AI Personalization for Cold Email: What to Personalize at Scale
AI personalisation in cold email works best when it is grounded in real facts about the prospect, such as their role, company, a recent change, or a problem common to their segment, and used to explain why you are writing now. It fails when AI invents details or writes flattery. Personalise the reason for writing, not just the first line.
Prompt libraries and AI writing tools make it easy to produce a unique email for every prospect. That does not mean every email is relevant. This guide covers what is worth personalising, the levels of effort that match different deal sizes, how to keep AI honest, and how to test whether personalisation is paying off.
Key takeaways
- Relevance beats uniqueness. A segment-level message that names a real problem outperforms a unique email full of trivia.
- Personalise the “why you, why now”. The reason for writing matters more than a custom opening line.
- Match effort to deal size. Large accounts deserve manual research; small deals deserve good segments.
- Never let AI invent facts. Give it a source for every claim, or it will make one up.
- Review before sending until the edit rate is low, then keep sampling.
What does personalisation actually mean?
There are three things you can personalise in a cold email, and they are not equally valuable:
| What | Example | Value |
|---|---|---|
| Tokens | First name, company name | Expected; adds almost nothing on its own |
| Opening line | “Loved your post on hiring” | Low to medium; often reads as flattery |
| Reason for writing | “You're hiring two SDRs, and teams at that point usually hit X” | High: it explains why this email, to this person, now |
Most AI personalisation effort goes into the second row, because it is easy to automate. The third row is where replies come from. A good rule: if you deleted the personalised sentence, would the email still make sense? If yes, the personalisation was decoration.
Four levels of personalisation
Choose a level per campaign based on what one deal is worth and how many prospects you have.
| Level | What changes per email | Effort | Best for |
|---|---|---|---|
| 1. Segment | Problem and proof matched to a tight segment (industry, role, size) | Low | Small deals, large lists |
| 2. Segment + trigger | Adds a reason for now: hiring, a new location, a recent change | Medium | Most small B2B outreach |
| 3. Account research | A sentence or two drawn from the company's own site or news | Medium to high | Mid-sized deals |
| 4. Fully manual | Written by a person for one named buyer | High | Large deals, small target lists |
Level 2 is the sweet spot for most small teams, and the one where AI helps most: it can draft variations quickly once you have defined the segment, the problem and the trigger. Our guide to signal-based selling covers how to find good triggers.
What to personalise, in order of value
- The problem. Name a challenge that people in this role and segment really have, in their words.
- The trigger. Something that makes now a sensible time: a hire, a launch, a season, a regulation.
- The proof. A customer or example from a similar business, with permission to mention it.
- The ask. Smaller asks for colder prospects: a question or a resource, not a 45-minute demo.
- The opening line. Last, and only when it connects to the reason for writing.
How to use AI without inventing things
Language models write fluent, plausible sentences, including plausible false ones. A cold email that congratulates someone on an award they never won is worse than a generic email.
| Risk | What it looks like | Control |
|---|---|---|
| Invented facts | “Congrats on the Series B” to a bootstrapped company | Only let it use facts from fields you supply; instruct it to omit rather than guess |
| Flattery | “I was blown away by your inspiring LinkedIn post” | Ban praise phrases in the instructions; ask for a business reason instead |
| Overclaiming | Promising results or features you do not offer | List what may be claimed; review drafts before sending |
| Same-sounding emails | Every email opens with the same structure | Vary structure across steps; read a batch side by side |
| Wrong tone | Too casual for the market, or too stiff | Give real examples of emails you like |
A prompt structure that stays grounded
Whatever tool you use, give the model the same four inputs:
- Facts: the prospect fields you trust (role, company, industry, trigger), each labelled.
- Message: the problem you solve for this segment and the one proof point you may use.
- Rules: length limit, no flattery, no facts beyond those given, one question as the call to action.
- Examples: two or three real emails you consider good.
“If a fact is not in the data provided, do not mention it. Write a shorter email instead.” A shorter accurate email beats a longer invented one every time.
Want to see what your own numbers look like? Start a workspace and price it against the table above.
Start free14-day free trial · No credit card required
Get the data right first
AI personalisation is only as good as the fields behind it. Before writing prompts:
- Verify emails so personalised messages are not bouncing.
- Standardise role and industry into a small set of values you can segment on.
- Record the trigger as a field with a date, so stale triggers are not used months later.
- Keep a source for any research note, so a reviewer can check it.
