What pays from the first month
Three use cases established themselves across all our stores: the support assistant (hundreds of conversations handled per month, human escalation on sensitive cases), catalogue enrichment (descriptions, translations, consistency of attributes) and monitoring — competitor prices, reviews, SEO rankings — with alerts that arrive already qualified.
What they have in common: repetitive, high-volume tasks where a mistake can be recovered from. That is where AI pays off, not in grand promises of "augmented strategy".
What we dropped
Content generation on autopilot, with no review: the tone drifts, approximations pile up, and Google eventually notices. "Shop window" chatbots plugged into nothing: a bot that knows neither the customer's order nor the stock does more damage than a form.
We also gave up on agents that act without approval on irreversible operations — refunds, price changes. In 2026 the right setting is still: AI proposes, a human approves. Auto mode has to be earned, scope by scope.
The framework that makes AI reliable
Every agent has a written scope: what it may read, what it may do, what triggers an escalation. Its actions are logged and reviewed by sampling. And above all, it is plugged into real data — orders, stock, CRM — not into generic answers.
That is exactly what we built into the suite: catalogue, order and SEO agents that run in alert mode by default, and in auto mode only on the plan that takes that on. Useful AI is engineering, not magic.