When Production Is Cheap: AI and the Economics of Being Wrong

In their 1998 paper describing Google, Sergey Brin and Lawrence Page devoted an appendix to the problem created by advertising-funded search. A ranking system financed by advertisers, they argued, would be biased toward advertisers and away from users, and they recommended that search remain transparent and academic. Google ultimately built an advertising business around the same intermediary position Brin and Page had identified: the system interpreting a buyer's intention could also sell influence over the result.

The same conflict appears wherever an intermediary sits between what a buyer wants and what a seller provides.

Broadcast media organized commerce around attention: ads captured it, posed a problem, promoted a solution, and sold it. Search introduced explicit queries, while social platforms inferred intent from behavior. Current AI systems can receive a far richer statement of what a buyer wants and retain that context across interactions.

The shift from captured attention toward increasingly legible intention is already visible. Its larger consequence is the market structure emerging as production and matching become inexpensive.

Where value moves as matching costs fall

AI is collapsing the cost of producing solutions and matching them to buyers faster than it reduces the consequences of a poor choice.
Buyer intent is becoming more detailed and persistent, while the economic bottleneck continues moving beyond matching.
Cheaply reversible decisions reward rapid iteration, while expensive or irreversible decisions push verification and controls ahead of commitment.
Relational channels persist because prior conduct, reputation, and an accountable counterparty reduce uncertainty before and after a deal.
Procurement intermediaries spread the cost of finding, comparing, vetting, and buying from suppliers across many transactions, while automation increases the relative value of standardization and accountability.
Whoever defines which claims, evidence, and remedies qualify for automated purchasing can become a consequential gatekeeper between abundant suppliers and increasingly selective buyers.

Demand becomes legible

Television sold a sequence: capture attention, pose a problem, promote a solution, sell it. Search received demand directly through the query, giving advertisers a commercial signal much closer to the moment of expressed need. Social platforms added behavioral histories and inferred the next want from the last action. All three preserved a discrete commercial moment in which sellers competed for placement before a person chose.

Amazon demonstrated a different posture earlier and without fanfare. Rather than wait for a query, it reads browsing, purchasing, and recurring-order patterns in real time and anticipates the need.

Conversational systems extend that posture to language: the buyer can state an outcome in their own words, with constraints, budget, and context attached, while repeated interactions can preserve substantially more context than a conventional query. The difference from search lies in the resolution and persistence of the demand signal.

A three-word query contains a sliver of the buyer's situation. A system with cross-context memory retains what was bought, what was rejected, what it cost, and why. The signal remains probabilistic while becoming more detailed and context-rich.

Agentic commerce can also convert an explicit instruction into a bounded purchasing mandate: renew this license, replenish this inventory, or solicit alternatives under specified conditions.

Google's Agent Payments Protocol encodes this requirement through typed mandates that record authorization, permitted merchants, spending limits, and transaction-specific approval, while the Universal Commerce Protocol supplies the surrounding commerce workflow. Their design reflects something procurement has always required: a match alone cannot complete a transaction. Authority, limits, and evidence of approval have to accompany it.

Supply floods the channels

The cost of producing answers is collapsing at the same time. The 2025 Stanford AI Index found that the price of querying a model at a fixed capability level fell more than two orders of magnitude between late 2022 and late 2024.

A field study published by the National Bureau of Economic Research found that an AI assistant raised customer-support issues resolved per hour by about 14 percent, with the largest gains among the least experienced workers.

Drafting, analysis, specification, creative production, and outreach are moving toward near-zero marginal cost in suitable workflows, with review and integration still owed.

The effects compound across go-to-market. Cold email and calling can be automated cheaply, raising outbound volume and reducing the informational value of another unsolicited message. Regulators are also tightening rules around synthetic commercial signals: the Federal Trade Commission's 2024 rule prohibits fake or false reviews and testimonials, including AI-generated examples that misrepresent actual experience. Beige Media's earlier analysis of the outsourced B2B sales market described the same pressure from the seller's side.

When anyone can produce a convincing product and a convincing pitch, in any volume, at any hour, convincingness stops discriminating. Polish no longer carries information about quality, because polish is free.

The seller's problem

Flattened production costs land hardest on the firms that sell what has been flattened. Software and software services now face competitors who can reach comparable output at a fraction of the historical cost, in a macro environment that leaves little room for initiatives that fail.

