TLDR: Agents need preferences, those preferences live in profiles and mandates, and consumers will only delegate with firm limits.
Your agent reaches the checkout with a full basket. One item, the oat milk you buy every week, is out of stock. Now what?
It could grab the supermarket’s own brand, pick a pricier organic carton, or drop the item and refund you. Without a rule, it guesses.
Guessing on oat milk is cheap; guessing on a £90 pair of running shoes is not.
This is the third piece in a series on the seven layers of agentic commerce:
This piece covers the layer above both: preferences. Once an agent can place any order, the question becomes what order, and in what order, and that takes rules.
The tools to encode those rules already exist. There’s a profile the agent reads, like ChatGPT’s memory and custom instructions, which already hold things like “I’m vegan” or “I’m allergic to peanuts.”
There’s a signed budget the agent can’t go past, like the price caps in Google’s AP2 Intent Mandate. There’s a brand list, and a substitution setting, the same kind Instacart shoppers have followed for years.
But people won’t hand over the basket without limits. In a six-market Checkout.com survey, the top conditions for delegating spend were spending caps (30%), instant revocation (29%), and easy cancellation (28%).
On average, consumers would let an agent spend £177 per purchase before asking again. Trust, not technology, is the constraint here.
| Preference | Where it lives | What it sets |
| Rules file | Agent profile, e.g. ChatGPT memory | Standing choices: brands, sizes, allergies |
| Budget | Signed mandate, e.g. AP2 | A hard spend cap per purchase |
| Brand preference | Profile, per category | A fixed brand or “best value” |
| Substitution rule | Per item | Replace, swap, or refund when out of stock |
| Approval rule | Agent settings or payment mandate | When the agent must ask before buying |
The rules file: what your agent can buy without asking you
Every preference layer starts with a place to keep standing rules. In ChatGPT, that place is custom instructions, guidance applied to every chat, plus memory, which holds details you’ve shared, like a dietary restriction. You tell it once, and it carries the rule forward.
A shopping agent needs the same type of stored rules, but with commerce fields: preferred brands, sizes, a delivery address, a price ceiling, allergies, ingredients to avoid, things never to buy.
Think of it as a shopping profile the agent consults before it fills a basket. The richer the file, the fewer questions it has to ask, and the closer its choices come to what you’d have picked yourself.
Google’s AP2 features an Intent Mandate—a signed record of price limits, timing, and conditions the agent has to honour. The mandate can hold a budget cap, an allowed list of merchants, and a category, so even a confused agent can’t buy outside the scope you set.
You are not as loyal as you think
Preferences present a commercially interesting angle. 57% of consumers said they’d let an agent switch brands if it found better value. So your brand loyalty is conditional on a machine’s read of the deal.
A preference file can pin a brand (“always this coffee”) or leave it open (“cheapest that fits”). Most shoppers will leave a lot of it open, which puts price and product data, not logos, in front of the agent.
For brands, that changes the job: get your value into the feed the agent reads, because the agent won’t see your packaging.
Substitution logic: out of stock, now what?
The oat milk substitution problem has a worked answer which predates AI. Instacart lets a shopper set, per item, whether to replace with a specific item, best match, or refund, and it saves that choice for future orders.
An agent needs the same setting, plus a budget tie-in. Replace, but not above the cap. Pick the store brand, unless it’s a named allergen. Refund anything more than a pound over the original.
These are small rules that together decide whether you trust the basket that turns up.
What happens when the market moves?
If you set a £3 cap on cereal, the rule holds until the box hits £4. After that, every cereal run fails. The agent will either keep dropping the item or refunding you. Do that for six months and you’ve effectively stopped buying cereal without deciding to.
What does an AI shopping agent do when your own budget rule blocks the purchase? A good one should flag the pattern, tell you the market has moved, and suggest a new ceiling.
"Your cereal failed eleven times this quarter. The brands you buy now start at £4.20. Raise the limit?" Preferences aren’t just instructions for the agent. They are also data about the buyer’s limits.
Every blocked purchase leaves a record. A signed budget cap is a number a merchant and payment network can both read, and each declined transaction has a reason: cart cleared, limit didn't.
Multiply that across thousands of similar baskets and you can unearth the price preferences of millions of shoppers; a map of what people will pay.
Say 65% of agent baskets cap cereal at £4. That tells the product team to hold cost-of-goods inside £4. It also tells the pricing team how much room there is above the line before demand falls away.
A firm could price a box at £4.50 and wait, betting that grocery inflation drags the average cap upward until it catches the price. Budget caps turn fuzzy willingness-to-pay into a hard threshold, measured per shopper, refreshed with every run.
Some of this data will surface the obvious way: a merchant who runs an agent checkout sees the cart, the cap, and the bounce. Some of it will surface sideways, inferred from declined-transaction patterns or passed up a supply chain that already trades demand data.
Your preference file describes a private rule, but at scale, the same files describe a market. The controls that protect you also reveal what you’ll tolerate, and the firms on the other side of the basket have every reason to read them closely.
What comes next
Adoption data tells us that UK consumers would hand over groceries first, with 40% comfortable letting an agent take the boring everyday purchases, since a wrong carton is a small loss. And since only 3% of reported transactions involve an agent today, most of this is still setup.
But suppose your preference file is full, the budget signed, and substitution rules set. Your AI shopping agent now knows what you’d buy in each case you thought of.
It still has to handle the edge cases you didn’t plan for, like a deal that’s good but not on your list, a tradeoff between price and delivery, or a choice no rule covers.
The agent was told to buy milk and eggs. It sees a 2-for-1 deal on toilet paper, knows from your history that you’ll need it in three days, and knows the deal ends tomorrow. Your rules didn’t cover that. Should it buy, ask, or ignore it?
Would this introduce the possibility of influenced agents—perhaps not through advertising per se, but through ambient information around your life, general preferences, and historical purchases that would justify inclusion in the basket?
These questions take us into the next layer of our series: judgment.