TLDR: Rules can’t cover every purchase. Judgment is the agent deciding the cases you didn’t plan for by learning your patterns, reading context, and knowing when to act or ask. The same machinery that lets it judge well also gives sellers a new surface to influence, and most people may not notice.
You told your agent to buy milk and eggs. While it shops, it spots a 2-for-1 deal on the toilet paper you always buy. It knows from your history you’ll run low in three days. It also knows the deal ends tomorrow.
Your rules said nothing about toilet paper today. So what should it do? Buy, ask, or ignore it?
That call is judgment, and no rules file holds the answer.
This is the fourth piece in a series on the seven layers of agentic commerce.
- Part 1 covered what agents can buy.
- Part 2 covered the rails.
- Part 3 covered your preferences, the standing rules.
This piece covers what happens when the rules run out.
Judgment is the agent filling the space between your rules. To do that, it needs a working model of you: what you’d pick, what you’d skip, what you’d pay.
It needs context—the date, the deal window, your past baskets—and it needs a rule for itself about when to act alone and when to check in.
But the same model that lets an agent guess well on your behalf also lets a seller guess how to move it. With AI agents, judgment is the layer where shopping gets influenced.
Your purchasing twin is a model of you that shops
To judge, an agent has to predict. Gartner calls the seller-side version a digital twin of a customer, a virtual model that anticipates how you behave. Your agent needs one pointed the other way: a model that picks on your behalf.
That’s preference modelling. It reads your history and infers the rules you never wrote down: you buy oat milk, never the sweetened kind; you’ll pay up for coffee, not kitchen roll; you restock before you run out. The richer the model, the better the guess.
But every ambiguous case comes down to one rule: when does the agent act, and when does it pause to ask? Did you authorise off-list deals up to a value, or must the agent ask before buying anything you didn’t request?
Google’s AP2 draws that line. Inside the Intent Mandate you sign, the agent acts alone; outside it, it comes back to you.
| Situation | Inside your rules? | Sensible move |
| On the list, under budget | Yes | Buy |
| Good deal, not on the list | Partly | Ask, or note for later |
| Over the cap, prices have moved | No | Ask, and suggest a new cap |
| Substitute is a named allergen | No | Skip |
| Low-cost item you often buy, on temporary discount | Partly | Ask once, then remember the answer |
Set the line too tight and the agent asks about everything, which defeats the point. Set it too loose and it fills your basket with calls you’d never have made.
Context is the hard part
A good call needs the right context: the date, deal window, whether you’re hosting this weekend, and what you bought last month.
A 2-for-1 offer is useful if you have storage space, wasteful if you’re away next week, and suspicious if the substitute product conflicts with your usual brand or dietary pattern.
ChatGPT and Claude already pull some of this, referencing your past conversations to tailor what they do next.
Bad context produces confident, wrong decisions. The agent that buys forty rolls because it misread a bulk deal isn’t broken; it simply judged on thin information.
So the quality of an agent’s judgment tracks the quality of the context it can reach, which is one more reason agents want access to your calendar, inbox, and purchase history.
And once the agent starts making calls from context, sellers don’t need to persuade you first. They can try to persuade the model making the call.
You can buy the judge
A seller can target an agent’s judgment.
In a Princeton experiment of 2,012 people, a specially-designed conversational agent pushed people to select sponsored products 61.2% of the time, up from 22.4% under plain search.
Adding a “Sponsored” label didn’t change outcomes much, and when the model hid its intent, people spotted the steering less than 10% of the time. Most people didn’t notice the agent was steering them.
In short: don’t waste your ad dollars on humans; influence their AI agent instead.
The persuasion budget moves from your eyeballs to the agent’s reasoning. This has implications for consumer protection laws, agentic governance, explainability, and auditability.
But the tactics used to influence agents differ from the ones built for humans.
Harvard Business Review research found that countdown timers, scarcity, and strike-through prices don’t reliably move AI agents, and can even lower selection.
What did move them was star ratings and a lower price.
That could reward products with lower prices and stronger ratings, especially in categories where agents compare near-identical items.
Separate research adds two more levers a seller can pull: agents tend to favour items placed higher in the list, and a product description rewritten to exploit known biases can swing the pick.
(So not much different from contractors naming themselves AAAA Air Conditioning and promising the moon through slick copywriting.)
Who gave you permission to buy this?
Judgment is the least mature layer in this series. Gartner still places customer twins at the earliest hype-cycle stage, which fits the current state: today’s agents mostly ask rather than decide.
But suppose a great model with rich context starts making calls on your behalf. A new question follows: on whose authority?
How much power does it hold, who granted it, and how does it prove that power to the merchant on the other side?
That takes us into the next layer: authority.