Use case · Data enrichment

Normalization for fractions of a cent, as a node in your existing flow.

A small model fine-tuned on your taxonomy — job titles, seniority bands, industries, ICP categories — exposed as an HTTP endpoint you drop into Clay or your own pipeline. Built for the runs where per-record API pricing stops making sense.

A data enrichment engine is a 1–3B model fine-tuned on your taxonomy and exposed as a plain HTTP endpoint inside your Clay flow. It answers in your categories — seniority bands, functions, industries, ICP tiers — deterministically, at hundredths of a cent per record (our estimate at public list prices, August 2026), with per-category accuracy measured on a frozen holdout first.

The unglamorous middle step, a million times a month.

Every GTM data pipeline has the same middle step: turning "VP, Rev Ops (EMEA)" into a seniority band and a function your routing actually uses. Teams do it with brittle regex, with spreadsheet passes, or with a frontier model that charges real money to answer a 40-token question a million times. The task is narrow, repetitive, and completely defined by your taxonomy — which makes it exactly the kind of work a small model does better and radically cheaper.

How it works · The Crit Path
01Scope — fix the taxonomyWe take your actual routing and segmentation logic — seniority bands, function map, industry list, ICP tiers — and turn it into an explicit target schema. We settle ambiguous cases once, in writing, instead of re-guessing them on every run.
02Prove — fine-tune a 1–3B on your labelsA few thousand examples from your own database, labeled against the schema, train a model that answers in your categories — not in a generic model's idea of them. We measure accuracy against a frozen holdout before anything ships.
03Deploy — one endpoint, drop-in nodeThe model runs in your VPC, or as a managed endpoint we operate if you don't want the infrastructure. It's a plain HTTP call: record in, normalized fields out — a drop-in node for Clay, or one function call in your own code.
04Assure — measure and extendPer-category precision dashboards show you where the model holds. When your taxonomy changes — new segment, new market — the model retrains against the updated schema instead of accumulating prompt duct tape.
The economics

A worked example, from public list prices. At 1,000,000 records normalized a month — well under 100 tokens per record — per-call pricing turns a 40-token question into a real line item. Both columns are our estimates at public list prices, August 2026; The Crit replaces them with your numbers. Gartner expects task-specific small models to be adopted three times more than general-purpose LLMs by 2027.

Frontier APIFine-tuned 1–3B model
Per record~$0.002–0.01 at published list prices, prompt overhead includedHundredths of a cent — one mid-range GPU handles tens of records a second
1M records a month~$2,000–10,000Flat instance, ~$300–700/mo at cloud list prices, or batch runs for less
ConsistencyGeneral model re-interprets your taxonomy on every callTrained on your taxonomy: deterministic schema output
Where this doesn't work

If you normalize a few thousand records a month, keep using a frontier model inside Clay — the math doesn't justify a build. This engine pays for itself on recurring six- and seven-figure record volumes, and The Crit will tell you which side of that line you're on.

What you get

One endpoint, one schema, one number.

  • A written normalization schema: your taxonomy, edge cases adjudicated, versioned
  • A fine-tuned 1–3B model with per-category accuracy measured against a frozen holdout
  • A deployed HTTP endpoint — in your VPC or managed — with authentication and rate limits
  • Ready-made integration: a Clay HTTP-node template and an HTTP integration recipe, documented
  • Batch mode for backfills — re-normalize the whole database on taxonomy changes
  • Precision monitoring per category and a retraining path when your taxonomy evolves
Who this is for
Clay at volumeGTM engineering and RevOps teams whose enrichment volume has made per-call pricing a line item someone noticed.
Routing that breaks on messy titlesTeams whose routing breaks on inconsistent titles and industries — and who are tired of regex archaeology.
Normalization as owned infrastructureData teams that want one endpoint, one schema, one accuracy number — versioned and theirs.
Questions, answered straight

What's the cheapest way to normalize job titles and industries at scale?

A small fine-tuned model (1–3B parameters) is typically the cheapest reliable option at scale: it handles a normalization call in under 100 tokens at a marginal cost of hundredths of a cent per record, versus roughly $0.002–0.01 per record through a frontier API at published list prices (our estimate, August 2026). At a million records a month, that's the difference between a flat ~$300–700 GPU instance and a $2,000–10,000 API bill.

Can I use a custom AI model as a node in Clay?

Yes. The enrichment engine deploys as a plain HTTP endpoint, which Clay consumes through its HTTP API column and any other tool through a standard HTTP request — we ship documented templates for both. The model runs in your VPC or as a managed endpoint, and returns fields in your own taxonomy rather than a generic model's guess at it.

Is a small model accurate enough for data normalization?

For a closed-taxonomy task like mapping titles to seniority bands or companies to industry categories, a small model fine-tuned on your labeled data is usually more consistent than a large general model, because it trains to answer only in your categories. We measure accuracy per category against a frozen holdout before deployment — you see the number, and the honest failure cases, before you rely on it.

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

One endpoint. One schema. One accuracy number.

Two weeks with a sample of your records, and you'll know the per-category accuracy and the per-record cost — before you re-route a single workflow. The numbers are yours either way.