what to expect from an ecommerce AI agent in the next 12 months

This is not speculation: today's conversational agents are no longer limited to text. They can receive a photo — a payment receipt, a garment the customer wants an exact match of — and understand it as part of the conversation, not as an ignored attachment. They can receive a voice note and reply taking into account what the customer actually said, with automatic transcription in between. heysteff runs both capabilities today: vision to read images and receipts, and voice-note transcription.

What is a forward-looking trend is that this capability stops feeling like an "extra" feature and becomes the baseline expectation for any serious agent, with no distinction between text, image or audio.

The Model Context Protocol (MCP) is an open standard, created by Anthropic, that lets an AI assistant use tools from other applications in a safe, auditable way. It is no longer a lab concept: heysteff exposes an MCP server today, through which a customer's own AI can configure the entire workspace — build flows, adjust escalation rules, review conversations — with every action recorded in an audit log.

The leap ahead in the coming months is for this way of operating software — asking an assistant instead of manually opening every dashboard — to spread across business tools in general, not just customer support platforms.

Another reality already here: no language-model provider is the single right choice forever. Different models have different strengths, costs and speeds, and that relationship shifts over time. heysteff already lets you choose the AI provider per workspace — Anthropic, OpenAI, DeepSeek or OpenRouter — instead of permanently tying a business to a single model.

What to expect is that this flexibility becomes a standard evaluation criterion when choosing AI software: any platform that forces a single model, with no way to switch when a better or cheaper one comes along, will be at a disadvantage against the ones that allow it.

Here we do enter projection territory, labeled as such: the idea that a shopping assistant representing the customer ("find me the best running-shoe option under a certain budget") talks directly to a brand's sales agent, negotiating availability, price or conditions without any person typing a message. The technical components for this to happen — agents that use tools, open protocols like MCP so one system can talk to another, models capable of negotiating within rules — already exist separately. Their coming together into an extended, trustworthy agent-to-agent purchase flow is, still, a trend in formation, not an established fact.

If that happens, the business that already has its catalog, its stock and its sales rules exposed in an orderly way to an agent of its own — instead of only to a website built for humans — will be in a better position to take part.

With so much marketing around "AI" in sales software, a simple criterion helps: ask whether the capability being sold to you exists today, verifiable and usable, or whether it's a roadmap promise. Multimodality, MCP and per-workspace model choice are real and evaluable right now. Agentic commerce is not yet a complete flow, though its pieces are. Any vendor should be able to tell you clearly which of the two categories each thing they offer falls into.

• The State of Conversational AI in LATAM Ecommerce (2026)
• The Model Context Protocol and the Future of Business Software
• MCP in heysteff
• What Is an AI Agent for Ecommerce?