AI personalisation uses a language model to write a lead-specific line or whole email from data about the prospect — their title, company, website, posts, job ads. Buyers care because generic templates get filtered and ignored, and because writing 1:1 lines by hand does not scale past a few dozen a day.
In October 2026 every sequencer in the feature matrix except EmailBison offers some version of it. The strategic question is no longer "does the tool have an AI writer" but who owns the research step that makes the line worth reading. The broader picture is in Copy and personalisation layer.
How it works
- Research — scrape or enrich: website, LinkedIn activity, news, hiring signals (Signals layer).
- Prompt — a template prompt per campaign ("write one sentence referencing their latest hire…").
- Generate — per lead, stored in a custom variable like
{{ai_first_line}}. - Insert — the sequencer merges it into the template; Spintax and message variants add further variation.
The sequencers bill step 3 (and sometimes 1) in credits. Clay bills both: fixed-price per AI task or variable token pricing for models like GPT-5.1, and bring-your-own API key runs cost one action and no data credit (Clay pricing, October 2026).
Who does it well
| Player | AI personalisation (as of Oct 2026) |
|---|---|
| Clay | AI columns and Claygent research; Launch from $167/mo (15,000 actions, 3,000 data credits) (pricing) |
| lemlist | "lemAgent" and AI agents included from $55/mo Email plan (pricing) |
| Outreach | "Personalization Agent" on Amplify Core and above, consuming AI credits (25,000 on Core) (pricing) |
| Salesforge | 1,000 personalisation credits/mo on the $80 Growth plan (pricing) |
| Instantly | AI features draw from Instantly credits shared with leads and replies (pricing) |
| Saleshandy | 100 one-time AI credits on Starter; full on Pro ($99/mo) (pricing) |
| Woodpecker | "AI email writer" in the base price (pricing) |
| Reply.io | "Full suite" AI on Multichannel; hyper-personalisation in Jason AI SDR (pricing) |
| EmailBison | None — spintax and Liquid templating instead (GTM Directory) |
Smartlead's and Apollo's native AI writing features were not verified in this pass. They are marked "unknown" in the matrix rather than guessed.
Where implementations differ
- Research depth. Twain sells depth explicitly: Base 1 credit per lead, High 2, "Ultra" 10+ credits for an agentic thesis per prospect (Twain pricing). Most sequencer writers only see the CSV columns you uploaded.
- Billing unit. Credits per generation (Instantly, Salesforge, Saleshandy), credits per agent action (Outreach), or included flat (lemlist, Woodpecker). Shared credit pools — Instantly's pays for data, verification and AI replies alike — make cost hard to predict.
- Model choice. Clay lets users pick the model and bring their own key. The sequencers mostly hide the model.
- Placement. First-line only (merged into a human template) versus full-body generation. Full-body increases variation, which helps against content fingerprinting, but raises the risk of off-brand or false claims.
Risks and abuse
No controlled, public study comparing AI-written with human-written cold email reply rates was found. Instantly's benchmark guide gives reply-rate bands but "no quantified impact metrics" for AI copy (Instantly blog). Lavender's homepage offers only a testimonial of "twice as many replies" (lavender.ai).
- Hallucination. A wrong fact about the prospect ("congrats on the Series B") is worse than no line.
- Sameness. When thousands of senders use the same "I noticed you…" prompt pattern, the AI line becomes a template again — see Content and tracking.
- Data protection. Scraping and sending a prospect's personal data to a US LLM is processing that needs a legal basis and a processor agreement under GDPR and ePrivacy.
What this means for an entrant
- Don't sell "AI writer" as a feature. It is table stakes and the commodity is the model.
- Do make the variable pipeline first-class: custom fields from Clay or any webhook, preview per lead, and reject-before-send rules (length, banned claims).
- Let users bring their own model key, as Clay does; price sending, not tokens.
- If you want a wedge here, it is research quality and verification of the AI line (fact-check against the source page), not generation.
- An EU-hosted model option is a credible differentiator for German buyers (Germany).