AI in ecommerce,
shipped — not slide-decked.

`Agentic commerce` is the shift from rule-based retail tech to multi-step AI agents that decide what to do at the per-request level — pick a recommendation, draft a recovery email, propose a re-order quantity, route a fraud signal. The agent observes a signal, enriches it with your data, decides, acts on a real system (Shopify, Klaviyo, NetSuite), and logs the trace. We're shipping production agentic-commerce workflows today on Shopify, BigCommerce, Magento, and WooCommerce — the constraint isn't model quality anymore, it's eval discipline and bimodal-load handling. If you skip the eval suite and the peak-mode runbook, you ship a demo that breaks on the first sale event.

It depends on your stack and SKU count, but the three patterns that consistently pay back in the audits we run are (1) abandoned-cart recovery agents on AOV >$120 — generic recovery is mostly ignored, personalized recovery clears 15–25% lift on the high-AOV cohort; (2) `ai inventory management` on long-tail SKUs (the 80% of SKUs your planner team can't re-buy weekly), where stock-out reduction translates directly to revenue; and (3) Tier-1 customer-support deflection with order context — the WISMO + returns workflow that's 60% of inbound. AI search and dynamic pricing pay back too but are slower to ship and need more guardrails. Recommendations look like the obvious win in slide decks; in practice they're the hardest to attribute uplift on.

`AI inventory management` is not a model staring at a stock table. The pattern we ship is a weekly-cadence agent that ingests four streams — 90-day sales velocity per SKU, supplier lead-time and MOQ, marketing-calendar overlap (promos, campaigns), and last-year same-week curve (or weather signal for relevant categories) — and drafts a buying plan grouped by supplier. A planner reviews, edits, or rejects line-by-line; the agent learns from the edits. Margin-floor and cash-flow guardrails are non-negotiable; the agent never executes a buy autonomously. Run cost is typically $400–$900/month for a 4–10K-SKU catalog including model + warehouse spend. The win isn't replacing the planner — it's letting one planner cover 5× the SKU tail they can today.

Buy if your support is mostly WISMO + returns + FAQ on a catalog under 2K SKUs and you don't have an engineering team. The off-the-shelf `ecommerce chatbot` platforms (Gorgias AI, Zendesk Answer Bot, Tidio) are fine and cheap on that profile. Build (or hire us to build) if any of: (1) you have a catalog over 5K SKUs and need real product retrieval, not a FAQ matcher; (2) your tone and brand voice matter — every off-the-shelf bot sounds the same and your customers will notice; (3) you're on Magento / headless / multi-storefront and the off-the-shelf integrations are weak; (4) you want the conversation data in your warehouse, not a vendor's. An `ai chatbot for ecommerce` built on Claude or GPT with order-history grounding typically runs $400–$800/month in model spend versus $1,500–$5K/month seat fees on the platforms — and you own the prompts.

A `shopify chatgpt integration` (or Claude variant — the pattern is identical) has three layers. (1) The signal layer is Shopify Webhooks plus Admin API polling for the events the agent reacts to (cart-abandoned, order-created, refund-initiated). (2) The decide layer is a small service — we usually run it on Cloudflare Workers or a Lambda — that holds the prompt, calls the model, and resolves the next action. (3) The write layer is Shopify Storefront / Admin API, plus your marketing stack (Klaviyo, Postscript) for the customer-facing send. Don't put the agent inside a Shopify theme app unless the merchandiser UX is the point — for most retail use cases the agent lives off-platform and the integration is webhooks + API. Token cost on a Shopify-grounded loop with prompt caching warm averages $0.001–$0.005 per decision depending on model and context size.

`Dynamic pricing AI` on long-tail SKUs runs in two cost buckets. The signal layer (competitor scraping, elasticity history, margin lookup) is typically $80–$200/month of infrastructure regardless of catalog size — it's the same scrapers and warehouse views whether you have 1K or 20K SKUs. The model layer scales with how often you reprice: at weekly cadence on 10K SKUs using Haiku 4.5 with prompt caching, model spend is roughly $0.0002 per pricing decision, so ≈ $8/month of model spend for the whole catalog. The honest cost ceiling is engineering: getting the data clean, getting the margin-floor guardrail tight, getting the price-change diff into your PIM. Total all-in is usually $400–$1,200/month for a mid-market catalog. The pattern only pays back on tail SKUs — hero SKUs stay manually priced.

Three things, all of which need to be in place — pick any two and you'll quietly contaminate your catalog. (1) The input is structured spec data, not just an image — material, weight, dimensions, use-case, target customer. Image-only generation drifts within 100 SKUs. (2) The prompt includes 3–5 voice examples from your existing best-rated descriptions, refreshed quarterly. (3) Every generation runs through a brand-tone classifier (we usually train a small classifier on 500 of your existing descriptions versus 500 negative examples) and only descriptions above threshold ship; the rest go to a copywriter queue. With those three, an `ai product description generator` workflow holds tone over thousands of SKUs. Without them, the output reads generic by SKU 300 and your category pages start ranking worse, not better.

Two weeks for the audit, then six to eight weeks to first workflow live behind a feature flag — assuming we're not waiting on platform access. The breakdown: weeks 3–4 are integration scaffolding (webhooks, API auth, eval harness), weeks 5–6 are building the decide step and tuning the model picks, week 7 is bimodal-config testing (steady-state + simulated peak), week 8 is production hardening and the runbook. Narrow workflows — say, a product-description generator on a clean catalog — can ship in 3 weeks. Wider workflows that touch ERP and merchandising (`ai inventory management`, dynamic pricing) take the full 8. We don't quote 2-week timelines for 8-week work, and the pilot has an explicit kill point at week 6: if the metric won't move, we stop before production hardening and you don't pay phase 2.