AI & commerce / 4 MIN READ
E-commerce for agentic shopping
What should a shopping experience become when an AI agent helps the customer choose — and the customer still needs control?
What happens to an online store when the person browsing its pages is no longer the only participant in the purchase?
Imagine asking an assistant to prepare a grocery basket for the week: stay within a budget, avoid certain ingredients, reuse what is already in the kitchen and choose a delivery slot after work. The design problem extends beyond helping someone find a product. It becomes a question of how a customer can delegate a task, understand the result and decide what happens next.
This is an exploratory draft about that possible experience, rather than a prediction that every customer will shop this way.
Start with the shopping mission
A search query such as “milk” leaves many decisions open. A shopping mission adds context: who the purchase is for, which constraints matter and what can be negotiated.
For an agent-assisted experience, I would start by separating three things:
- Requirements: constraints that must be respected, such as an allergy or a spending limit.
- Preferences: choices that can be traded off, such as brand, pack size or delivery speed.
- Permissions: actions the customer allows the assistant to take, from building a shortlist to placing an order.
The interface should make these distinctions easy to review and change. A preference should never quietly become permission to spend money.
Product information becomes part of the experience
A useful product page helps people understand what they are buying. An agent-assisted journey would also need dependable information that software can interpret consistently.
As a design principle, I would prioritize clear product attributes, comparable units, explicit availability and transparent costs. Grocery shopping makes the gaps particularly visible: a similar-looking product may have different ingredients, a different weight or a different storage requirement.
The question for a product team is practical: Could someone explain why this item fits the request using the information we actually provide? If the answer is unclear, an attractive interface cannot resolve the uncertainty on its own.
Design the handover, not just the automation
Delegation needs boundaries. A customer might allow an assistant to suggest a basket but want to approve substitutions, a higher delivery charge or any purchase above a set amount.
I would make those moments explicit. Before an order is placed, the customer should be able to see the final basket, total cost, delivery conditions and any deviations from the original request.
When a decision needs approval, the explanation should say what changed and why it matters. “The requested item is unavailable; this alternative costs more and contains a different ingredient” is more useful than a generic request to continue.
Make recommendations open to correction
An explanation should help a customer challenge a choice. Why this product? Which preference influenced the recommendation? What was excluded? What is still uncertain?
A useful starting point would be to offer simple corrections: keep the brand, lower the budget, remove the substitute or compare another option. The customer should be able to adjust one decision without starting the entire task again.
Measure the quality of delegation
For a pilot, I would look beyond conversion. A completed order alone would not tell us whether the assistant respected the customer’s intent.
Research questions could include:
- Did the basket meet the stated requirements?
- Were costs and substitutions understood before approval?
- Could customers correct mistakes easily?
- Did the experience reduce effort without reducing control?
- What happened when the assistant could not complete the task?
These are proposed evaluation questions, not results from a project I have delivered.
A starting point for product teams
I would begin with a bounded task, such as helping a customer prepare a repeat order for review. Define the permissions, map the exceptions and test the handover before expanding the assistant’s autonomy.
For me, the interesting design challenge is making delegation understandable: what the system can do, what it has done and which decisions still belong to the customer.