Use case · Outbound personalization

100,000 personalized emails on a model you run yourself.

A 7B model fine-tuned on your voice and your best-performing snippets turns enrichment data into two or three grounded sentences per contact — deployed on your hardware or VPC, wired into Clay or Lemlist. Your account data stays in your infrastructure.

Outbound personalization on a local model runs a fine-tuned 7B in your infrastructure that drafts grounded, on-voice personalization from your enrichment data. The task is narrow and repetitive — read the data, write two sentences in a fixed voice — which is exactly the profile where a small model matches a frontier one at a fraction of the cost. High-value accounts still route to a frontier model when the account earns it.

Personalization works. The invoice rations it.

Personalized outbound works, but the way most teams build it — three or four frontier-model calls per contact, chained through Clay — means the cost per email quietly climbs until "personalize everything" becomes a budget decision instead of a default. So teams ration it: personalize the top tier, template the rest. The task itself is narrow and repetitive, which is precisely the profile where a fine-tuned 7B model does the work for a rounding error.

How it works · The Crit Path
01Scope — codify what good looks likeWe take your best-performing campaigns and turn them into an explicit rubric and training set: which angles, which structure, what your voice sounds like, which claims the model may make. Personalization quality stops being a vibe and becomes a spec.
02Prove — fine-tune a 7B on your voiceAn open-weights 7B trains on your snippet structure and rubric, with rejection sampling against the rubric during training — the model learns your bar, not just your phrasing. It benchmarks against a frozen eval set before launch.
03Deploy — wire into the stack you already runThe model runs on a local GPU box or a VPC instance, exposed as an endpoint that Clay or Lemlist calls as the generation step. High-value accounts route to a frontier model instead — a per-request router decides, so you spend frontier money only where the account earns it.
04Assure — measure replies, retrain on winnersWe track variants to reply and positive-reply rate per segment. The training set grows from what actually worked, and the model retrains on it periodically.
The economics

A worked example, from public list prices. At 100,000 personalized emails a month — enrichment context in, two or three grounded sentences out — a multi-step frontier chain pays frontier prices three or four times per contact. All figures below are our estimates at published API list prices and measured self-hosted throughput, August 2026 — the ~250× gap is the arithmetic of those estimates, not a vendor benchmark. The Crit replaces them with your numbers.

Frontier API, multi-step chainFine-tuned local 7B
Per email~$0.025–0.15~$0.00004–0.0001
100k emails a month~$2,500–15,000Single-digit dollars of inference, plus a GPU instance at ~$300–700/mo cloud list price
Data exposureAccount and contact data processed by a third party on every callStays in your infrastructure
Published pattern66–80% lower inference cost on routine workloads — Forethought, on AWS; long narrow chains sit at the far end
Where this doesn't work

Cheap generation does not fix outbound. If list quality or deliverability is your bottleneck, a better-written email lands in the same spam folder — and The Crit will say so rather than sell you a model.

What you get

Your voice, as infrastructure.

  • A personalization spec: rubric, allowed claims, voice guide — derived from your winning campaigns, not invented
  • A fine-tuned 7B generation model, benchmarked against your rubric on a frozen eval set before launch
  • Deployment on your hardware or VPC, with an endpoint wired into Clay or Lemlist
  • An SLM-to-frontier router for high-value segments — frontier spend only where justified
  • A/B measurement plumbing: variant tracking to reply and positive-reply rate per segment
  • Retraining loop on winning variants, documented
Who this is for
50k+ emails a month through Clay or LemlistOutbound teams where per-contact frontier chains have become a visible line item.
GTM engineers who want to own the stackModel, rubric, and eval set in their hands — not rented per token.
Data that can't ride alongTeams with data-handling constraints on pushing enriched account data through third-party AI APIs.
Questions, answered straight

How much does it cost to personalize 100,000 outbound emails with AI?

Through a frontier API with typical multi-step chains, roughly $2,500–15,000 at published list prices; on a fine-tuned local 7B model, single-digit dollars of inference plus a GPU instance at $300–700 a month — our estimates, August 2026, a gap of around 250×. A rubric and a frozen eval set built from your best-performing campaigns hold the quality bar, not a bigger per-token price.

Is a 7B model good enough to write personalized cold emails?

For the narrow task of turning enrichment data into two or three grounded sentences in a fixed voice, a fine-tuned 7B model performs comparably to a frontier model — the task is repetitive and well-specified, which is where small models are strongest. Published fine-tuning work shows small models matching or beating larger ones on narrow tasks (distil labs × Knowunity: accuracy from 81% to 93%, vendor-reported). High-value accounts still route to a frontier model automatically.

Does AI personalization at scale hurt deliverability?

Generation cost and deliverability are separate problems: a local model makes each email nearly free to personalize, but sending volume, domain health, and list quality still decide whether it lands. Varied, grounded personalization generally helps engagement signals compared to static templates, but if deliverability is the bottleneck we address that before scaling generation — cheaper drafts of an email nobody sees are worth exactly nothing.

Security · by architecture

Your CRM never feeds Big AI.

  • Pipeline data stays in your VPC
  • No third-party enrichment APIs
  • Your DPA stays unchanged
  • Prospect data never resold
Architected to deploy inside your
GDPR CCPA SOC 2
Small models · critical hits

Personalize everything. Ration nothing.

Two weeks inside your campaigns and your stack, and you'll know what a model trained on your winners can draft — and what it costs to run. The numbers are yours either way.