AI Agents Will Need A Credit Score For Companies: How Contradictory Records Get Suppliers Dropped From AI Shortlists
Joseph Byrum is an accomplished executive leader, innovator and cross-domain strategist with a track record across multiple industries.
gettyA supplier can be dropped from a buyer’s shortlist by software, and no one at the company will ever know.
The AI agent building that list has to decide whether each company is who it claims to be. One supplier was acquired last year, and its website carries the new name while an official registry still shows the old one and a public file names a former owner. The agent has no neutral way to settle the conflict, so it moves to the next supplier.
The company never learns it was evaluated.
That scene is a projection, and the pieces that would make it routine are being built now.
Advertising offers a public window; its agent registry and standards are open to inspection and comment. The same logic reaches procurement, where Gartner predicts that by 2028, 90% of B2B buying will be AI agent intermediated, with verifiability a prerequisite.
The infrastructure is moving faster than adoption. IAB Tech Lab went from naming its agentic framework to publishing a third-version draft in about seven months, while an IAB survey found the share of U.S. buyers planning more focus on agentic buying fell from 66% to 54% during 2026.
An agent decides who a company is by comparing what independent sources say about it. When they agree, the match is fast and confident. When they conflict, the agent can guess or move on. In my judgment, a procurement agent has little reason to guess.
Human buyers already behave this way. McKinsey’s 2026 B2B survey names inconsistent information across a supplier’s teams as the top reason buyers switch suppliers, and ranks generative AI among the top five channels for researching them. I expect an agent to apply the same intolerance for mismatch, only faster.
The mechanism favors the company whose name, address and owner read the same everywhere and penalizes the one that changed records in one place only, the normal state after a merger.
In a recent advisory engagement, a private equity-backed data company acquired a second firm and merged it under its own name. AI systems confused the combined company’s leadership and products, and favored the acquired firm’s descriptions over those of the surviving brand.
Three conditions combined: By my count, 87% of the independent sources still described the acquired firm, old product and leadership pages stayed live and were never updated, and the acquired firm’s machine-readable records stayed intact and separate, so the systems saw two companies where there was one.
The name on the door did not decide whose description the systems believed.
Three groups are building partial checks. IAB Tech Lab’s agent registry checks advertising agents against domain, legal-entity and privacy records. The foundation behind the global company-identification code has published a discussion paper on a credential that lets a counterparty verify who stands behind an agent. Accountants and an AI underwriting firm are rolling out assurance on how agents behave.
Each system verifies a slice. None verifies the whole.
Visibility dashboards answer a different question: Is the company mentioned? Because AI answers vary from run to run, a single ranking partly measures noise, and a preprint review of 45 studies notes low source overlap and substantial run-to-run variability in commercial audits. None of it tells a machine whether a company’s facts agree with each other.
Credit scores settled who to trust with money. Machines will need an answer to whom to trust with a purchase. The shorthand here is an entity authority score, a measure of whether the facts about a company agree and whether an independent party has checked.
• Independence: Builders should not grade their own work, and platforms carry a conflict because they would grade their own market.
• A Range Of Results: The result should arrive as a range, since one run’s number inherits that run’s noise.
• A Way For Companies To Contest: A wrong score nobody can correct is a shortlist drop with no appeal.
The inputs are plain: whether official records, app-store accounts and the website agree on name, address and owner, and whether the company’s own AI agent repeats the public record.
In my view, some form of entity authority score arrives within one to three years. The best-prepared companies will inventory every place their identity is recorded, from registries and app stores to domains and public files. They will assign an owner to each, correct the mismatches and run an identity check before any acquisition or rename closes.
The forecast fails if buyers and platforms never require such a check by 2028, or if free platform-verified profiles make a separate score unnecessary.
The question is whether your company’s records will agree when the first score is computed.
Forbes Technology is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?