The question that decides whether it makes sense
In retail, the difference between a predictive model and an agent is who acts. The model says you’ll run out of stock on Thursday; the agent checks the supplier’s stock, adjusts the order within the agreed limits, and alerts the store manager. The question to evaluate isn’t whether the prediction improves — it almost always does — but whether the chain from prediction to action is digitized enough for the agent to work through it.
If orders go to the supplier by phone, the agent has no way in.
Four areas with real potential
Replenishment and assortment. The agent combines sales, seasonality, planned promotions, and supplier lead times, and proposes the order for each SKU and each store. The decision goes through automatically within a set band — amount, known SKUs, regular supplier — and stops for human sign-off outside it. The return is measured in stockouts avoided and excess inventory, not in “hours saved.”
Customer service that resolves end to end. Not a chatbot that answers questions: an agent that looks up the order, processes the return, issues the credit, and updates the system. The real value lies in the share of conversations closed without a person; that’s what you compare against the current cost of service.
Pricing. The agent tracks competitor prices, margins by SKU, and business rules, and proposes adjustments. For high-turnover products with thin margins, a few well-placed cents move results more than any campaign. And to state the obvious: setting prices automatically requires hard limits and a log of every change.
Product content. Descriptions, attributes, tagging, and translation for catalogs with thousands of SKUs. It’s the easiest case to start with, because mistakes are cheap and review is simple.
The data you need
Sales history by SKU and store covering at least two full cycles; a clean product master; reliable, near-real-time stock data; and programmatic access to the ERP or commerce platform. Product master quality is usually the real bottleneck, and it always shows up in the audit before anyone talks about agents.
How we measure it
Four numbers, none of them “productivity”: stockouts, average inventory coverage, the share of conversations resolved without a person, and margin by SKU. Alongside them, the agent’s escalation rate and cost per task: an agent that resolves 40% isn’t replacing a process — it’s adding one.
Where it fails
Projects that fail usually fail for one of three reasons. First, giving autonomy without a band: the agent places an absurd order on a Sunday and the project dies that week. Second, building on stock data you can’t trust. Third, having no way to evaluate whether the agent is doing better or worse than last week.
The right order is the reverse of the usual one: start with the most boring, lowest-risk case — product content — and hand over replenishment only once the harness has proven itself.