Artificial intelligence deserves a tougher business model: payment for results, not applause for a demonstration. In the United States, the practical conversation is already reaching ordinary business operations. A Small Business Administration listing for a September 17, 2026, online workshop focused on missed calls, reminders, neglected follow-ups and repeated data entry. That is a useful starting point for the economics of AI: identify an expensive problem before choosing the software. A clever machine without a paying customer is still an expense. The available adoption evidence suggests an opportunity, but not an untouched gold mine. In its May 26, 2026, analysis, the United States Census Bureau reported that business AI use fluctuated between 17% and 20% across survey periods from December 2025 through early May 2026. Those are historical observations, not a September adoption estimate. The bureau also explained that it broadened its question in November 2025 from AI used in producing goods or services to AI used in any business function. That distinction matters: drafting an occasional email and building a profitable operation around automation are not equivalent achievements. Theory one is production arbitrage: use AI to reduce the time required to deliver something customers already buy, then retain the difference. A hypothetical specialist could sell edited product descriptions, translated training materials or tailored sales presentations at an agreed project price. If production and checking become cheaper without damaging quality, the margin improves. The attraction is straightforward: a service can be tested before an elaborate product is built. The weakness is equally straightforward. Faster production creates no automatic advantage in finding customers, earning trust or defending prices against competitors with similar tools. This model makes most sense when the seller contributes expertise, distribution or judgment that the customer cannot obtain from a prompt alone. Theory two is the automation agency: sell the installation and maintenance of systems that make another business more productive. Its strongest argument is that productivity gains can be real. A study published in the May 2025 issue of The Quarterly Journal of Economics examined 5,172 customer-support agents and found that AI assistance increased issues resolved per hour by 15% on average, with substantial differences between workers. That finding supports a specific workplace use, not a universal earnings promise. An agency would still need to establish whether its particular installation saves paid overtime, expands useful capacity or improves customer retention. Minutes saved are not automatically dollars earned. The weak version of either theory is the passive-income pitch: buy access, switch on the technology and wait for money. Official enforcement records offer a reason for skepticism. In April 2026 congressional testimony, the Federal Trade Commission described obtaining bans against several business-opportunity operators that had made AI-related claims, including Ascend Ecom. The testimony said the commission had accused that operation of taking at least $25 million from consumers through deceptive promises. That history does not condemn AI entrepreneurship. It condemns the substitution of impressive terminology for evidence about customer outcomes and business economics. The more interesting alternative is an AI-assisted purchasing recovery service: find money a business has already overpaid, document the discrepancy and earn a negotiated share only when the recovery is realized. A proposed pilot could concentrate on one wholesaler's supplier invoices, payment records, purchase orders and credit notes. AI would help extract information from inconsistent documents and identify possible mismatches; ordinary calculations and human review would determine whether a duplicate payment or missing credit actually existed. Every candidate recovery would need an evidence packet linking the original documents. This is a proposed application, not a claim that an off-the-shelf model can reliably audit a business without supervision. The commercial difference is accountability. Rather than selling a chatbot or a vague promise of efficiency, the operator would sell a documented recovery process. The agreement should exclude discrepancies already identified by the customer's staff, establish a lookback period and define whether payment depends on a cash refund or an actually applied credit. A hypothetical engagement illustrates the arithmetic: $20,000 in eligible, realized recoveries at a negotiated 20% success fee would produce $4,000 in revenue. If total engagement costs, including sales effort, software, human review and allocated overhead, were $1,500, the remainder would be $2,500 before tax. Every figure is an assumption, not a market rate or an earnings forecast. An engagement with no recovery could lose money. The proposed safeguards are as important as the sales pitch. Access should begin with authorized, read-only records, limited to the information necessary for the engagement. The operator should agree retention and deletion terms, test extraction accuracy and require a qualified reviewer to approve every claim before a supplier receives it. A language model should not authorize payments, alter bank details or invent the contractual basis for a refund. Where a spreadsheet or conventional matching rule solves the problem, it should replace the AI step. The service should also avoid presenting commercial invoice reconciliation as a regulated audit or offering legal or tax conclusions beyond its competence. The hardest question is repeatability. A first review could uncover a backlog that never returns, while successful prevention could reduce future recovery fees. The operator would therefore need to test whether a separately priced monitoring service delivers enough continuing value to earn renewal. Prospective customers should be selected for sufficient transaction volume, accessible records and a willingness to validate results, not merely enthusiasm for AI. The Census Bureau's May 7, 2026, release made supplemental data available on adoption by industry and business size, as well as business uses and supported tasks. Such evidence can help frame market research, but only a paid pilot can establish whether this particular service has a business.