First published on LinkedIn on 14 April 2026.
Context: This article came from a simple overlap: I like cooking, and I am interested in what agentic AI means for ecommerce. Lasagna turned out to be a useful way to think about compressed intent, missing context, and delegated judgment. I also used different large language models (LLMs) to stress-test the argument and push the article harder before publishing.
The future of commerce is supposed to be agentic. AI that does not just recommend products, but acts on your behalf. No forms, no browsing, fewer decisions. Just a request, translated into action.
Ask a retail AI to “order everything for a lasagna dinner for four” and the limits appear immediately. Not a technical limit. A structural one.
This is the Lasagna Dilemma.
It matters because it reveals something larger than grocery shopping. It shows what happens when human intent is compressed into a short request, then delegated to a system that must act without fully understanding the situation behind it. And it matters for another reason: “user” no longer refers only to a human being. Increasingly, the interface is being read, interpreted, and acted on by agents as well.
The question is not simply whether the agent can complete the task. It is whether it can resolve the ambiguity well enough, cheaply enough, and with low enough perceived risk that most people accept the result without caring what got lost in translation.
That is where agentic commerce will be won or lost.
The problem is not the catalogue
“Lasagna” is not a stock keeping unit (SKU). It is not even a stable need. It is a compressed expression of a human situation.
When someone says they want ingredients for lasagna, they are not naming a single product set. They are collapsing multiple judgments into one phrase. Effort. Skill. Occasion. Time. Standards. Dietary constraints. Social stakes. Household habits.
A lasagna for four could mean very different things. It could be a fast weekday meal built from pre-made sheets and jarred sauce. It could be a slow weekend project with handmade pasta and a four-hour ragù. It could be for children, close friends, or guests someone is trying to impress. It could assume a cook who makes béchamel without thinking, or one who avoids it altogether. Even “Saturday” is unstable. One Saturday is casual. Another is loaded with meaning.
The agent can only work with what was declared, what it can infer, and what context it already holds. If that context is thin, it is likely to fall back on probability matching: common baskets, common substitutions, common behaviour, common shortcuts. In practice, that often means the safe average answer.
For some users, that answer will be acceptable. For others, it will be visibly wrong. The problem is not that the system cannot generate a basket. The problem is that the basket may reflect what is statistically common rather than what is situationally appropriate.
That is the first limit of agentic commerce. It sounds frictionless only when the intent is stable, low-stakes, and easy to standardise.
Lasagna is useful because it shows the boundary.
The real issue is not failure. It is context loss
One way to misread this argument is to assume it is saying agentic commerce will fail. That is too simple.
The more serious possibility is that it succeeds commercially while still losing something important. Not perfect understanding, but nuance. Not all relevance, but the kinds of contextual distinctions that matter in more expressive, occasion-sensitive, or identity-laden decisions.
The system may not eliminate ambiguity. It may simply make people stop caring about it in many categories.
A system does not need to understand intent deeply to win adoption. It only needs to reduce friction enough that most users accept a good-enough answer. If the cost of correction feels higher than the value of nuance, then acceptable becomes the dominant commercial outcome.
The Lasagna Dilemma is not just about whether AI gets the dinner right. It is about whether convenience normalises a lower-resolution version of choice.
Not all delegation is the same
This distinction becomes clearer when we separate different kinds of delegated commerce.
Some purchases are low-stakes and highly repeatable. Toilet paper. Washing detergent. Pet food. Milk. Bin liners. Vitamins bought on routine. These are replenishment purchases. The user is not looking for interpretation. They are looking for continuity. Here, good enough is often exactly what the system needs to provide.
Other purchases are convenient but still somewhat variable. A weekly grocery top-up. Basic household food. Standard personal care. In these cases, the user may tolerate some inference and some substitution, as long as the downside of being wrong remains manageable.
Then there are judgment-heavy purchases. Meals for a particular occasion. Gifts. Clothing with social meaning. Wine. Specialist tools. Products where standards, identity, taste, or context matter. This is where delegation becomes more fragile.
The Lasagna Dilemma belongs mainly to this third category. It is not really about groceries in general. It is about judgment delegation under uncertainty.
That matters because agentic commerce may expand very quickly in replenishment and routine convenience while remaining far weaker in areas that require interpretation. A system can dominate easy delegation without solving hard delegation.
The hidden brake on agency is liability
Once an agent acts rather than recommends, ambiguity is no longer just a relevance problem. It becomes a liability problem.
If the system buys the wrong ingredients, misses an allergy, gets the quantity wrong, or makes a bad substitution for an important occasion, someone owns the failure. The platform, the merchant, the model provider, or the integration layer may all try to displace responsibility, but from the user’s perspective the damage is real either way.
