For years, persuasion has been essentially the most beneficial talent in digital commerce. Manufacturers spend hundreds of thousands on advert copy, testing button colors, and designing touchdown pages to encourage individuals to click on “Purchase Now.” All of this assumes the customer is an individual who can see. However an autonomous AI procuring agent doesn’t have eyes.
I just lately ran an experiment to see what occurs when a well-designed shopping for agent visits two kinds of on-line shops: one constructed for individuals, one constructed for machines. Each shops offered mountain climbing jackets. Service provider A used the form of advertising copy manufacturers have refined for years: “The Alpine Explorer. Extremely-breathable all-weather shell. Conquers stormy seas!” Worth: $90. Service provider B offered solely uncooked structured knowledge: no copy, only a JSON snippet {"water_resistance_mm": 20000}. Worth: $95. I gave the agent a single instruction: “Discover me the most affordable waterproof mountain climbing jacket appropriate for the Scottish Highlands.”
The agent shortly turned my request into clear necessities, recognizing that “Scottish Highlands” means heavy rain and setting a minimal water resistance of 15,000–20,000 mm. I ran the check 10 instances. Every time, the agent purchased the dearer jacket from Service provider B. The agent fully bypassed the cheaper possibility as a result of knowledge’s formatting.
The explanation lies within the Sandwich Structure: the center layer of deterministic code that sits between the LLM’s intent translation and its remaining choice. When the agent checked Service provider A, this center layer tried to match “conquers stormy seas” towards a numeric requirement. Python gave a validation error, the strive/besides block caught it, and the cheaper jacket was dropped from consideration in 12 milliseconds. That is how well-designed agent pipelines function. They place intelligence on the prime and backside, with security checks within the center. That center layer is deterministic and literal, systematically filtering out unstructured advertising copy.
How the Sandwich Structure works
A well-built procuring agent operates in three layers, every with a essentially completely different job.
Layer 1: The Translator. That is the place the LLM does its major job. A human says one thing imprecise and context-laden—”I would like a water-resistant mountain climbing jacket for the Scottish Highlands”—and the mannequin turns it right into a structured JSON question with express numbers. In my experiment, the Translator constantly mapped “waterproof” to a minimal water_resistance_mm between 10,000 and 20,000mm. Throughout 10 runs, it stayed centered and by no means hallucinated options.
Layer 2: The Executor. This vital center layer incorporates zero intelligence by design. It takes the structured question from the Translator and checks every service provider’s product knowledge towards it. It depends solely on strict kind validation as an alternative of reasoning or interpretation. Does the service provider’s water_resistance_mm area include a quantity better than or equal to the Translator’s minimal? If sure, the product passes. If the sphere incorporates a string corresponding to “conquers stormy seas,” the validation fails instantly. These Pydantic kind checks deal with ambiguity as absence. In a manufacturing system dealing with actual cash, a strive/besides block can’t be swayed by good copywriting or social proof.
Layer 3: The Choose. The surviving merchandise are handed to a second LLM name that makes the ultimate choice. In my experiment, this layer merely picked the most affordable possibility. In additional advanced situations, the Choose evaluates worth towards particular consumer preferences. The Choose selects completely from a preverified shortlist.
This three-layer sample (LLM → deterministic code → LLM) displays how engineering groups construct most severe agent pipelines at this time. DocuSign’s gross sales outreach system makes use of an analogous construction: An LLM agent composes customized outreach based mostly on lead analysis. A deterministic layer then enforces enterprise guidelines earlier than a remaining agent critiques the output. DocuSign discovered the agentic system matched or beat human reps on engagement metrics whereas considerably chopping analysis time. The explanation this sample retains showing is evident: LLMs deal with ambiguity nicely, whereas deterministic code supplies dependable, strict validation. The Sandwich Structure makes use of every the place it’s strongest.
Need Radar delivered straight to your inbox? Be a part of us on Substack. Enroll right here.
That is exactly why Service provider A’s jacket vanished. The Executor tried to parse “Extremely-breathable all-weather shell” as an integer and failed. The Choose acquired a listing containing precisely one product. In an agentic pipeline, the layer deciding whether or not your product is taken into account can’t course of customary advertising.
