"I loved your post about remote onboarding" is now a negative signal. Not neutral — negative. Five years ago that line proved a human had spent four minutes on you. Today it proves the opposite: it is the single most automatable sentence in B2B sales, generated at a rate of thousands per hour by tools that scrape a LinkedIn feed and pattern-match a compliment. Buyers have learned this. The line that once said I did my homework now says you are number 4,000 in a sequence.
So the honest answer to "which AI tool writes the most personalised cold email" is that you are asking the wrong question. Personalisation is no longer scarce, which means it is no longer valuable. What is still scarce: being obviously relevant, landing in the inbox at all, and being short enough to read. This piece is about the tools that serve those three things — Clay, Apollo, Outreach, Lavender, Instantly and Smartlead — and about the constraint that beats all of them.
The real constraint is deliverability, not copy
You can write the best cold email in the category and it will do nothing from the spam folder.
In 2024 Google and Yahoo introduced bulk sender requirements that changed the economics of outbound permanently. The published guidance is specific and worth knowing rather than paraphrasing:
- Authenticate your domain. SPF and DKIM, plus DMARC for bulk senders. Not optional any more.
- One-click unsubscribe on bulk mail, honoured within a short window.
- Keep spam complaints low. Google's sender guidelines say to keep the reported-spam rate below 0.3%, and ideally under 0.1%. That is Google's number, published in its own documentation — and it is a brutally low bar to clear. It means a few complaints per thousand emails puts your domain in trouble.
Everything else follows from that third bullet. Volume is a liability now. Irrelevance is expensive. The mass-blast strategy that AI made cheap is exactly the strategy those rules were written to kill.
The operational consequences most teams still have not internalised:
Never send from your primary domain. Buy lookalike sending domains, warm them, and keep the domain your invoices and contracts run through completely out of the blast radius. If you burn your main domain, the damage is to the whole company, not to the campaign.
Volume per inbox stays low. A modest daily send per mailbox, ramped slowly. Tools that let you attach fifty inboxes and rotate through them are solving for the symptom; the underlying reason you need fifty inboxes is that you are sending too much.
Plain text, few links, no images. HTML-heavy sales mail with a tracking pixel and three links looks like exactly what it is. And open tracking, since Apple's Mail Privacy Protection, is close to useless as a metric anyway — you are paying a deliverability tax for a number you cannot trust.
Clean your list. Bounces hurt. Verify before you send.
What actually still works
Cut personalisation into two things that get confused constantly.
Cosmetic personalisation — the merge field, the flattering reference to a post, the "congrats on the funding round". Every competitor can now do this, at scale, for a fraction of a cent. It has been fully commoditised, and worse, it has become a tell. When a buyer sees a compliment about their content in the first line, the modern inference is not "this person read my work." It is "this is automated."
Structural relevance — the email is obviously, checkably about a problem this specific company has right now. They are hiring three integration engineers, which means they are drowning in bespoke customer integrations. Their pricing page just added a usage tier, which means metering is suddenly somebody's problem.
The second kind cannot be faked, because it requires a hypothesis, and models are bad at hypotheses about businesses they have no context on. It also does not need to be flattering. The best cold email is often almost rude in its brevity: here is what I think is true about your situation, here is what we do about it, worth ten minutes?
And the unglamorous truth underneath all of it: a boring, short, obviously relevant email to 200 tightly-qualified people will beat a smarmy AI-flattered email to 5,000 loosely-qualified ones, on every metric including the one that matters to your domain reputation. Targeting is the lever. Copy is downstream.
The tools, by the layer they actually own
Cold email is four jobs, and almost nobody is good at more than two. Buy per layer.
| Tool | Layer it owns | Best for | Where it disappoints |
|---|---|---|---|
| Clay | Data enrichment, research, list building | RevOps teams who will genuinely build | Steep learning curve; credit costs escalate fast |
| Apollo | Contact database plus sequencing, all in one | Small teams wanting one cheap tool | Everyone emails the same contacts; data quality is uneven |
| Outreach | Enterprise sequencing, governance, reporting | Large teams needing process control | Expensive, heavy, no data of its own |
| Lavender | Real-time coaching on the writing | Improving reps who write badly | Does not find, send, or route anything |
| Instantly | Sending infrastructure, warmup, inbox rotation | High-volume operators | Makes it very easy to do the wrong thing at scale |
| Smartlead | Sending infrastructure, whitelabel, API | Agencies running many clients | Same risk; power tool, no seatbelt |
Swipe the table sideways to see every column →
Clay — the one that changes how you think
Clay is a spreadsheet crossed with an enrichment engine. You build a list, then run each row through a waterfall of data providers and AI research agents that go and check things on the open web — does this company have a careers page mentioning Kubernetes, do they use Shopify, did they announce a funding round, is there a security page.
That capability, used correctly, is what makes structural relevance possible at more than ten emails a day. Used lazily, it is also the industrial-scale generator of the fake compliments described above. The tool is genuinely neutral. Almost nobody uses it neutrally.
Who should use it: a team with someone who enjoys building systems and will invest real hours. Who should not: a founder sending fifty emails a week. You will spend more time configuring waterfalls than you would have spent reading fifty websites — and reading fifty websites would have produced better email. The trade: credit-based pricing that looks cheap in the demo and does not stay cheap. Model your volume before you commit.
Apollo — cheap, capable, and everybody has it
Apollo bundles a large B2B contact database with sequencing and a Gmail extension. For a small team it is the pragmatic default, and the free tier is real enough to evaluate properly.
