How AI SDRs personalise cold emails: signals, worked examples and the checks that matter

Key takeaways

• Good AI personalisation is one checkable fact plus the inference it supports, not a compliment.
• Pick the signal by seniority: Gong Labs found company-level signals tripled replies from directors and above, while personal-level signals worked best below manager level.
• Every claim about a prospect needs a source and a freshness window; no source, no line.
• Test signal types, not sentences, and plan on roughly 750 sends per signal type to see a 3% to 6% reply lift.

AI SDR personalisation is the process of researching each prospect for a specific, recent fact and turning it into a first line that shows why you are writing to this person now. That sounds simple. Most of what goes wrong happens in the step nobody shows you: the jump from the fact to what the fact means.

Buyers can tell the difference. In a Gong Labs study, 87% of buyers said the sales emails they receive don't address a relevant challenge facing their organisation (Gong Labs). A congratulations on a funding round doesn't fix that. A line that connects the funding round to a problem you solve might.

What does an AI SDR actually do when it personalises an email?

An AI SDR runs five steps for every prospect: it gathers data, picks one signal, infers what that signal means for the buyer, writes an opener of one sentence, and bridges to the offer. Merge tags only do the first step. The inference is where real personalisation lives, and it's also where models make things up.

In our setup, the Researcher agent pulls the company and contact data (in Outbound Labs, contact data comes from B2B data provider APIs). The Writer agent drafts the email. Then a person, you, approves the campaign before anything sends. You can see how the six agents split the work on the Outbound Labs agents page, and the wider workflow on our AI SDR overview.

Which signals do AI SDRs use to personalise cold emails?

AI SDRs mostly use public company signals: job posts, funding, leadership changes, product or pricing changes, technology changes and published content. The best signal is recent, specific to the buyer's job and connected to the problem you solve. The freshness windows below are our working rules, not industry standards. Adjust them to your sales cycle.

Notice what is missing: hobbies, family, holidays. A machine that finds those at scale reads as surveillance. We would skip them even when they're public.

Should the personalisation change with the prospect's seniority?

Yes. Gong Labs analysed more than 30,000 prospecting emails from over 250 companies and found that company-based personalisation (company news, launches, milestones) tripled reply rates from directors and above, with little effect on lower-level buyers. Individual-based personalisation more than doubled replies from non-managers but lifted director-and-above replies by only 50% (Gong Labs).

So tell your AI SDR which signal to prefer for each persona. For a CFO, lead with the funding round or the expansion. For a sales ops manager, lead with their new role or the post they wrote about a tool migration. Gong also found that industry-based personalisation, meaning proof from similar companies, correlated with an 88% lift in replies, which makes it a useful fallback when no individual signal exists. These are Gong's own platform figures, not an independent trial.

What does a worked example look like, from signal to sent email?

A worked example shows each step, because each step can fail. Say you sell inventory software and the prospect is Priya, Head of Operations at a homeware brand. The Researcher finds a live job ad for an "Inventory Analyst" that mentions weekly spreadsheet reconciliation across three warehouses.

• Signal: the job ad, with its URL and the date it was found.
• Inference: reconciliation is manual and painful enough to hire for.
• First line: "Your inventory analyst ad mentions reconciling stock across three warehouses every week."
• Bridge: "Teams we work with usually automate that before they hire for it."
• Ask: "Worth a 15-minute look before you fill the role?"

Now the version a model can plausibly produce on a bad day:

Priya, loved your talk at RetailTech last month on warehouse automation. Your team's move to three new warehouses must be keeping you busy.

Both details sound right and neither is in the source. There was no talk, and the ad said three warehouses, not three new ones. The check that catches it is boring and reliable: for every claim about the prospect, the reviewer asks "where does this come from?" and opens the link. If the answer is "nowhere", the line goes. Jason Lemkin described the same discipline after SaaStr's AI SDR sent 4,495 emails in two weeks: read a sample of what it sent every day and ask "would I have written this email myself?" (SaaStr).

How do you stop an AI SDR from inventing facts or sounding creepy?

You stop invention with rules the agent can't skip and a person who checks samples. Require a source for every claim about the prospect, ban invented familiarity, use one signal per email and write an honest problem-led opener when no strong signal exists.

Our rules, in the order we'd enforce them:

1. Source every claim. Each fact about the prospect keeps the link it came from, so a reviewer can check it in seconds.
2. No invented familiarity. The email never says you read, attended, watched or met anything unless a person really did.
3. One signal per email. Three facts in a row reads like a dossier.
4. Freshness check. A signal past its window (see the table above) is treated as no signal.
5. Plain fallback. "Most ops leads at brands your size reconcile stock by hand" beats a forced compliment.

In Outbound Labs you approve every campaign before it sends, which is where these checks happen. Our guide to what to approve before an AI SDR sends covers the review itself.

How much does personalisation lift reply rates?

Published data points to a meaningful lift, roughly double in one vendor's numbers. Woodpecker reports that emails with advanced personalisation, meaning snippets beyond first name and company, average about 17–18% replies against 7–9% for basic or non-personalised emails (Woodpecker).

For a baseline, Instantly's 2026 benchmark across its platform puts the average cold email reply rate at 3.43%, with the top 10% of senders above 10%. It also found 58% of replies come from the first email in a sequence, which is why the opener carries so much weight (Instantly). Both are vendor datasets.

Personalisation also can't rescue an email that lands in spam. Google asks senders to keep the spam rate reported in Postmaster Tools below 0.3% (Google sender guidelines). Get the sending side right first with warmed mailboxes and the limits in our post on AI cold email deliverability. Outbound Labs caps sending at 40 emails per inbox per day and pauses a campaign automatically above a 2% bounce rate.

How do you test personalised emails when every line is different?

Test the signal type, not the sentence. Group emails by the signal they open with (job post, funding, new leader, fallback) and compare replies and meetings per group. A single first line is sent once, so it can never be tested. A signal type is sent hundreds of times, so it can.

The catch is volume. At a 3% baseline reply rate, you need about 750 sends per signal type to reliably tell 3% from 6% (95% confidence, 80% power, standard two-proportion test). To spot a smaller lift, 3% to 4.5%, you need about 2,500 per group. That's our arithmetic, and it changes how you plan:

If you send a few hundred emails a week, compare two signal types at a time, not five. Judge on meetings booked as well as replies, since a curious "who are you?" reply isn't pipeline.

When is AI personalisation the wrong tool?

AI personalisation is the wrong tool when your list is wrong, your offer is vague or you have fewer than a few dozen target accounts. A brilliant opener sent to the wrong person still gets no reply, and for a short list of big accounts a rep's own research beats any agent.

It's also weaker when the prospect has already shown intent. Someone who read your pricing page doesn't need a clever line about their job post; they need a fast, relevant follow-up, which we cover in what to send after a website visit. And if you only need a sending tool with your own copy, cold email software may suit you better than a full AI SDR.

If you want to see the research-to-draft flow on your own market, the trial covers 750 emails with no card. Start here and review the first campaign before a single email goes out.

Frequently asked questions

Sources

1. Gong Labs: 4 data-backed ways to increase your email reply rate (2023, checked 2026-10-17)
2. Woodpecker: Cold email statistics based on sending over 20M cold emails (updated 2026-06-23, checked 2026-10-17)
3. Instantly: Cold Email Benchmark Report 2026 (checked 2026-10-17)
4. Google Workspace Admin Help: Email sender guidelines (checked 2026-10-17)
5. SaaStr: We sent 4,495 AI SDR emails in 2 weeks (checked 2026-10-17)