At thin margins, pursuing the wrong vertical, specifying the wrong product, or solving the wrong problem can consume the financial room needed to recover.

Sellers therefore have to present the product against a specific, recognized, immediate customer problem. As the buyer's need becomes more legible, mismatches in positioning and product fit become easier to identify.

Mismatches between claims and outcomes can surface faster through feedback cycles that now iterate positioning and go-to-market at automated speed. Thin margins leave less capacity to survive the discovery.

These firms are also caught between two clocks. AI accelerates the informational work surrounding a transaction far faster than the organizational and physical commitments that follow it. A team can draft in an afternoon what its organization can deploy only in a quarter.

Where the deliverable is physical, the gap is widest: a manufacturer can generate twenty design alternatives before lunch and still get one economically meaningful chance to commission a production line correctly. NIST's manufacturing digital thread work uses standardized product information to improve information flow across design and production, directly targeting the lag between digital specification and physical execution.

Brownfield organizations often adopt automation incrementally because existing systems, processes, and obligations limit the pace of deployment.

Most firms straddle both regimes. Marketing copy changes within an hour; the production contract holds for years. A prototype is rebuilt in days; the data, legal obligations, and operating dependencies around it unwind over quarters.

Where costs remain

As production and matching costs fall, the consequences of a bad decision consume a larger share of transaction economics and can rise where automation increases the scale or speed of failure. AI cheapens remediation: rollback, simulation, and insurance can reverse or redistribute some errors. It also cheapens error itself: more convincing fraud, faster propagation of a defect through automated supply chains.

Options can now be generated far faster than many contractual, organizational, financial, and physical commitments can be changed. The cost of generating another option therefore falls relative to the expected cost of a poor decision, especially when detection is slow or reversal is difficult.

Low-cost, reversible decisions reward rapid shipping, measurement, and revision. Where error is expensive or irreversible, value moves upstream of commitment, into verification and controlled authorization before the fact.

Cryptocurrency is the limit case, and the posture there was not chosen by taste. On-chain transfers can become effectively irreversible long before legal remedies can recover assets, particularly once funds have crossed counterparties or jurisdictions. Verification, security, and proactive operational defense are therefore foundational rather than additive: the field cannot rely primarily on after-the-fact remedies because recovery may be impossible by the time misconduct is detected.

The same logic governs wherever remedy is slow relative to harm, which is why the operational-proactivity theses in security, diligence, and systems monitoring describe one phenomenon in different domains. Assurance built before commitment is the general form. Crypto is simply where it was forced earliest, because the cost of being wrong is measured in the millions and billions and there is no second channel.

The same economics help explain the persistence of relational channels. High-consequence deals continue to move through relational channels because a relationship carries what an automated pitch cannot: a record of prior conduct and an accountable party when something breaks.

The introducer stakes future credibility on the recommendation, and a signal that costs its sender something is a signal that carries information. Personalization may improve the odds of engagement, while an established relationship also provides reputation, accountability, and a channel for resolving failure.

Partnerships apply the same mechanism at scale: an established intermediary brings prior customer relationships, reputation, and an existing channel for accountability.

Procurement is the prehistory

The institutions this regime requires already exist in sectors where buyers carry formal obligations and errors have long been expensive.

A K-12 food broker does not primarily sell food; the district's real problem is answering an auditor, and the broker maintains a book of vendors pre-vetted against nutrition rules and bid thresholds.

The federal Multiple Award Schedule gives public buyers access to products and services at pre-negotiated prices through defined procedures, with compliance as a stated feature. Carahsoft, as Beige Media has covered, built an aggregation position on contract access and channel coordination between technology vendors and public buyers.

Amazon combines standardized product attributes with signals such as fulfillment eligibility, ratings, and brand identity, allowing buyers to narrow unfamiliar sellers without independently investigating each one.

Each uses standardization to spread the costs of discovery, comparison, compliance, and risk evaluation across many transactions. A maintained catalog reuses product definitions, contract terms, and evidence requirements instead of rebuilding them for every order.

The relevant question for any intermediary is which of those costs AI removes and which become more concentrated. Discovery and comparison are being automated directly; an intermediary whose advantage is knowing who sells what is squeezed.

The more valuable functions become making heterogeneous suppliers comparable and providing accountability when a transaction fails. What matters is the maintained connection among a supplier's claim, the evidence supporting it, and the terms under which a buyer can rely on it.