The catch is structural, not technical. If you can buy a list of every VP of Engineering at a Series B company for a low monthly fee, so can your six competitors — and they did, this quarter. The contacts most likely to be in Apollo are the contacts most likely to be fatigued. Data accuracy is also variable, and stale records mean bounces, and bounces mean deliverability damage.
Who should use it: early-stage teams who need to start somewhere without a data budget. Who should not: anyone selling into a niche where the list needs to be built rather than bought.
Outreach — process, not magic
Outreach is a sales engagement platform for organisations big enough that consistency and reporting matter more than cleverness. It sequences, it enforces process, it tells the VP what is happening.
Who should use it: teams of fifteen reps and up where the problem is that everyone does outbound differently. Who should not: a team of three. It is expensive, it takes real implementation effort, and it brings no contact data of its own — you still need to solve the data layer separately. Buying it early is buying governance for a process you have not designed yet. The rest of that stack is mapped out in our AI sales tools guide.
Lavender — the coach, not the writer
Lavender scores your draft while you write it, in the compose window, and tells you the thing your manager will not: the email is too long, it is all about you, it is written at a reading level nobody skimming on a phone will finish, and the ask is unclear.
It is the only tool on this list that makes the human better rather than making the machine faster, which is a strange thing to find undervalued in this market. It also does exactly nothing about your list, your domain reputation, or your sending infrastructure. It is a layer, and it knows it. (There is a free individual tier, historically — check what is current; we cover it in the free Gmail AI assistants roundup.)
Instantly and Smartlead — infrastructure with no seatbelt
Both are sending-infrastructure tools: mailbox rotation, automated warmup, deliverability tooling, high volume at low cost. Smartlead leans agency, with whitelabel and API depth; Instantly leans operator.
They do their job well. The problem is what their job makes easy. These tools remove the natural friction that used to cap how much bad outbound one person could do, and the mailbox providers responded by making bad outbound existentially dangerous. A tool that lets you attach forty inboxes is not a strategy. It is a way to distribute the consequences of not having one.
Who should use them: operators who have already fixed targeting and are scaling something that demonstrably works. Who should not: anyone whose reply rate is bad. Sending more of a bad email from more inboxes is the single most common and most expensive mistake in outbound.
The week before you send anything
- Define the segment narrowly enough to say it in one sentence. If you cannot, your email will not be relevant, no matter which model writes it.
- Write the hypothesis, not the compliment. One sentence: what do I believe is true about this company's problem, and what fact makes me believe it? If you cannot fill that in, remove them from the list.
- Set up authentication. SPF, DKIM, DMARC on your sending domains. Warm them. Do not skip this and hope.
- Kill open tracking. It costs deliverability and, post-MPP, buys you a number you cannot trust. Measure replies.
- Cap volume per inbox and ramp slowly. Resist every tool that encourages otherwise.
- Read the email aloud on a phone screen. If it takes more than fifteen seconds, cut it in half.
- Check the law before the copy. In the US, CAN-SPAM requires accurate headers, a real postal address and a working opt-out. In the EU and UK, GDPR requires a lawful basis — legitimate interest is arguable for B2B and it is not a free pass, and it obliges you to honour objections. This is not legal advice; it is a prompt to get some.
FAQ
Does AI-personalised cold email still work?
Cosmetic personalisation — the merge field, the reference to a recent post — has stopped working, because it stopped being evidence of effort once every competitor could automate it. Structural relevance still works: an email that shows you understand a specific problem this company has right now, backed by a checkable fact. The tools have not stopped helping. What they help with has changed.
What is the best AI tool for cold email personalisation?
There is no single one, because the job is four jobs. Clay owns research and list building, Apollo owns cheap data plus sequencing, Outreach owns enterprise process, Lavender owns the writing, and Instantly or Smartlead own the sending infrastructure. Teams that buy one tool expecting it to cover all four end up mediocre at all four. Buy per layer, starting with the layer that is actually broken.
How many cold emails can I send per day safely?
Fewer than the tools imply, and there is no universal number — it depends on domain age, warmup, list quality and, above all, your complaint rate. The binding constraint is Google's published guidance to keep reported-spam rates below 0.3%, ideally under 0.1%, which is a very small number of complaints per thousand sends. Anyone quoting you a confident daily figure without asking about your list is selling something.
Should I use my main domain for cold email?
No. Use dedicated sending domains that look like your brand but are not it, warm them properly, and keep your primary domain — the one your invoices, contracts and support mail depend on — entirely out of the campaign. A burnt sending domain is a bad week. A burnt primary domain is a bad quarter for the whole company.
Is AI cold email legal?
The tool does not change the rules; the sending does. Under CAN-SPAM in the US you need accurate headers, identification, a physical postal address and a working opt-out. In the EU and UK, GDPR requires a lawful basis for processing, and while legitimate interest is commonly relied on for B2B, it needs to be justifiable and you must honour objections. Get advice for your jurisdiction rather than trusting a vendor's blog.
What replaces "I loved your post about X"?
A fact and a hypothesis. "You are hiring three integration engineers" is a fact. "Which usually means custom integrations are eating your roadmap" is a hypothesis. Together they take one line, prove you looked, and — unlike a compliment — invite a reply even when the hypothesis is wrong, because people enjoy correcting you about their own business.
Where to go next
The wider field is in AI sales tools and AI email tools, with the CRM layer that everything eventually has to write back into over in AI CRM tools. If you are building the outbound function rather than just the sequence, the sales teams hub is the place to start. And if you are running something in outbound that is genuinely working, submit it — the tools that work rarely have the biggest ad budgets.