AI for fashion ecommerce, built on the client book.
Fashion ecommerce margin lives in the client book — and leaks out through the returns queue. We build the AI agents that work both: your clienteling rules, sizing conventions, and outreach tone, encoded inside the data and email stack you already run.
Official services partner of the platforms defining AI
NVIDIA
Anthropic
Fashion ecommerce margin leaks one return — and one lapsed client — at a time.
From client-experience VPs, styling-team leads, and data leads inside a luxury fashion e-commerce retailer — designer apparel, trunk shows, personal-styling tiers, ~12.5K client profiles under management, and a Snowflake-and-Looker data spine with Klaviyo carrying the sends.
Returns are the board’s number one opportunity
“Mitigating returns is our number one opportunity.” The follow-up window is 48 hours, and the outreach should offer the next size down or a similar piece in the client’s palette — not re-show the item she sent back. Half of it is sizing: UK 10, French 34, Italian 38, and designers who don’t conform to standard size-scale conventions, so her most consistently purchased and kept size has to be computed per designer.
Top clients drift before anyone notices
High value drift: a top-spending client starts going quiet — not lost yet, but drifting. Stylists carrying ~350-client books spend a disproportionate part of their time in a very small group and miss the rest; a concierge rep carrying ~2,000 can’t see it at all. And the $5K tier line misses people — there are clients below the threshold who might be billionaires, and a household name can buy on the site without anyone being told.
Outreach that can’t read like AI wrote it
“Make sure this doesn’t sound like I’m speaking to someone I don’t know” — it should not look like it’s written by AI. And if the queue raises the cognitive load — what is the action, is this accurate — stylists will just close it and do what they usually do. The pilot can’t be a feedback loop that goes on forever; land the plane.
The $1–6-a-profile enrichment tax
~12.5K client profiles under management, ~3,000 enriched nightly, at $1–6 a profile — that $75,000 figure, run daily. The target is a private deployment for about a thousand dollars a month on open-source models — 90–95% cheaper — because in terms of client information and security, the data isn’t going anywhere.
A data spine held together by hand
The warehouse is kind of chaos — the orchestration layer got built by hand. The dashboards carry site-credit values 24 hours late, and the email platform alone carries 617 distinct sends across 584 scheduled campaigns — which is how clients get emailed about pieces that aren’t available.
Where an AI agent for ecommerce pays first.
AI in fashion ecommerce doesn’t pay on the homepage — it pays in the book. Each card is a workflow or capability built on the same pattern: your clienteling rules, sizing conventions, and outreach tone, encoded inside the systems you already run.
Mitigating returns is our number one opportunity.
Led by Peter Enestrom.
Asked by client experience and styling leads.
The straight answers, before you book anything.
It reads the signals your stack already produces — returns, carts over $10,000, back-in-stock pieces, expiring site credit, who purchased this day last year — and turns them into a daily opportunity queue per stylist: who to contact, why now, and a draft in that stylist’s voice. What the rules resolve gets queued; what they can’t goes to a person. No chatbots, no site widgets — conversational commerce answers anonymous traffic; this works the client book.
The engine personalizes the site for traffic; it doesn’t work the book. It doesn’t know she kept the Italian 38, that trunk-show season is when she spends, or that she’s going to the Hamptons next month. A personalized shopping experience at this level is 1:1 outreach from a named stylist — a different motion, run on your client data, not on clickstreams.
Tone is the build, not an afterthought. Drafts train on each stylist’s own sent edits, kept-item history, and palette notes, and the stylist approves anything before it sends. “Make sure this doesn’t sound like I’m speaking to someone I don’t know” is the spec — AI email personalization that reads like the person the client actually knows, or it doesn’t ship.
No. Enrichment and drafting run on open-source models in a private deployment inside your environment — 90–95% cheaper than per-profile API pricing, with a four-figure monthly run cost as the sizing target. In terms of client information and security, you’re not sending the data anywhere; that’s the requirement and the architecture.
It’s the normal starting point. The orchestration layer over the warehouse gets built as part of the engagement; site-credit values that land 24 hours late and an email platform carrying 617 distinct sends are inputs to encode, not blockers. The system reads what your stack actually produces.
Two weeks inside one tier’s outreach and returns flow, fixed fee: we map the rules, quantify the leak in your own numbers, and hand you a ranked opportunity map — yours whether or not you continue. From there the build is scoped to land the plane: a pilot that ships and gets used, not a feedback loop that goes on forever.
Start with one workflow.
Tell us where your team loses hours. We will come back with a straight answer on whether AI can help, what it would take, and what it would pay.