Path 1: Agentic discovery (the LLM interface)
Path 1 is the LLM interface – this is what happens when your customer is actively present, interacting with an AI assistant like ChatGPT or Gemini, and that assistant helps them discover, evaluate, and move towards purchasing a product. The customer is in the loop, and the agent is a sophisticated, context-aware intermediary.
You’ve probably shopped like this yourself. It’s happening now, it’s growing fast, and it has clear implications for how you structure your product data and your digital presence.
How the model operates
When a customer opens ChatGPT and asks, “What’s the best commercial espresso machine for a small café with a budget of around €3,000?” – that’s agentic discovery.
The model then pulls from its training data, live web browsing, structured product feeds, and whatever other signals it can find to construct a recommendation. Your product either appears in that answer, or it doesn’t.
This isn’t hypothetical. It’s already how a meaningful and growing segment of your customers are starting to shop.
The platforms driving the agentic shift
Perplexity has built an entire product around shopping discovery, Google's AI Overviews are reshaping what appears above the fold in search results, and ChatGPT is still exploring its shopping capabilities. In all of these scenarios, the customer still makes the final call – but the agent dictates exactly what they see and consider.
Think of the AI less like a transaction tool and more like an encyclopaedic sales consultant. It has read everything ever written about your product category and can have a nuanced conversation about it with your customer. Except this consultant is available 24 hours a day, speaks every language, and serves thousands of customers simultaneously.
Path 2: Agentic transactions (the delegated buyer)
This is the delegated agentic transaction. A customer gives an AI agent a standing instruction and a budget, and the agent executes purchases autonomously in the background. The customer is entirely out of the loop, and the agent is making financial decisions on their behalf.
How the delegated agentic model works
A customer decides they want to stop thinking about a category of recurring purchases – household supplies, office consumables, standard B2B inputs – and delegates the decision to an AI. They set the parameters: “Find me a cheaper supplier for coffee; you’ve got a €50 monthly budget, make sure it’s a sustainable brand, and always find the lowest price.” The agent then monitors, selects, and executes purchases autonomously.
For the customer, this is genuinely useful. It eliminates low-value cognitive load. For businesses who reliably meet the agent’s criteria, it represents recurring revenue that doesn’t depend on winning the customer’s attention every month. It shifts purchasing from an active chore to a background process that’s taken care of on their behalf.
Why the delegated infrastructure isn't ready
As a concept, delegated commerce is genuinely compelling. But the infrastructure required to make it work safely does not yet exist at scale.
The core problem is authentication and fraud detection. When a payment arrives at your payment processor today, sophisticated systems determine whether it’s legitimate based on human behaviour patterns. An AI agent executing a payment generates signals that look completely different – and in many cases, look exactly like fraud. False declines are a real risk, as are fraudulent transactions that slip through by mimicking agent behaviour.
Banks are not yet equipped to handle this distinction. The protocols that would allow a payment to carry verified ‘this was authorised by a human and executed by a trusted agent’ signals are still being developed.
Then there is the unresolved chargeback question. If an AI agent makes a purchase and the customer says, “I didn’t authorise that specific transaction,” the liability framework becomes incredibly murky. Payment providers and regulators are still working through these implications. Until they do, building significant infrastructure around delegated agentic payments is a risk.