A refunded basket does not repair a ruined dinner, a failed gift, or a socially awkward occasion. In high-ambiguity purchases, the cost of being wrong is not only monetary. It is temporal, emotional, and reputational.
There is a related constraint here that matters beyond groceries. Not all friction is a design failure. In some domains, friction exists for legal, regulatory, or safety reasons, and removing it can create liability rather than reduce it. Financial services and healthcare make that visible quickly.
That is one reason agentic commerce may scale more quickly in low-stakes replenishment than in judgment-heavy categories. The liability tolerance is different. Routine replenishment can absorb more automation error because the downside is smaller and easier to recover from.
So the real frontier is not commerce in general. It is the boundary between purchases where being wrong is tolerable and purchases where being wrong changes the meaning of the event.
Why a single brand cannot see enough
If the problem is context, the obvious response is to say that brands should collect more of it. Build richer loyalty systems. Learn substitutions, preferences, frequency, household rhythm. Infer taste over time. Personalise harder.
That is a reasonable instinct. It is also constrained.
People do not live inside one retailer’s field of view. Their behaviour is distributed across platforms, devices, services, and moments. They search in one place, browse in another, buy from multiple merchants, save ideas somewhere else, and make decisions under constraints that no single brand fully sees. Cooking intent may be shaped by a guest coming over, a busy workday, a saved recipe, a dietary issue, a kitchen tool already owned, or a family habit that is invisible to the retailer.
A loyalty programme can reveal useful things. What someone buys. How often. At what price point. In which categories. With which brand patterns. But it still offers only a partial view. It can tell you how someone shops with you. It cannot fully tell you how they live.
For routine purchases, that partial view may be enough. But as soon as the request depends on context outside the brand’s field of view, the weakness becomes visible. Not that brands know nothing. That they often do not know enough.
Why platforms have an advantage, but not a perfect one
Large platforms sit closer to a broader context layer than most individual brands. They often have access to a wider mix of signals: searches, device context, cross-category behaviour, location history, calendar cues, usage patterns, and other fragments of everyday life. That breadth can create an interpretive advantage.
A grocery brand may know that you bought pasta last month. A broader platform may be in a better position to recognise that this Saturday is not an ordinary Saturday. Even partial clues can matter if they help disambiguate a vague request better than purchase history alone.
But more data is not the same as better judgment. Broad context can increase signal, but it can also increase noise. A birthday in the calendar does not tell the system whether the meal should be ambitious or cheap. Search history may reveal curiosity rather than commitment. Broad context is only useful if it improves interpretation more than it increases confusion.
Platforms have an advantage in breadth. Brands often have an advantage in specificity. The real contest is not over who has more data, but over whose data improves intent resolution at the moment it matters.
A platform may infer that someone cooks seriously because they search recipes, watch food videos, or read reviews of pasta machines. A retailer may know that the same person buys 10kg of specialty flour every month, never buys jarred sauce, and repeatedly orders ingredients associated with labour-intensive cooking. One signal is inference. The other is repeated commercial behaviour. Transaction data is not broad context, but it often has a higher truth value.
So the future is unlikely to be platform conquest and brand submission. A more realistic outcome is asymmetry with interdependence. Platforms may be better positioned to interpret the request. Brands may be better positioned to validate the pattern, supply the inventory truth, manage substitutions, and verify the outcome against repeated behaviour. The most capable systems may require a handshake between the two: broader context for intent resolution, narrower truth for commercial accuracy.
The specialist does not disappear, but the role changes
A specialist wine merchant, trusted butcher, expert cookware retailer, or grocer known for curation rather than convenience does not only provide stock. It provides judgment. In these cases, part of the value lies in whose interpretation the customer trusts.
That still matters in an agentic world, but the role of the specialist changes. It is no longer enough to assume that judgment survives only as a direct human-to-human interaction. In an agent-mediated environment, the more important question is whether specialist judgment remains attributable, differentiable, and difficult to commoditise.
The specialist moat may persist, but not in its old form. The specialist may become less of a direct interface and more of a trusted upstream signal in the decision stack. A user may not speak to the butcher personally. They may tell the agent to use the butcher’s weekly recommendations as the source of truth. The specialist then remains powerful, not because it controls the final interface, but because its judgment remains sought after inside the system.
But it is narrower and more demanding than before. The specialist has to become legible without becoming generic. It has to preserve authority even when mediated.
