AI Vendor Checklists Have Quietly Become a Buying Standard

← Back · ACT III · Chapter 22: Embedding Buyer Governance

Enterprise buyers assembled a governance instrument for AI vendors one RFP at a time. It already gates payment on evidence, and it still has no name.

An enterprise procurement checklist for an AI agent vendor in mid 2026 is a different document than it was eighteen months ago. It specifies a kill switch and names who administers it, it requires audit trails that are append-only and tagged to agent identity and model version, and it treats acceptance testing as a statistical exercise against a failure budget rather than a single pass or fail. A quarterly governance review with a board reporting pack now sits in the deliverables schedule instead of in the closing formalities of a sales cycle.

No regulator wrote that document, and no vendor consortium convened to draft it. Buyers wrote it themselves, one RFP at a time, because the alternative was accepting a vendor’s account of what its own product does. Zylos Research, surveying enterprise AI agent deployments in July 2026, found that procurement checklists now “routinely demand kill switches, evidentiary audit trails, human-in-the-loop boundaries, model change control, outcome-based SLAs, and ISO/IEC 42001 or SOC 2 attestations as gating conditions.” A gating condition is a term the vendor meets before it is paid, which is a different thing from a preference on a wish list.

I call that instrument a Request for Value. The buyer defines what counts as proof, sets the threshold before the money moves, and pays against evidence rather than against a promise. That is what these checklists do, so the mechanism is already in the market and running. What it lacks is the name, and the name matters because an instrument without one is hard to improve or to pass on to the next buyer.

The Buyer Filled the Gap the Seller Left Open

The clearest evidence sits in contract law rather than in marketing. Mayer Brown’s technology practice published guidance in February 2026 arguing that standard SaaS terms are “a poor fit for agentic AI solutions,” and proposing a hybrid built from business process outsourcing precedent: a defined delegation of authority, explicit policy guardrails, outcome SLAs measured against business results rather than uptime, and audit rights over decision logs so a buyer can ask why the system did what it did and get a structured answer. One sample metric in that guidance is 99 percent of invoices processed correctly against the purchase order, which is a different promise from 99.99 percent platform availability.

The move from a license to something closer to a services contract changes where the burden sits. The vendor’s obligation stops being to make a tool available and becomes to deliver a verifiable outcome, which means the record that proves the outcome is the vendor’s to produce. A vendor that cannot produce the decision log has failed to perform on terms the buyer wrote, which is a more serious matter than a disappointed customer.

Proof Replaced the Attestation

The Cloud Security Alliance’s AI Controls Matrix, released in July 2025 with 243 control objectives across 18 domains and carrying 247 in its 2026 version, assigns ownership across four layers of the AI supply chain: the model provider, the orchestration layer, the application provider, and the cloud infrastructure underneath. That granularity exists so a hard question cannot be answered with a soft one. A buyer asking about training data provenance is no longer satisfied by a paragraph in a trust center, because the question now routes to a specific control, owned by a specific party in the stack, auditable on its own terms rather than absorbed into one vendor’s blanket assurance.

The market has already seen what happens without that architecture. A Texas healthcare AI vendor advertised a critical hallucination rate below 0.001 percent, a figure the state attorney general’s investigation found was likely inaccurate. The 2024 settlement carried no financial penalty, and what it did require was disclosure of the definition and the calculation method behind any accuracy metric the company advertises, which is the part a buyer needs in order to check a claim at all.

The Buyer Sets the Threshold

The same survey notes something the sellers have left unfinished. A shared responsibility matrix is the one document that would answer a buyer’s ownership question before it is asked. Microsoft publishes an AI shared responsibility model for its own platform, and Zylos still found that no major vendor has done this well yet. Buyers built the requirement anyway and are procuring around the gap rather than waiting for the market to close it. That is how buyer power has always accumulated, one clause and one refused purchase order at a time, until the accumulation functions as a standard whether or not anyone convened to write one.

I wrote earlier about a market that adopted the word value before it agreed on a definition. This is the same failure running in the other direction, and it is the more useful one, because an instrument that already works can be named later. When the Linux Foundation stood up a body to standardize the economics of AI, it opened with a measure of substitution rather than a measure of realized outcome. The buyers writing these checklists are already asking the harder question, in the one document a vendor cannot ignore.

The instrument is in use. Naming it is the part still outstanding, and it decides whether the next buyer inherits this work or rebuilds it from scratch.

References: Mayer Brown, “Contracting for Agentic AI Solutions: Shifting the Model from SaaS to Services,” February 17, 2026, mayerbrown.com. Cloud Security Alliance, “Introducing the CSA AI Controls Matrix,” July 10, 2025, cloudsecurityalliance.org, with the 247-objective figure from the 2026 version of the matrix. Zylos Research, “Buyer-Side Governance: What Enterprise Customers Now Demand From AI Agent Vendors,” July 2, 2026, zylos.ai. Texas Attorney General, settlement with Pieces Technologies, Inc., 2024, texasattorneygeneral.gov.

Keep reading: ACT III · Chapter 22: Embedding Buyer Governance