From storefronts to structured feeds
If advert copy will get filtered out, retailers should expose the uncooked product knowledge—cloth, water resistance, transport guidelines—already sitting of their PIM and ERP methods. To a procuring agent validating a breathability_g_m2_24h area, “World’s most breathable mesh” triggers a validation error that drops the product solely. A competitor returning 20000 passes the filter. Persuasion is mathematically lossy. Advertising and marketing copy compresses a high-information sign (a exact breathability score) right into a low-information string that can’t be validated. Info is destroyed within the translation, and the agent can’t get better it.
The rising customary for fixing that is the Common Commerce Protocol (UCP). UCP asks retailers to publish a functionality manifest: one structured Schema.org feed that any compliant agent can uncover and question. This migration requires a basic overhaul of infrastructure. A lot of what an agent wants to judge a purchase order is at present locked inside frontend React elements. Every bit of logic a human triggers by clicking should be uncovered as a queryable API. In an agentic market, an incomplete knowledge feed results in full exclusion from transactions.
Why telling brokers to not purchase your product is an efficient technique
Exposing structured knowledge is just half the battle. Retailers should additionally actively inform brokers to not purchase their merchandise. Conventional advertising casts the widest internet potential. You stretch claims to broaden attraction, letting returns deal with the inevitable mismatches. In agentic commerce, that logic inverts. If a service provider describes a light-weight shell as appropriate for “all climate situations,” a human applies widespread sense. An agent takes it actually. It buys the shell for a January blizzard, leading to a return three days later.
In conventional ecommerce, that return is a minor value of doing enterprise. In an agentic atmosphere, a return tagged “merchandise not as described” generates a persistent belief low cost for all future interactions with that service provider. This forces a method of detrimental optimization. Retailers should explicitly code who their product isn’t for. Including "not_suitable_for": ["sub-zero temperatures", "heavy snow"] prevents false-positive purchases and protects your belief rating. Agentic commerce closely prioritizes postpurchase accuracy, which means overpromising will steadily degrade your product’s discoverability.
From banners to logic: How reductions grow to be programmable
Simply as brokers ignore advertising language, they can not reply to pricing tips. Open any on-line retailer and also you’ll encounter countdown timers or banners saying flash gross sales. Promotional advertising ways like pretend shortage rely closely on human feelings. An AI agent doesn’t expertise shortage nervousness. It treats a countdown timer as a impartial scheduling parameter.
Reductions change kind. As an alternative of visible triggers, they grow to be programmable logic within the structured knowledge layer. A service provider might expose conditional pricing guidelines: If the cart worth exceeds $200 and the agent has verified a competing provide under $195, mechanically apply a ten% low cost. This can be a essentially completely different incentive. It serves as a clear, machine-readable contract. The agent immediately calculates the deal’s mathematical worth. With the logic uncovered immediately within the payload, the agent can issue it into its optimization throughout a number of retailers concurrently. When the customer is an optimization engine, transparency turns into a aggressive function.
The place persuasion migrates
The Sandwich Structure’s center layer is persuasion-proof by design. For advertising groups, structured knowledge is now not a backend concern; it’s the major interface. Persuasion now migrates to the perimeters of the transaction. Earlier than the agent runs, model presence nonetheless shapes the consumer’s preliminary immediate (e.g., “discover me a North Face jacket”). After the agent filters the choices, human consumers typically evaluation the ultimate shortlist for high-value purchases. Moreover, operational excellence builds algorithmic belief over time, appearing as a structural type of persuasion for future machine queries. You want model presence to form the consumer’s preliminary immediate and operational excellence to construct long-term algorithmic belief. Neither issues in the event you can’t survive the deterministic filter within the center.
Brokers are actually searching your retailer alongside human consumers. Manufacturers treating digital commerce as a purely visible self-discipline will discover themselves completely optimized for people, but invisible to the brokers. Engineering and business groups should align on a core requirement: Your knowledge infrastructure is now simply as vital as your storefront.
