Personalisation built from what a customer told you works. Personalisation built from what a model inferred about them produces complaints.
That line is the entire subject, and it does not move as the technology does. The risk is not that the inference is bad. It is that it is confident, specific and occasionally wrong — and a recipient reading a confident wrong statement about themselves concludes one of two things: that you were careless, or that you have been watching.
Complaints follow from watching much faster than from generic. And complaints are the metric receiving providers act on directly, with Google's bulk sender requirements stating they should stay below 0.3%.
The ladder
Personalisation data sits on a ladder, and value stops rising several rungs before risk starts.
| Rung | Example | Verdict |
|---|---|---|
| 1. Explicitly given | Name, preferences chosen at signup, stated interests | Use freely |
| 2. Directly observed on your own property | What they bought, what they downloaded, pages they visited while logged in | Use freely |
| 3. Aggregate behaviour | Segment membership, purchase frequency band | Use, carefully worded |
| 4. Inferred from thin signals | Industry guessed from email domain, seniority from job title text | Rarely worth it |
| 5. Enriched from third parties | Company size, revenue, tech stack, appended demographics | Do not |
| 6. Predicted | "Likely to churn", "probably interested in X" | Use for your own decisions, never state to the recipient |
Almost all the achievable value is on rungs one to three, and almost all of the risk is on four to six.
Rung six deserves its own note. A prediction is a legitimate input to a decision you make internally — which segment to send to, what to write about. Stating it to the recipient is where it goes wrong. "We noticed you might be losing interest" is a model's guess delivered as an observation about somebody's inner life.
Why inferred personalisation misfires
Four specific ways, each producing a different bad outcome.
1. It is wrong about a fact. Wrong industry, wrong role, wrong location, wrong assumed problem. The recipient learns your data about them is unreliable, which undermines every message after it.
2. It is right and reveals more than they shared. Correct inferences from data they did not knowingly provide are the ones that generate "how do they know that". Accuracy makes this worse, not better.
3. It is generic wearing specific clothing. "As a [industry] business, you probably struggle with [generic problem]" reads as a template with a slot filled — which is more insulting than an honestly general message, because it attempted personalisation and visibly failed.
4. It is stale. Inferred a year ago, still being asserted. A person who changed jobs, moved, or stopped doing the thing receives a message confidently describing someone they no longer are.
What is actually worth personalising
Five things, in order of return.
1. What the message is about. Segment-level relevance beats token-level personalisation by a wide margin. Sending the right topic to the right group does more than putting a first name in a subject line, and it carries none of the risk.
2. What they bought or downloaded. Concrete, verifiable, theirs. Reference the actual thing — not "your recent purchase" but the item.
3. Where they are in a process. Onboarding step, subscription stage, unfinished action. Useful and factual.
4. Timing relative to their action. Sent because they did something, not because it is Tuesday. This is the highest-value automation available and it is not really personalisation — it is relevance through timing.
5. Their name. Modest value, low risk, and only with a fallback that reads naturally in the whole sentence. "Hi there" works; "Thanks for your order, there" does not.
What is not worth personalising: anything requiring the recipient to be impressed that you know it. If the personalisation is the point of the sentence, cut the sentence.
Where AI is genuinely useful here
Three applications, none of which involve asserting anything about an individual.
1. Classifying your own data into segments. Reading a year of replies, support tickets or survey responses and grouping them into the themes that actually exist rather than the ones you assumed. This is the highest-value AI application in email and it is not personalisation — it is finding out who your list is.
2. Writing segment-level variants. One email, five versions for five segments. The variation is at segment level where it is defensible, not at individual level where it is a guess.
3. Drafting from data you already hold. Turning an order record into a sentence, correctly, at volume. Deterministic input, checkable output.
The rules that hold
Six, and they are stable regardless of what tooling appears.
- Only assert what the customer gave you or did on your own property
- Never state an inference to the recipient, however confident
- Never use appended third-party data in the message text
- Every token needs a fallback that reads naturally in the full sentence
- Fewer tokens is more reliable than better tokens. A message that works with none cannot break
- Personalise the topic before the text
One more, on data: anything you store about a person is subject to the rules of wherever they are, and inferred profiling attracts more scrutiny than data they handed you. Check the obligations for your own market rather than assuming, and prefer holding less.
Frequently asked questions
Does AI personalisation improve email performance?
Segment-level relevance does — sending the right topic to the right group. Individual-level inferred personalisation generally does not, and it carries a real cost: a confident wrong statement about somebody produces complaints, which is the metric providers act on.
What data is safe to use for email personalisation?
What the customer explicitly gave you and what they did on your own property — purchases, downloads, logged-in activity. Aggregate segment membership works if carefully worded. Inferred, enriched and predicted data should not be stated to the recipient at all.
Why does personalisation feel creepy?
Because a correct inference from data the person did not knowingly provide reveals that you have more than they shared. Accuracy makes it worse rather than better — a wrong guess reads as careless, a right one reads as surveillance.
Can AI predict which customers will churn?
Predictions are a legitimate input to your own decisions about who to send what. Where it goes wrong is stating the prediction to the recipient — "we noticed you might be losing interest" is a model's guess delivered as an observation about somebody's inner life.
What is the highest-value use of AI in email personalisation?
Classifying your own replies, tickets and survey responses into the themes that actually exist, so you can segment by what people asked rather than what you assumed. It is not personalisation at all — it is finding out who your list is.
Do personalisation tokens need fallbacks?
Yes, and the fallback has to read naturally in the complete sentence rather than on its own. Using fewer tokens is more reliable than configuring more of them: a message that works with no tokens cannot break.
Free: The 60-Minute Email Authentication Fix
A no-fluff checklist to set up SPF, DKIM & DMARC correctly and pass Gmail & Yahoo's sender requirements.
Muhammad Basim has worked in digital marketing since 2013, focused on email deliverability and AI-assisted content production. He is the author of The Email Deliverability Playbook and The Email Copywriting Playbook.
Related Articles
Does Send Time Optimisation Work?
Send time optimisation models are trained on when recipients opened messages, and open timestamps now include machine activity that has nothing to do with when anyone was reading. Privacy protection pre-fetches message content on its own schedule, regardless of whether the recipient opened it. Security gateways fetch content during pre-delivery scanning, which happens before the […]
Prompting for Marketers: The Part That Actually Matters
Output quality is set mostly by the material you supply, not the phrasing you use. That is the finding people take longest to accept, because it is less interesting than the alternative. A carefully worded request with no context produces a competent generic answer. A plainly worded request with your actual customer language, your positioning […]
What You Grant When You Connect a Tool
Your security perimeter includes every vendor you have ever connected, including the ones you stopped using and never disconnected. A connection is a standing grant. It does not expire because you stopped logging in, it does not lapse because the trial ended, and it does not disappear when you delete the app from your phone. […]