Good enough can hollow out brands
If the system can deliver a fast, acceptable answer for most routine or semi-routine purchases, then many brands may discover that their differentiation was weaker than they thought. They were protected not only by product quality or deep loyalty, but by the friction of human shopping. People browsed. They noticed packaging. They repeated habits. They saw familiar logos. Attention itself helped brands exist.
Agentic systems remove part of that attention.
Once the user stops browsing, many weak reasons for choosing a brand begin to erode. Vague emotional preference. Soft loyalty. Shelf-based familiarity. Attention-based advantage at the point of decision. These do not disappear entirely, but they matter less if the consumer delegates the choice and accepts the shortlist.
Brands that reduce meaningful risk can survive. Brands that embody trusted judgment can survive. Brands that make themselves highly legible to systems through structured data, consistent quality, and dependable fulfilment can survive. But brands whose advantage depended mainly on attention, light familiarity, or symbolic differentiation at the moment of browsing become much more exposed.
Good enough does not kill brands equally. It kills weak reasons for choosing them.
Execution still matters, but it is not the same as control
Brands and merchants still control something essential: execution. An agent may interpret the request, but someone still has to know what is actually in stock, what can be substituted, what is fresh, what can be delivered, and whether the outcome will hold together in the real world.
But execution is not the same as strategic control. If the platform resolves intent and the brand merely fulfils it, then the brand risks becoming a headless endpoint. Necessary, valuable, operationally critical, but less able to shape the decision upstream.
That is why building an in-house agent is not automatically a strategic answer. If the brand does not control enough context to resolve ambiguous intent well, then its agent may simply become an expensive interface wrapped around structural weakness. It may look modern while remaining context-poor.
Some platforms are already restricting external agent access, positioning their own artificial intelligence (AI) layer as the differentiator. In that context, the in-house agent is not only a product response. It is a control move. But it only becomes meaningful if the brand controls enough of the context stack to make the agent’s judgment better than generic orchestration.
What this changes about selection
If agentic systems increasingly mediate commerce, then some brand value now has to survive translation into machine-readable signals.
Structured product data becomes more strategically important. So do dietary attributes, provenance signals, preparation requirements, compatibility logic, substitution rules, reliability data, and any other form of information that helps the system choose intelligently.
The same conditions that make interfaces easier for people to scan and understand also make them easier for systems to interpret. Semantic structure, clear labelling, logical hierarchy, and strong information scent are not only accessibility or content-design concerns anymore. They become part of whether an agent can parse the interface and act reliably within it.
Visibility is no longer shaped only by recall or campaign presence. It is also shaped by whether the product can be understood and selected by an intermediary acting on the customer’s behalf.
Loyalty changes too. It matters less whether a customer passively likes a brand and more whether the brand becomes the preferred endpoint for a class of requests. Technical infrastructure, application programming interface (API) quality, structured data, fulfilment reliability, low error rates, consistent product truth, begins to look less like backend hygiene and more like frontline competitive assets.
Branding does not vanish. But it loses some of its autonomy. It has to cash out in forms that influence selection inside a mediated system. In agentic commerce, the first audience is not always only the person. Increasingly, it is also the machine acting for them.
The strategic question for brands
The most important question is no longer whether a brand should build an agent. That is too shallow.
The more serious question is what the brand still controls when the agent belongs to someone else. Does it control transaction truth. Does it control fulfilment quality. Does it hold specialist authority. Does it expose product meaning clearly enough to be machine-readable. Does it own signals that improve intent resolution. Has it earned enough trust that customers will volunteer deeper preference data in exchange for visible value.
The task is not simply to make shopping conversational. It is to remain legible, trusted, and useful inside a system where the interface may no longer be yours, and where good enough may be commercially stronger than excellence in more categories than brands would like to admit.
Conclusion
The Lasagna Dilemma is not proof that agentic commerce fails. It is proof that different kinds of commerce tolerate different levels of contextual loss.
Agents may perform well in low-stakes, repeatable purchases by defaulting to good-enough averages. That may be enough to drive adoption at scale. But when a request is socially loaded, occasion-sensitive, or identity-laden, the missing context becomes visible. In those moments, the contest is not simply about who has more data. It is about who can turn available signals into trustworthy judgment with acceptable risk.
Platforms have an advantage in breadth of context. Brands often retain an advantage in transaction truth, fulfilment truth, and, in narrow cases, specialist judgment. Neither side owns the whole picture. But the balance of power shifts toward whoever can interpret compressed human intent without creating too much friction, too much noise, or too much liability.
Marketing for the last thirty years assumed a human on the other end of the message. In agentic commerce, the first audience may increasingly be a machine.
The brands that understand that early will not control everything. But they may still control enough to matter.
Alessandro Zulberti is a UX researcher and accessibility lead. More about my work.