Sales Agents

Personalisation at Scale: What Actually Moves Reply Rates

The short answer

Merge fields are not personalisation. What raises reply rates is relevance: evidence in the first line that you understood something specific and true about this company's situation, followed by a reason that matters to them and a small ask. Doing this by hand costs five to ten minutes per prospect, which is why it stops happening above about thirty a day — and why it is the part of outbound worth automating properly rather than faking.

Open the folder where cold email goes to die and the pattern is immediate. "Hi Sarah, I came across Acme and was really impressed by what you're building in the workflow automation space." Nobody wrote that sentence about Acme. It was generated for four thousand companies, and every recipient knows it.

The uncomfortable part is that this performs worse than a plain message with no personalisation at all. A blunt, honest email is at least honest. A template wearing a costume signals that the sender wanted the appearance of effort without the effort, which is not a good opening argument for a business relationship.

Three levels, only one of which works

Merge fields. Name, company, sometimes title. This is not personalisation; it is mail merge, and it has been legible as mail merge since roughly 2015. It costs nothing and it buys nothing.

Scraped detail. A line pulled from a LinkedIn bio or an About page, dropped into a template slot. Slightly better, and still recognisable, because the detail sits in the sentence without connecting to anything. "I saw you're passionate about customer experience — anyway, we help companies with…" The seam is visible.

Genuine relevance. Evidence that the sender understood something about this company's situation and wrote to them because of it. This is the only level that changes reply rates, and the difference is not marginal.

What separates the third level from the second is not how much detail is used. It is whether the detail is load-bearing. If you could swap in a different company's fact and the email would still make sense, the detail was decoration.

What relevance is made of

Useful raw material is almost always public and usually recent:

  • Hiring. A job posting is a company saying out loud what it is struggling with. Two operations roles open at once is a bottleneck being admitted to in public.
  • Funding. Not as flattery, but as timing — a round means a plan and a deadline, and both create the kind of problem that gets budget.
  • Product changes. A launch, a new pricing page, a market they have just entered.
  • People changes. A new head of the function that owns your problem is a person with ninety days to demonstrate something.
  • Their own words. What the website claims the product does, in the terms the company uses for it.

The shape that works is: specific observation about them, the implication you have drawn from it, the reason you are therefore writing, a small ask. In that order, with the observation first — because the first sentence is the only one guaranteed to be read.

Why this stops happening around thirty a day

Doing the above properly takes five to ten minutes per prospect: read the site, check recent news, look at open roles, work out what any of it implies, write something that is not generic. At fifty prospects that is four to eight hours. Nobody sustains it.

So the compromise happens. A template is written once, merge fields are added, volume goes up, reply rate goes down, and the response to the falling reply rate is more volume. That loop is how a domain gets burned and how a team concludes that outbound does not work.

The honest framing is that personalisation at scale was never a copywriting problem. It was a research problem with a copywriting step at the end, and the research is what does not scale by hand.

What AI does and does not fix here

An AI that is handed a template and asked to reword it produces variation, not relevance. Four thousand differently-phrased versions of a message that was never about the recipient is the same email with extra steps, and recipients spot it just as quickly.

An AI that reads the company's site, its recent announcements and its open roles, forms a view about what that company is dealing with, and writes from that view is doing the thing a good rep does — just for the two-thousandth account as carefully as the first.

The test is the swap test, and it is worth applying to any tool that claims to personalise. Take two generated emails, exchange their opening lines, and see whether either still makes sense. If they do, nothing was personalised.

The parts people get wrong after the first line

Length. Fifty to a hundred and twenty words. Anything longer is being read on a phone by someone standing up.

The ask. "Do you have 30 minutes on Thursday?" is a large request from a stranger. "Is this something you're dealing with?" is a small one, and a yes to a small question is how the conversation starts.

Follow-up. Most replies come after the first message, and most sequences waste that by sending "just bumping this to the top of your inbox". A follow-up should add something — a different angle, a relevant example, a more specific question — or not be sent.

Stopping. Someone who replies should never receive the next scheduled message. It is the most basic thing a sequence can get wrong and it happens constantly.

Where this leaves the work

Relevance is not a writing style you adopt. It is the output of research you either did or did not do, and at any real volume the research is the constraint.

Operater's Sales agent does that research per person before it writes: what the company sells, what it has just announced, who it is hiring, what any of that implies for the thing you sell. Every message is written from that, one at a time, rather than assembled from a template — and when someone replies, the sequence stops on its own.

Key takeaways

  • A merge field is legible as a merge field. It reads as less effort than no personalisation at all.
  • Relevance beats flattery: the opening should be about their situation, not about how impressive they are.
  • The first sentence carries almost all of the weight. Everything after it is read only if it survives.
  • Ask for something small. A meeting request in a first cold email is asking a stranger for half an hour.

Frequently asked questions

Does personalisation actually increase cold email reply rates?

Genuine relevance does, consistently and substantially. Superficial personalisation — a first name, a company name, a scraped compliment — does not, and can perform worse than a plain, honest message, because recipients recognise the pattern and read it as automated flattery.

What should the first line of a cold email be?

Something specific and true about the recipient's situation that explains why you are writing to them in particular. A recent funding round, a role they are hiring for, a change in their product. Not a compliment, not your name, and not what your company does.

Can AI personalise cold email well?

It can if it is given real material to work with. An AI that only rewrites a template in different words produces variation without relevance. An AI that reads the company's site, recent news and hiring activity, and writes from that, produces the kind of opening a person would write if they had time to research every prospect.

How long should a cold email be?

Between fifty and a hundred and twenty words for a first touch. Long enough to say what the problem is, why you are writing to them and what you want; short enough to read on a phone without scrolling.