AI Personalization That Does Not Sound Like AI

Almost everybody who has run AI-written cold email has had the same experience. The output is grammatically flawless, structurally sensible, and completely dead on arrival.

The reflex is to blame the model. The model is rarely the problem.

The actual failure

Ask yourself what the model was given. In most setups it is: a first name, a company name, an industry, and an instruction to "write a personalised cold email".

Now imagine giving a talented copywriter exactly that and nothing else. They cannot say anything specific, because they do not know anything specific. So they do what anyone would do with no facts: they compliment, they generalise, and they hedge.

"Hi Sarah, I came across Brightside Dental and was really impressed by the work you are doing in the community. Many practices like yours are looking for ways to modernise their patient experience while keeping costs manageable..."

That paragraph could be sent to fifty thousand businesses. The model did not fail. The brief did.

The four tells

Before fixing it, learn to recognise it. AI cold email gives itself away structurally, not grammatically, and once you can see these you cannot unsee them.

1. It opens with a compliment. "I was impressed by", "I love what you are doing with", "your work in the space is fantastic." Real people do not open emails to strangers with praise. They open with a reason for writing.

2. It hedges. "Many businesses like yours", "this could potentially help", "I thought it might be worth exploring." Hedging is what a writer does when they are not sure the claim is true, which is exactly the situation the model is in.

3. It arrives in threes. "Faster, simpler and more affordable." "Save time, reduce costs and improve retention." Tricolons are a stylistic fingerprint of language models, and once you notice how often the third item is filler, you notice it everywhere.

4. It closes by asking for time. "Would you be open to a quick 15-minute call next week?" This asks the recipient to spend a scarce resource before they know why. It is the single most deletable sentence in outbound.

The fix is context, not prompting

There is a whole genre of advice about prompt phrasing. It is largely a distraction. Two inputs change the output more than any wording ever will.

Input one: something true about them

Not a compliment. An observation, checkable, ideally slightly uncomfortable.

Every sentence on the right is a fact. Facts are what make an email feel written rather than sent, and they are also what make the rest of the message land, because you have earned the next line.

The practical implication is that personalisation quality is a data problem before it is a writing problem. An agent connected to real business records, with website presence, review signals, categories and recency, has raw material. An agent handed a spreadsheet of names is being asked to invent, and inventing is how you get flattery.

Input two: how you actually talk

The second input is your own voice, captured once. Not a tone instruction like "friendly but professional", which every model interprets as bland. Concretely:

• What you sell, in the words you use with a customer at a pub, not the words on your homepage.
• Your real pricing logic, including what you will not do.
• The three objections you get every week, and how you answer them.
• Two examples of results, with actual numbers.
• Something you are deliberately not: "we do not do rebrands", "we are not the cheapest".

That last one does more work than people expect. Constraints give a model edges, and writing without edges is what produces mush.

This is why the useful pattern is to train an agent once on a business brain rather than rewriting prompts campaign by campaign. You are not tuning a prompt, you are teaching a colleague.

Constrain the output hard

With good inputs, the remaining improvements are almost all subtractive.

Four lines maximum. Observation. Implication. Offer. Question.

Ban the tells explicitly. No opening compliment. No "I hope this finds you well". No lists of three. No "quick call".

Ask a closed question. "Do you get many calls after 6pm?" can be answered in three words from a phone. "Would you be open to a 15-minute call?" requires the recipient to make a commitment before they have a reason to.

Delete the first and last sentence. Genuinely. Take any AI draft, remove the opening line and the closing line, and read what is left. It is usually better, and usually the actual message.

Here is the earlier example, rewritten with real context and those constraints:

Sarah, your booking page sends people to a phone number, and two reviews this month mention nobody picking up after 5.

We put online booking on dental sites; the practices we do it for typically recover 6 to 10 appointments a month that were previously lost to voicemail.

Do you know roughly how many after-hours calls you miss?

Ninety words shorter, one checkable fact, one number, one question answerable in a sentence. No adjective anywhere doing decorative work.

Where AI still beats manual writing

It would be dishonest to end without saying this plainly: for your twenty best prospects, write the emails yourself. You will always do it better, and those twenty are worth the hour.

For the next two hundred, a trained agent working from real business data produces mail closer to your best work than you will at 4pm on a Thursday writing your fortieth message of the day. Consistency is the AI's actual advantage, not creativity. It never gets tired, never phones in the fifth follow-up, and never sends the version where you forgot to change the company name.

That is worth having. It is just not worth having if you feed it nothing.