Large language models changed cold email in three waves. First (2023–24) they made "personalised" first lines free, and buyers bought AI SDRs on the promise of infinite SDR capacity. Second (2025) the bill came due: 11x was reported to have lost 70–80% of early customers, and the industry learned that more AI-written email is not more pipeline. Third (2026) AI moved to the recipient's side: Gmail now summarises threads for every user and is testing an "AI Inbox" that filters low-priority mail.
The finding: AI lowered the cost of producing an email to near zero, so the scarce resources are now the recipient's attention, the sender's reputation and the reason to write at all. A new platform should be built around those three, not around generating more text.
Layer by layer
| Layer | Before LLMs | With LLMs (2026) | Who captures value |
|---|---|---|---|
| Contact data and enrichment | Buy a list, verify it | Agents research each row on the web (Claygent); waterfalls across providers | Clay (est. $150M ARR, May 2026) |
| Signals layer | Intent data for enterprise ABM | Signals feed automated plays; cheap self-serve visitor ID | Clay, Unify, RB2B, Warmly |
| Copy and personalisation layer | Templates + spintax | Per-lead copy; per-signal copy | Commoditised — every tool ships it |
| Sequencing & sending | Human builds campaign | AI campaign builders; agents run whole campaigns | Sequencers absorbing AI-SDR features |
| Inbox & domain infrastructure | A few inboxes per rep | Volume rises as generation cost falls; more inboxes and domains | Inbox resellers, until providers tighten |
| Replies, booking and handoff | SDR reads every reply | AI classifies, drafts and routes replies | Built into sequencers (see AI reply agent) |
| Recipient inbox | Subject line + preview | AI summaries, priority filtering, suggested replies | Google, Microsoft, Apple |
1. Personalisation at scale stopped meaning anything
When every tool can write "Saw your post on X — loved the point about Y" for every lead, the line stops signalling effort. Practitioner criticism of AI SDRs quoted by data vendor Prospeo sums it up as "more volume, less signal". Instantly's 2026 benchmark (2025 data, "700k+ businesses") recommends emails under 80 words, one call to action and problem-first openings — short, plain email, not elaborate personalisation.
Whether Gmail or Microsoft filters detect LLM-written text as such could not be verified. No primary source from either provider says so. What is documented is that filters act on engagement and complaints (see How filtering works and Google and Yahoo sender rules) — so AI copy hurts when it drives complaints or silence, not because it is AI.
2. Volume inflation
Cheap generation plus cheap inboxes push senders toward more volume. AI SDR plans are sold by monthly volume — Artisan's Team plan is ~2,500 leads a month, Scale ~6,000; AiSDR's tiers cover 1,200–4,500 AI messages a month. An operator analysis claims AI-SDR programmes hit spam-complaint rates of 0.2–0.3% by weeks two to three and see reply rates fall to 0.5–1.5% after week six (leadgen-economy) — unverified, but consistent with how provider rules punish complaints (see Provider crackdowns).
The counter-movement is signal-based outbound: fewer, timelier sends triggered by events. Unify dropped to $20–60 per seat in 2026, and Clay now sends from its own sequencer. Both point to a market where the winning sender sends less.
3. Reply classification became table stakes
Classifying replies (interested, not now, wrong person, out of office, unsubscribe) is the most reliably useful LLM job in outbound: bounded, checkable, high volume. Clay's sequencer ships sentiment analytics and Slack approval for replies; Artisan handles replies autonomously with escalation rules. The harder job — answering objections and pricing questions — is where the 2025 complaints about AI SDRs concentrate. See Replies, booking and handoff and Unified inbox (Unibox).
4. Agents running whole campaigns
The autonomous AI SDR — 11x, Artisan, AiSDR — was the bet that one agent could do list-building, copy, sending and replies. The evidence (see AI SDR layer) shows heavy early churn and a 2026 retreat to approvals and human handoff. The agent is being absorbed into platforms from both sides:
- From the data side. Clay's Sequencer, generally available from 1 September 2026 per Clay's blog, runs on Smartlead's sending engine and warmup pool (Clay docs). Clay's July 2026 changelog added open-weight models (Kimi K2.6, GLM 5.2) to Claygent for cheaper research (changelog) — model cost is now a line item vendors optimise.
