The copy layer is everything that turns a lead row into words: template writing, first-line personalisation, full AI-written emails, and coaching tools that score drafts. Two years ago this was a product category of its own. By October 2026 every sequencer ships an AI writer, the models underneath are interchangeable, and the money has moved to the research step that feeds the prompt — dominated by Clay.
The uncomfortable finding for anyone selling "AI copy": no public, controlled evidence shows that AI-written cold email lifts reply rates. Vendors publish benchmark bands and testimonials, not experiments.
Who sits in the layer
| Tool | What it does | Price (as of Oct 2026) |
|---|---|---|
| Clay | AI columns and Claygent research over 150+ data providers; output pushed to any sequencer | Launch from $167/mo (15,000 actions, 3,000 data credits); Growth from $446/mo (pricing) |
| Twain | Per-prospect research and 1:1 sequences for email, LinkedIn, calls | Starter $49 (500 credits), Pro $101 (2,000); Base 1 credit/lead, High 2, Ultra 10+ (pricing) |
| Lavender | Real-time email coach in the inbox; new AI sales agent "Ora"; SOC2 (lavender.ai) | Not shown on homepage; /pricing returned 404 |
| Instantly | AI writer and AI Sales Agent drawing on shared Instantly credits | Credits from $47/mo for 1,500; bundles $85–555 (pricing) |
| lemlist | lemAgent and AI agents writing sequences | Included from $55/mo (pricing) |
| Salesforge | Personalisation credits, Agent Frank | 1,000 credits/mo on $80 Growth; Agent Frank from $499/mo (pricing) |
| Outreach | Personalization Agent on AI credits | Amplify Core and above, custom quote (pricing) |
| Woodpecker, Saleshandy, Reply.io | Built-in AI writers | Included or plan-gated (Woodpecker, Saleshandy, Reply.io) |
| Custom GPT/Claude prompts | DIY: CSV → model → CSV, or n8n workflows | Model API cost only |
Two notable absences: EmailBison ships no native AI and relies on spintax and Liquid templating (GTM Directory); and the sequencers' MCP servers (API, webhooks and MCP) mean the customer's own assistant can now write copy and push it straight into a campaign.
Copy.ai was in scope for this pass but its current cold-email offering and pricing were not verified (search budget exhausted). Smartlead's native AI writer was also not verified.
How the layer works
- Data — a lead row from Contact data and enrichment plus enrichment.
- Research — scrape the website, recent posts, job ads, news; Signals layer add timing.
- Prompt — a campaign-level instruction ("one sentence tying their hiring of SDRs to our offer").
- Generate — one variable per lead (
{{first_line}}) or a whole email. - Check — length, banned phrases, factual sanity. This step is usually missing.
- Insert — the sequencer merges variables; Spintax and message variants add surface variation.
Steps 4 and 6 are commodities. Step 2 is where Clay and Twain charge — Twain's "Ultra" research costs 10× its "Base" per lead — and step 5 is where almost nobody competes.
First-line personalisation
The dominant pattern in volume outbound is a human-written body with an AI-written opening line. It is cheap, keeps the offer and claims under human control, and limits hallucination damage to one sentence. Its weakness is sameness: when thousands of senders run near-identical "I saw that you…" prompts, the personalised line becomes a recognisable template — to recipients first, and plausibly to filters.
Full-body generation increases variety and lets the email follow the prospect's context, at the cost of brand control and a higher chance of a false claim.
What AI copy does to reply rates
What can be sourced:
- Reply-rate bands: Instantly calls 5–10% "solid" and 10–15% "excellent", citing Backlinko (~8.5% average) and a Belkins 2025 study of 16.5M emails where replies cluster "around the mid single digits" (Instantly blog). Other published 2025–26 averages run from 0.45% (Belkins) to 3.7% (Saleshandy) on different denominators. See Reply benchmarks.
- Length: the same source cites Boomerang research that 50–125 words correlates with higher response.
- Coaching: Lavender's site carries a customer quote of "twice as many replies" with no dataset (lavender.ai).
- AI reply handling (a different step) is where the strongest vendor claim sits: Instantly says its reply agent lifts booked meetings "20 to 40 percent" (Instantly blog).
No controlled public study (human vs AI copy, same list, same infrastructure) was found. The Instantly benchmark article itself has "no quantified impact metrics" for AI copy. Every uplift claim for AI personalisation in this market should be read as marketing until someone publishes an A/B with a disclosed sample.
List quality, timing and offer explain more of the variance in reply rates than wording. AI copy is most valuable where it is fed a real signal (a hire, a launch, a tech change) — which makes it a data problem wearing a copy costume.
How AI copy interacts with spam filtering
Filters score sender reputation, authentication and engagement before content (How filtering works, Reputation and blocklists). Content still matters at the margins: repeated near-identical bodies across many mailboxes are a classic bulk signal, and links, tracking domains and attachments add risk (Content and tracking).
AI copy cuts both ways:
- Helps: per-lead generation produces more unique bodies than spintax, weakening fingerprint-based bulk detection.
- Hurts: AI models converge on the same phrases and structures; mass-produced "personalised" emails can look more alike than vendors assume. And making each email cheaper to write encourages higher volume per domain, which is what Google and Yahoo sender rules and Microsoft rules: Outlook.com and Microsoft 365 limits punish through complaint rates.
No source in this pass measured how Google or Microsoft filters treat LLM-generated text specifically. The two points above are mechanism-level reasoning, not measured effects.
Instantly's guidance for AI-driven campaigns — 30 days of warmup, 30 sends a day per inbox, bounces under 1% (Instantly blog) — is a reminder that infrastructure discipline, not prose, decides placement.
Who keeps the money
- Model providers keep the generation margin. Clay exposes this directly with variable token pricing and bring-your-own-key options (Clay pricing).
- Clay keeps the research margin by sitting between 150+ data providers and the prompt.
- Sequencers monetise copy indirectly, through credit pools shared with leads and AI replies (Instantly) or plan upgrades (Saleshandy, PlusVibe).
- Point tools (Lavender, Twain) must keep moving toward agents — Lavender's "Ora" and Twain's agentic "Ultra" research show both doing so.
Legal and data angle
Personalisation means processing more personal data: scraping a prospect's posts and sending them to a US model provider. In the EU that needs a legal basis and documented processors (GDPR and ePrivacy, Data sourcing law); in Germany the cold email itself is already the bigger problem (Germany). An EU-hosted model path is a cheap differentiator nobody in this layer advertises.
What this means for an entrant
- Don't sell an AI writer. It is table stakes and the model providers own the margin.
- Make your sequencer the best destination for copy made elsewhere: arbitrary custom variables, per-lead preview, Clay and n8n push, MCP tools that write variables directly.
- Compete on step 5 — checking. Fact-verify AI lines against their source URL, block unsupported claims, enforce length and banned phrases. Nobody does this well.
- Let customers bring their own model key and choose an EU-hosted model; price sends, not tokens.
- Instrument the experiment the market lacks: built-in holdouts (AI line vs none) with honest reply and positive-reply reporting (Campaign analytics, A/B testing). Published results would be marketing no incumbent can match.