Per-ticket AI costs that survive retail margins.
Small, fine-tuned models for tier-1 support, catalog search, and product-data cleanup — because at retail volume, frontier-API pricing is a margin problem, not a rounding error.
Retail AI survives on unit economics. A fine-tuned small model answers tier-1 tickets, powers catalog search, and normalizes product data at a flat infrastructure cost, with a router escalating the genuinely ambiguous cases. The reference evidence comes from support automation itself: Forethought cut inference costs 66–80% moving routine workloads to fine-tuned small models (published on AWS).
Retail AI lives and dies on unit economics. Support tickets, catalog queries, and product-data fixes arrive in the hundreds of thousands, margins are thin, and a per-request frontier-API fee that looks harmless in a demo compounds into a real line item at volume. The workloads themselves are narrow: order-status questions, returns policy, size and availability lookups, attribute extraction from supplier feeds, title and category normalization. Narrow, high-volume, repetitive — the exact profile where a fine-tuned small model wins.
The reference evidence comes from support automation itself: Forethought, an AI support company, cut inference costs 66–80% by moving routine workloads to fine-tuned small models (published on AWS). The architecture we build follows the same logic — a local model handles the routine bulk of tier-1 traffic, and a router escalates the genuinely ambiguous cases to a frontier model or a human agent. The same extraction pattern cleans supplier feeds and powers natural-language catalog search without a per-query API bill.
We haven't shipped this vertical yet; The Crit is how we'd prove the numbers on your ticket history before you commit.
Support economics, fixed at the model layer.
Retail's constraints are consumer privacy and payment data. Customer conversations fall under GDPR and CCPA, and the simplest posture is to keep them on your own infrastructure rather than adding another processor to the disclosure list. Payment details are simpler still: they should never appear in a model prompt at all, and we design pipelines so they can't — keeping your AI systems entirely outside PCI scope. We're engineers, not your compliance advisers — we just keep the disclosure list from growing.
We haven't shipped this vertical yet. Here is the published evidence the economics hold — led by a result from support automation itself — and here is how we'd prove it on your ticket history before you commit.
| Published result | Source |
|---|---|
| Inference costs fell 66–80% after moving routine support workloads to fine-tuned small models | Forethought, published on AWS |
| Task accuracy rose from 81% to 93% while inference costs fell 50–68% | distil labs × Knowunity, vendor-reported |
| Serving costs fell roughly 50% on dedicated small-model infrastructure | distil labs on Cerebrium, vendor-reported |
| Task-specific small models to see 3× the adoption of general-purpose LLMs by 2027 | Gartner forecast |
What does tier-1 support automation cost at retail volume?
A flat infrastructure cost instead of a per-request fee. A fine-tuned small model on your own hardware handles the routine bulk of tier-1 traffic — order status, returns policy, size and availability — at a marginal cost per ticket that rounds to zero. The reference evidence comes from support automation itself: Forethought cut inference costs 66–80% moving routine workloads to fine-tuned small models (published on AWS).
Can the same model clean our product catalog?
Yes — the same extraction pattern normalizes titles and categories, pulls attributes out of supplier feeds, and powers natural-language catalog search without a per-query API bill. Narrow, high-volume, repetitive: the exact profile where a fine-tuned small model wins.
What about payment data?
Payment details should never appear in a model prompt at all, and we design pipelines so they can't — keeping your AI systems entirely outside PCI scope. Customer conversations stay on your own infrastructure, which also keeps the GDPR and CCPA disclosure list short.
Customer data stays home.
- Conversations never leave
- Card data can’t reach a prompt
- No new subprocessor
- Region-pinned by design
See if the numbers hold at your volume.
The Crit is a teardown of your AI spend and workflows: where a small model wins, where an API is fine, and where AI shouldn't be used at all. You get the numbers either way.
If your workload doesn't clear roughly 50M tokens a month and you have no privacy constraint, a frontier API is probably fine — and we'll tell you so.