Will AI write product copy that sounds like our brand?

Only if the brand voice can be written down and evaluated. We build a labelled set of good and bad examples first; if the output cannot clear that bar consistently, we recommend keeping copywriting human.

Does anything publish to our store automatically?

No. Generated output lands in a review queue with a visible diff, and a person publishes it. Automatic writes to a live storefront are not something we build.

How do you stop it inventing product details?

Output is grounded in your own catalogue, specifications and policy content through retrieval, and the evaluation set specifically tests for fabricated attributes before launch.

What does this cost to run?

We model token and latency cost per item during scoping and enforce budgets in the pipeline, with caching and cheaper-model fallbacks so cost per catalogue run stays predictable.

Is our data used to train models?

We use API tiers that exclude your data from training, and we document exactly which data leaves your systems and where it goes before anything is connected.

That is the only way we do it. One narrow use case, an evaluation set and a measurable saving first — expansion is a decision you make with evidence, not upfront.

Service · Optimise & Scale

AI Commerce & Automation

Narrow, evaluated AI applied to merchandising, catalogue enrichment and support workflows. No autonomous agents writing to your live store.

OpenAI / Anthropic APIs
Embeddings & vector search
Node.js
Shopify Flow

Group: Optimise & Scale
Starts with: Technical audit
Delivery: Git + staging theme
Handover: Written documentation

01Overview

Most AI proposals for Shopify merchants start from the technology and work backwards to a use case. We start from the operational cost. If a task is high volume, repetitive, and currently done by a person copying between systems, it is a candidate. If it needs judgement, brand voice nobody has written down, or carries legal risk, it stays with a human.

Everything we ship is evaluated before it reaches customers. That means a labelled test set, an agreed accuracy threshold, and a measurement you can repeat after a model or prompt change. Without that, quality drifts silently and nobody notices until a customer does.

Nothing writes to the storefront unattended. Generated catalogue copy, attributes and recommendations land in a review queue with a clear diff and a one-click revert, so your merchandising team stays accountable for what customers actually see.

02The problem

AI in commerce is mostly demos. The value sits in unglamorous places: catalogue data quality, internal search relevance, support deflection and operational triage — work that is repetitive, measurable, and expensive when done by hand.

03You will recognise this if

Catalogue copy and attributes are entered by hand
Support answers the same questions repeatedly
Merchandising decisions rely on manual review
Product data quality varies by whoever entered it
An AI pilot produced impressive demos and no measurable saving

04How we work on it

01
Pick boring problems
We target repetitive, high-volume work with a measurable cost, not headline use cases. The first question is always what this currently costs in hours.
02
Evaluate before shipping
A labelled evaluation set and an accuracy threshold are agreed before anything reaches customers. If the threshold is not met, the feature does not ship.
03
Human in the loop
Generated output is reviewable, diffable and reversible. Publishing to the storefront is an explicit action by a person, never a side effect.
04
Grounded, not guessed
Answers and enrichment are grounded in your own catalogue, policies and help content through retrieval, which is what keeps output factual rather than plausible.
05
Cost control
Token and latency budgets, caching and cheaper-model fallbacks so the feature stays economical at catalogue scale, not just in a pilot.
06
Monitoring and rollback
Quality, spend and failure rates are tracked after launch, with a documented path to turn any of it off without touching the theme.

04What you get

Deliverables, not decks.

Use-case assessment with an hours-and-cost model
Labelled evaluation set and agreed accuracy thresholds
Retrieval layer grounded in your catalogue and policies
Production pipeline with a human review UI
Shopify Flow and Admin API automations
Quality and spend monitoring dashboard
Prompt and model change process
Documented rollback path

ai/enrich-catalogue.ts

1// field-data first — never lab scores alone2const budget = { lcp: 1800, tbt: 120, cls: 0.05 };3 4onCLS(send); onINP(send); onLCP(send);5// regressions fail the deploy, not the quarter

Delivery pipeline

Audit
Scope
Build
QA
Measure

05What changes

Hours returned: Repetitive catalogue and support work moves to a reviewed pipeline, so your team spends its time on decisions rather than data entry.
Measurable quality: Accuracy is a number against a fixed evaluation set, not an impression, which makes model and prompt changes safe to make.
Controlled spend: Token, latency and failure budgets are monitored from day one, so cost scales with the catalogue predictably.

06How an engagement runs

01
Audit
We map the operational workflows that are actually expensive and score them for volume, risk and measurability. You get a shortlist with a cost model attached.
02
Scope
One use case is chosen, with an evaluation set, accuracy threshold and cost ceiling agreed in writing before build starts.
03
Delivery
Built with a review interface from the first iteration, evaluated against the test set, then monitored for quality and spend after launch.

07Stack

OpenAI / Anthropic APIs
Embeddings & vector search
Node.js
Shopify Flow
Admin API

09Questions

Send us the store and the symptom.

A senior engineer reads it, tells you what is actually going on, and scopes only what the evidence supports.

Book Consultation Request Shopify Audit

AI Commerce & Automation

Pick boring problems

Evaluate before shipping

Human in the loop

Grounded, not guessed

Cost control

Monitoring and rollback

Deliverables, not decks.

Audit

Scope

Delivery