Two storefronts can share a language and still need different checkout rules. Map legal entities, taxes and processor eligibility separately from localization. Multiple payment accounts add flexibility; they do not replace compliance or make every basket eligible.
An AI recommendation can be wrong without being invented: the catalog changed after retrieval. Use incremental updates, expose sync status, revalidate price and stock before purchase, and trace the recommendation through checkout. Fix the failing layer—not just the prompt.
We’re exploring agent-to-agent commerce at https://t.co/ToMMAyhHdd. But for a merchant’s first deployment, I’d keep checkout human-confirmed until inventory rules, payment-abuse controls and fulfillment capacity are tested. Autonomy should follow readiness.
Before adding a second POS workflow at an event, strengthen the first: independent internet links, tested failover and backup power. More bandwidth is not the goal. Keeping checkout connected is. Provider outages still need a separate response plan.
Exact part-number lookup is the baseline. Cross-brand equivalents and fitment come next—but only when the data supports them. A similar part number is not proof of compatibility. AI can help retrieve the right relationship; it should not invent one.
Read-only does not mean risk-free. Before connecting an AI assistant to commerce data, define which fields it needs and which it must never receive. Grant the minimum access, then test it. You can add access later; you cannot undo a disclosure.
The @Square update I care about: shared menu controls and organization-wide reporting. For franchise operators, growth should not mean repeating the same update location by location. This is the kind of operational support Square needs as businesses scale.
I prefer Claude for planning, Codex for implementation, and a separate agent for QA. Not because more models are better, but because these are different jobs. The real test is cost per accepted result, including handoffs, tests and rework.
Selling faster is not the same as serving faster. Add more POS devices without checking preparation and pickup capacity, and you may just move the queue. Map the order's route, preserve its identity, and measure capacity at every handoff.
Un agente de IA no arregla un proceso que nadie puede explicar. La empresa pequeña necesita sacarlo de la cabeza del fundador. La grande, hacer explícitas sus decisiones y excepciones. Cambiar quién ejecuta no resuelve la falta de información.
An AI agent cannot fix a process nobody can explain. Small businesses need to get workflows out of the founder's head. Larger ones need to make unwritten decisions and exceptions explicit. Switching to an agent does not solve insufficient information.