check that the personal line is true for five recipients.
- Volume within the ramp. Daily cap per mailbox set in the tool, not left to memory.
- A test send to yourself. Send one to your own Gmail and Outlook addresses and confirm it lands in the inbox, reads well on a phone and shows your name, not the tool's.
Rules on cold email you need to check
Rules about unsolicited email vary a lot depending on where recipients are, and on whether they're businesses or individuals. Some places allow emailing business contacts with a clear opt-out; others require prior consent for many recipients, including some people at small firms. The safe habits everywhere:
- Say who you are and give your business's real contact details.
- Use honest subject lines that match the content.
- Include a plain way to opt out, and honour it straight away.
- Keep a suppression list and load it into every tool you use.
- Record where each contact came from, because work emails are personal data under laws such as the GDPR, and people can ask.
If you're emailing people in several countries, or individuals rather than companies, ask a data-protection adviser which rules apply before you start.
Warning signs in the first fortnight
| What you see | Likely cause | What to do |
|---|---|---|
| Bounces above about 3% | Stale or guessed addresses | Stop; verify the rest of the list; drop the source if it's consistently bad |
| Replies asking "how did you get my email?" | Targeting too loose, or no visible reason for contact | Tighten the research; say in the first line why you're writing to them |
| Your own test email lands in spam | Reputation dipping | Pause cold sending for a week, keep warm-up running, halve the volume when you restart |
| Bounces mentioning "550 5.7.515" | Outlook authentication check failing | Check SPF, DKIM and DMARC alignment with your provider's checker |
| Replies suddenly stop altogether | Emails being filtered, or a merge-field error | Send yourself a copy through the tool and read it as a recipient would |
| Everyday mail from your main domain going to junk | Cold email sent from the main domain | Move all outreach to the separate domain now |
Once first emails are landing and getting replies, the next job is following up without becoming a nuisance, which has its own rules on timing and when to stop. Automating sales follow-ups without being pushy covers that part.
Cold email and spam filters: common questions
Do spam trigger words still matter?
Much less than people think. Modern filters weigh your domain's reputation, authentication, complaint rate and how recipients react far more than individual words. A message full of 'FREE!!!' and 'guaranteed' still looks bad, but removing a word won't rescue an email sent from an unauthenticated new domain to a stale list. Spend your effort on set-up, list quality and relevance first.
Can I send cold emails from a free Gmail or Outlook address?
Don't. Free consumer addresses look unprofessional to business buyers, have lower sending limits and can be suspended for bulk sending. Use a mailbox on a domain you own, on a business email service such as Google Workspace or Microsoft 365, set up with SPF, DKIM and DMARC. A separate domain close to your brand name keeps your main domain safe if something goes wrong.
Do Google or Microsoft filter emails for being written by AI?
Neither says it filters emails for being AI-written. What gets filtered is the pattern that careless AI use produces: hundreds of near-identical messages, generic wording people ignore or report, and heavy links and tracking. An AI-drafted email that is short, specific to the recipient and sent at low volume from a properly authenticated domain is treated like any other one-to-one business email.
Should I pay for an email warm-up service?
Warm-up is included in most cold-email tools, such as Instantly and lemlist, so you rarely need a separate service. You can also warm up by hand: for two or three weeks, send normal emails from the new mailbox to colleagues, suppliers and customers who will reply. Real two-way conversations are the best reputation signal. Warm-up helps a new domain; it can't fix a bad list or poor emails.
Further reads
- AI Lead Generation vs Hiring a Salesperson: A Cost Comparison — Whether cold outreach with AI beats hiring someone to do it.
- How to Write Sales Call Scripts With AI That Don't Sound Scripted — What to say when a cold email turns into a call.
- How to Segment Your Email List With AI, Even a Small One — Sending different messages to different kinds of prospect.
- Email Newsletter QA Checklist Before You Hit Send — A pre-send checklist habit that also suits outreach.
- AI Lead Generation for Small Businesses: What Actually Works — Where cold email fits among lead sources that work for small firms.
- Can AI Help a Small Business Grow Revenue, or Only Cut Costs? — When AI grows revenue and when it only trims costs: diagnose your bottleneck, pick from five revenue levers, and check two worked examples with numbers.
- AI for Recruitment Business Development: Winning New Clients — A weekly business-development routine for agencies: hiring signals, five-minute account briefs, outreach that names the vacancy, and consented spec profiles.
- How to Write a Sales Deck With AI That Wins Meetings — Research one buyer, draft the storyline in plain text, turn every headline into a claim, and only then let an AI slide tool do the design.
- What Does an AI SDR Cost, and What Does It Actually Do? — Published AI SDR prices from about $1 a lead to $3,750 a month, the mailbox and staff costs around them, and how to work out cost per booked meeting.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google's Gmail email sender guidelines and sender guidelines FAQ (spam-rate thresholds, bulk sender rules); Microsoft Defender for Office 365 announcement of Outlook's high-volume sender requirements; Instantly and lemlist pricing pages. Checked September 2026.