Grocery ordering is one of those errands that sounds simple until you actually do it. The list is messy, the store has substitutions, delivery windows change, prices move, and one wrong click can turn “bananas” into a family-sized banana emergency. A Vellum assistant can handle the cart-building work while keeping the human in control of the purchase.
What you delegate
You give the assistant a grocery list, pantry restock request, recipe plan, or recurring household staples. The assistant opens Amazon Fresh in Chrome, searches for each item, picks reasonable matches, handles unavailable products, and prepares the cart. Before checkout, it sends a clear summary of items, substitutions, delivery window, fees, tip, and total.
Can you build my Amazon Fresh cart from this grocery list?
Get the usual versions where possible, prefer organic produce when the price difference is reasonable, and choose substitutions only when they are close enough.
Before checkout, send me:
1. The final cart
2. Any substitutions or unavailable items
3. Delivery window options
4. Fees, tip, and total
5. Anything you need me to confirm
Do not place the order until I approve the final cart.How it works
The assistant uses browser control through the local desktop and Chrome workflow to operate the website like a person would. It can search, compare item names, read sizes and prices, add items to the cart, and keep track of what it has already handled. For custom-rendered or awkward web UIs, the same pattern can lean on lower-level app control when that capability is enabled.
The useful part is judgment. “Greek yogurt” might need a brand, fat percentage, size, and quantity. “Stuff for tacos” might imply tortillas, protein, salsa, avocado, limes, cheese, and cilantro. The assistant can turn vague intent into a cart, but it can also pause when a choice is too personal or expensive to guess.
The checkout boundary stays explicit. The assistant can prepare the order, summarize the tradeoffs, and leave the cart ready. It should not spend money, choose a delivery window, or complete checkout until the user approves the final cart. That makes the workflow practical for real household purchasing instead of a cute demo that accidentally buys twelve jars of peanut butter. Not that anyone would complain.
A real example
Becky has used this pattern to order from Amazon Fresh through the Chrome extension workflow. The assistant handled the tedious browser work, built the cart, and kept the final purchase behind approval. That is the repeatable pattern: let the assistant do the clicking and comparison, but keep the human decision at the money boundary.
Why this pattern matters
Browser-based errands are hard to automate with brittle scripts because every site has different menus, popups, substitutions, and checkout flows. A personal assistant does not need a perfect API for every service. It can use the same interface the user uses, combine it with memory and preferences, and stop when approval is required.
The outcome
A scattered grocery request becomes a ready-to-review Amazon Fresh cart. The user avoids the tedious search-and-click loop, sees the substitutions and total, approves the final order, and gets groceries without losing twenty minutes to browser errands.