Much of Carahsoft's paperwork can be automated, while its position as an established channel between public-sector buyers and outside vendors may become more valuable as purchasing systems demand standardized, machine-readable supplier information.

The assurance ladder

Every assurance signal becomes an attack surface once enough value rides on it. Fake reviews proliferated until the FTC's 2024 rule banned their sale, purchase, and AI-generated fabrication outright.

A verification badge that anyone can purchase converts trust into placement; X's 2022 paid checkmarks produced corporate impersonations, including a fraudulent Eli Lilly account, within weeks of launch. Credit-rating conflicts provide a larger-scale example. Standard & Poor's business depended on firms issuing the securities it rated, and in 2015 S&P and its parent agreed to pay $1.375 billion to resolve federal and state litigation over ratings of residential mortgage-backed securities and collateralized debt obligations issued before the financial crisis.

Beige Media's work on credential extraction describes the mechanism at the level of the individual career: displayed affiliation circulates more easily than demonstrated capacity. The pattern is consistent. Once a trust signal becomes easy to imitate, buyers begin discounting it and place more weight on evidence, independent review, enforceable rules, and meaningful consequences for failure.

Verification remains only one layer of assurance. A transaction can pass every check and still fail, leaving a second question: who is responsible when it does? Warranties, escrow, insurance, reputation, and contractual liability place something of value behind the transaction. The final backstop is risk-bearing itself, because a promise of compensation matters only when the responsible party has the capital, authority, or operational capacity to honor it. A ten-million-dollar warranty from an insolvent shell provides little protection.

Beige Media's analysis of security underwriting shows the condition under which evidence matters: continuous operational evidence changes outcomes only when a carrier acts on it and reprices accordingly.

Evidentiary infrastructure such as append-only records and attributable audit trails can narrow later disputes, while its greatest operational value comes when the evidence reaches decision-makers early enough to stop or correct a decision.

The compensation model of whoever operates the funnel belongs on the ladder too. SEC staff guidance treats a conflict as any interest that may incline a recommendation away from the disinterested one.

The relevant conflict arises when compensation can move a recommendation away from the principal's stated objective, especially when that influence is undisclosed or difficult to audit.

Finance illustrates how compensation models vary with the customer, product, transaction, and degree of discretion entrusted to the intermediary, with commission-based and fiduciary relationships continuing to coexist. An intermediary becomes fiduciary-like when its objective and controls keep the principal's interest ahead of whoever is selling inventory.

Who defines admissibility

Automated purchasing does not invent the demand for standards; it inherits it. A standard determines which attributes can be represented, what evidence supports them, and which remedies accompany the transaction.

Sellers outside the schema can remain visible to people while becoming difficult to transact with through the channels where volume lives. The position can concentrate markets: a supplier flood meets buyers whose thin margins make them risk-averse, and the assurance layer between them can become a chokepoint.

Admission rules favor incumbents who can produce the required evidence, whether or not they are the best performers. Open schemas keep vendors portable across purchasing systems; closed ones bind a buyer's purchasing history and assurance record to a single operator.

The institutional design of a standard can matter as much as its technical quality, whether the standard governs agent protocols or procurement schedules.

After intention

Intention increasingly mediates commerce, but it does not remain the principal scarcity once matching itself becomes inexpensive. Broadcast made attention the binding constraint, and it held for decades because human bandwidth is fixed.

As AI lowers the cost of production and matching, less value remains in identifying a plausible fit and more shifts toward evidence of fit, authority to transact, and credible responsibility when the transaction fails.

The intention era is the hallway between the attention economy and an economy organized around assurance. Procurement, security, underwriting, and other high-consequence fields already provide working examples of institutions built around verification, authorization, accountability, and recovery.

What remains scarce through each round of automation is assurance capacity: capital at risk, legal authority, and the ability to remedy a failed transaction. These cannot be produced by the tools that collapsed everything else, which is why they are the floor of the new structure rather than another layer of it.

As production and matching become inexpensive, commercial power moves toward the systems that determine which claims qualify, which evidence counts, who may authorize a transaction, and what recourse exists when it fails. Attention once determined who entered the buyer's field of view. Increasingly, the consequential question is who defines the conditions under which a buyer, or an agent acting for one, is willing to commit.