Our guide to finding B2B leads covers verification and enrichment in detail.
Variation is not personalisation, but it matters
Sending the identical text to many people is a pattern mailbox providers can notice, independent of your bounce and complaint rates. Slight rewording per recipient helps the message look less like a template blast. That is a deliverability measure, not personalisation: a reworded irrelevant email is still irrelevant. Treat both separately, and keep the basics from our email deliverability checklist in place.
Review before you trust it
Roll out AI-written emails the same way you would roll out any agent:
- Draft only: generate a batch of 20, read them all, and fix the prompt.
- Approve each: send only after a person approves; track how often you edit.
- Sample: once edits are rare, review a random sample of each batch rather than every email.
Keep human review permanently for your highest-value accounts. The guide to setting up an AI sales agent covers approvals in depth.
How to test whether personalisation pays
Personalisation costs time or credits, so check it earns its keep. Split a segment in two: one half gets the segment-level email, the other the more personalised version. Compare:
- Positive reply rate, not opens. Opens are unreliable as a measure.
- Meetings booked from each half.
- Negative signals: unsubscribes, complaints and annoyed replies.
- Cost per meeting, including your review time.
With small lists, differences will be noisy; repeat the test over several batches before drawing conclusions. For the overall sequence structure, see cold email sequences for small business, and for message templates, sales email templates that get replies.
Compliance does not change
Personalised or not, cold email rules depend on where your recipients are. Identify yourself honestly, include a working unsubscribe and your postal address where required, honour opt-outs everywhere, and have a lawful basis for contacting each person. Using AI does not change any of that. This is not legal advice; check the rules for your markets.
How Scalry handles this
Scalry's cold email sequences pull merge fields from any CRM field, not just a first name, so a segment's problem, trigger and proof can be stored on the contact and used in the copy. By default, each sequence step is lightly reworded per recipient so the same step does not read as one template sent to everyone; you can switch that off for a hand-personalised sequence that should be sent exactly as written.
For research-led emails, AI sales agents can research a prospect and draft in your voice, with sending set to wait in the approval queue until a person approves the full draft. Every send carries a working unsubscribe and your business name and postal address, and opt-outs are suppressed across campaigns and agents. You send from your own mailbox with per-mailbox daily caps. Cold email outreach and the Lead Finder are on every plan; see pricing.
Merge fields, mailbox connections and sending caps are covered on the cold email page.
Questions to ask before you commit
Whichever way you are leaning, these are the questions that change the answer. Work through them against your own numbers rather than anyone's feature matrix — including this one.
- Which level of personalisation does one deal justify?
- Which prospect fields do you trust enough for AI to use?
- Who reviews drafts, and how will you track edits?
- How will you measure whether personalisation earns its cost?
If you cannot answer the first three, you are not ready to choose between Scalry and writing every email by hand yet — you are still deciding what you need, which is a cheaper problem to solve first.
Frequently asked questions
Does AI personalisation improve cold email replies?
It can, when it makes the email more relevant: a real problem, a real trigger and credible proof. AI-written flattery or invented facts usually hurt. Test personalised and segment-level versions on the same audience and compare positive replies and meetings, not opens.
What should I personalise in a cold email?
In order of value: the problem you address, the reason you are writing now, the proof you offer, and the size of the ask. A custom opening line is the least important, and only useful when it connects to the reason for writing.
How do I stop AI from making things up in emails?
Give it only the facts you trust, labelled clearly, and tell it not to mention anything not in that data. Ban flattery, set a length limit, include examples, and review drafts before sending until the edit rate is low.
Is personalising first lines worth it?
Usually less than people think. A personalised first line that has nothing to do with your offer reads as flattery. Spend the effort on why you are writing to this person now, which can be done well at segment level.
How many emails can I personalise per day?
That depends on your sending limits and review capacity, not the AI. Deliverability limits per mailbox apply regardless of how personalised the emails are, and every email you send should be one you would be happy to have reviewed.
Does varying email copy help deliverability?
Sending identical text to many people is a pattern mailbox providers can detect, so light per-recipient variation can help. It does not replace authentication, verified lists, low volume per mailbox, working unsubscribes and relevant content.
Build one segment-level sequence with merge fields from your CRM and approve the first batch by hand. You can rebuild one real campaign inside Scalry before moving anything else.
Start your free trialSee pricing14-day free trial · No credit card required
Build a workspace, import a real list, and run one campaign end to end before you decide anything.
14-day free trial · No credit card required