- From the sending side. Instantly, Smartlead, Apollo.io and Amplemarket ship AI campaign builders and reply agents (see AI campaign builder).
- From the CRM side. HubSpot took on Trigify's team in September 2026, bringing "signals, data and context" expertise in-house; see HubSpot Sequences.
5. The recipient has AI too
On 8 January 2026 Google announced that Gmail is "entering the Gemini era":
| Feature | What it does | Who gets it |
|---|---|---|
| AI Overviews | Summarises threads | Free for all users |
| Suggested Replies | Context-aware one-click replies | Free |
| Help Me Write | Drafts emails | Free |
| AI Inbox | "Filters out low-priority messages", highlights VIPs based on communication patterns | Paid (AI Pro/Ultra), trusted testers first |
| Proofread, inbox Q&A | Grammar; questions across the inbox | Paid |
Rollout started in the US in English, with more languages and regions to follow.
What this means for a cold email: a summary may be read instead of the email, so padding and flattery are stripped before a human sees them; and a prioritised inbox keyed to "communication patterns" favours people you already correspond with — by definition not a cold sender. Suggested replies may also make short "not interested" answers easier to send, which helps list hygiene but cuts the silent maybes.
Apple Intelligence also summarises and prioritises mail on Apple devices, but no Apple source could be fetched for this page to confirm current behaviour and availability. Microsoft Copilot's equivalents in Outlook were not researched here.
No data yet shows how Gemini summaries or AI Inbox change cold-email reply rates. Open tracking already undercounts (see Content and tracking); AI summaries will make opens even less meaningful as a metric.
Where money and power move
- Generation is worth nothing. Every sequencer ships AI copy; LLM cost per email is a fraction of a cent. Charging for "AI writes your emails" is not a business. AiSDR at ~$0.56–0.75 per message is the visible price that will be undercut.
- Context is worth more. Clay's valuation went from $3.1B (August 2025) to $7.1B (September 2026) on the back of research and data workflows, not sending.
- Reputation is the scarce input. If AI inflates volume and inboxes filter harder, the sender that protects domain reputation and complaint rates wins the right to reach a human at all. See Reputation and blocklists and Volume math: what 10k and 100k emails a day cost.
- Trust is a market. The 11x scandal and the unsourced "50–70% churn" stats circulating through vendor blogs (see AI SDR layer) mean buyers discount vendor claims. Verified measurement is a selling point.
AI did not make cold email easier; it made it cheaper to do badly. The opening for a new platform is to be the system that makes each send more defensible — triggered by a real reason, checked by a human, sent from a protected reputation, measured on replies — rather than the one that sends the most.
What this means for an entrant
- Treat AI copy as a free feature, not the product. Bundle it; compete on what it cannot do.
- Build for triggers and approval. Webhook-in (Clay, RB2B, CRM), draft per trigger, one-click human approval, send within the hour. This serves the signal-based buyer the volume tools were not designed for — see Signals layer.
- Write for the summary. Offer a pre-send check that shows what Gemini-style summarisation would keep from an email; make short, plain, specific copy the default.
- Measure replies and meetings, not opens. Inbox AI and privacy features erode opens further; build analytics around positive replies and booked meetings.
- Be the engine AI tools license. Clay licensed Smartlead's sending. AI SDRs and signal tools need deliverable, API-first sending; that is a B2B2B channel a new entrant can win (see API, webhooks and MCP and White-label and agency portals).
- Make volume restraint a feature. Per-domain caps, complaint-rate kill switches and CRM-aware suppression protect customers from the AI-era failure mode — and protect the platform from Provider rules: Google, Microsoft and the ESPs trouble.