Every retail AI vendor has a demo. Very few have a planner who clicks “approve” on a Monday morning and watches real inventory move because of it.

That gap — between a convincing demo and a live production system that merchandising teams actually trust — is where most retail AI projects quietly die. Over the past quarter, several of our customers crossed it. This is what that journey looks like from the inside.

The milestone that matters

A few weeks ago, a premium apparel brand running Intelo agents hit the moment every deployment builds toward: a complete allocation run, generated by the agent, reviewed and approved by their planning team, and pushed directly into their WMS. Live orders. Real stores. No spreadsheet intermediary.

On paper it reads like a routine systems integration. In practice, it's the culmination of a deliberate trust-building process — one we've now seen repeat across apparel, footwear, and luxury deployments.

Phase one: the agent watches, the planner decides

No retailer should hand inventory decisions to an AI system on day one, and none do. Field trials start with the agent running in parallel: generating recommendations against live data while the existing process continues untouched.

This phase does two things. It validates the math — demand forecasts, target weeks of supply, size curves — against reality the planning team knows intimately. And it surfaces every data quirk that never appears in a clean demo environment: transaction-type exclusions, inferred floor sets, warehouse inventory splits. The unglamorous work of making the numbers match is the work.

One customer's planning team spent this phase interrogating individual recommendations — why this store, why this quantity, why not that size. Every question the agent could answer clearly built credit. That credit is the actual product.

Phase two: the planner edits, the system executes

The second phase is where our product philosophy shows. Recommendations become actionable, but every unit stays editable. Planners adjust quantities, exclude items, override store selections — and then release.

This is not a limitation to be engineered away. It's the design. Retail planning teams carry context no model sees: a store manager's flooring renovation, a marketing push landing next week, a fit issue coming back from customer service. The agent's job is to do the exhaustive part — scoring every SKU-store combination, every size, every week — so the planner's judgment is spent only where it adds value.

The tooling reflects that. Independent filters for sending and receiving stores. Bulk actions across thousands of items. Grid views that persist how each planner works. None of it is headline AI, and all of it determines whether the AI gets used.

Phase three: live execution, expanding scope

Once approved runs flow to production systems, something predictable happens: scope expands. The brand that started with replenishment invited us into initial allocation. A luxury customer that began with store-to-store rebalancing is now quantifying the financial gains of executed transfers with their steering committee and mapping the next categories.

And the hardest problems come into range. One of our retail customers runs size-level optimization in production across roughly 1,000 stores — six age-based size ranges crossed with gender divisions and fit variants. Size is where allocation math gets combinatorially brutal, and where broken sizes silently kill full-price sell-through. It's exactly the kind of problem that justifies an agent: too granular for any human team to work SKU by SKU, too commercially important to leave to averages.

What we've learned

Three lessons repeat across every deployment that reached production:

• Autonomy is earned in increments. The path is watch, then recommend, then execute-with-approval. Teams that trust the system at step three trust it because they stress-tested it at steps one and two.
• The workflow is the product. Model quality gets you into the trial. Editability, transparency, and bulk tooling get you into production.
• Outcomes compound. The first live run is the hard one. After that, each approved run adds evidence, each answered question adds trust, and each new agent starts from a higher baseline.

The demo era of retail AI is ending. What replaces it is quieter and far more valuable: planners approving live runs on Monday mornings, and inventory landing where the demand actually is.

If you're evaluating how AI agents could fit your in-season process, we're happy to show you what a field trial looks like — real data, your planners in control.