Launching an AI Model Proves Little Until the Results Show Up

← Back · ACT II · Chapter 9: The Value Control Loop

Domino Data Lab’s fifth annual enterprise AI report came out on July 21, 2026. The survey behind it, run by BARC Research in April, asked 639 senior AI leaders in North America, the UK and continental Europe how their AI work is going. Thomas Robinson, Domino’s COO, opened the release with a line most AI teams will recognize: “Getting a model into production used to be the milestone that mattered.” The milestone that counts now, he said, is “the moment a business user can act on what the model found.” Domino sells a platform for building and governing exactly these applications, so it has a stake in that conclusion. The figures below are from BARC’s fieldwork.

The report calls the distance between those two moments the last-mile gap: models running in production while the people who need their output cannot get to it, or cannot act on it. Forty percent of the organizations surveyed still rely entirely on at least one mediated route to AI output, either a scheduled report from the data science team or a request to an analyst who runs the analysis and sends the results back.

I read that gap as the point where a claimed capability meets the definition of a real one and falls short. An organizational capability is a demonstrated ability to produce a defined outcome repeatedly, without dependence on specific individuals, at a cost the organization knows, under conditions that change. Demonstrated means evidence exists at the point of outcome. A model in production has demonstrated nothing yet. It is an ingredient in position.

The Milestone Was Always Self-Graded

Deployment is a milestone the deploying team awards itself. Every check that earns it, from the model running to the release going out, happens inside that team, which is why the milestone feels so achievable and proves so little. The outcome sits at someone else’s desk, with the underwriter who prices a risk differently or the controller who closes the books a day sooner. Until there is evidence at that desk, the organization is holding a claim. Celebrating at deployment grades the half of the work the team already controlled and skips the half the definition actually tests.

Claims Are Scaling Faster Than Verification

The same survey shows where this goes next. Expanding agentic AI ties with upskilling business users as the top enterprise AI priority for 2026, each named by 38.5 percent of respondents. Meanwhile 41 percent of organizations are piloting or scaling agentic AI without the governance to manage it, and the ones scaling outnumber the ones piloting by more than two to one, so claims about what agents deliver are growing where the checking is missing.

The organizations that did build the checking look different. Where governance fully keeps pace with AI activity, 67.5 percent have agentic AI running in governed production, against 17.2 percent where it only partly keeps pace, which is where the report’s figure of 3.9 times comes from. Where AI governance is fully integrated, 75 percent report significantly improved delivery velocity, against 23 percent where governance is falling behind. Financial services, banking and insurance, among the most regulated sectors in the survey, lead every vertical on both governance maturity and production velocity, and Domino’s own reading is that the enterprises succeeding there “built governance infrastructure first, then scaled.” A survey like this shows the two moving together without proving that one caused the other. What it does not show is checking slowing anyone down.

Verification Is Part of the Ability

The instinct is to treat governance as overhead bolted onto capability. The definition points the other way: if a capability has to be demonstrated, the checking is part of the ability itself. An organization that cannot show its outcome at the point of use, at a cost it knows, has an unverified claim with infrastructure behind it. And unverified claims compound, because each new deployment built on an undemonstrated one adds to a stack of assertions no one has checked.

The test is short and anyone can run it: name the outcome and the desk where it lands, then show evidence that it landed there more than once, without the two or three people who currently make it work.

What Running The Test Looks Like

The test has been run in public on one narrow kind of capability, and the result is worth knowing. SkillsBench, an academic benchmark first published in February 2026 and revised in June, took 87 tasks across eight domains and ran each one with and without a curated skill meant to help an AI agent, across 18 combinations of model and agent software, with a deterministic check deciding whether each task was done. Curated skills raised the average pass rate from 33.9 percent to 50.5 percent. The gain ranged from 4.1 to 25.7 points depending on the setup, so the only way to know what a given skill was worth in a given setup was to run it.

A second study shows what happens when nobody runs the test. Researchers at Palo Alto Networks examined 49,943 skills in the OpenClaw registry and found that 80.0 percent deviate from the behavior they declare, mostly through developer oversight rather than malice. Palo Alto sells security products and has an interest in that finding, but the point here does not depend on the motive: a capability that has only been described, once it is checked, usually turns out to do something other than what was described.

The same gap appears at the scale of a warship, in a buyer’s acceptance step. Malaysia’s National Audit Department reported this month on the navy’s Littoral Combat Ship program. As Bernama reported it on October 5, the audit found that “LCS 1 will be handed over without full operational capability, as training and weapon systems have yet to be completed,” and recommended that “any conditional acceptance of the vessels clearly sets out the contractor’s obligations, completion timelines”. A handover date works like a deployment date. It records that the ship arrived, and the auditors’ recommendation keeps the buyer’s acceptance tied to what the ship still has to be able to do.

A claim that stops at deployment is a loop left open, and the loop must close before the capability is real.

An audited case of the same gap in installed security equipment, where every delivery metric read complete and the auditors could not verify the upgrades, is in An Australian Audit Found Upgrades Marked Complete but Unverified.

References: Domino Data Lab, “AI ROI Still Fails to Outpace Spend for 57%, as 93% Improve Production”, press release for the Fifth Annual Domino Enterprise AI Report, July 21, 2026. Survey conducted independently by BARC Research on behalf of Domino Data Lab, April 2026, 639 senior enterprise AI leaders at Director level and above in organizations with annual revenue of $100 million or more. Xiangyi Li and others, “SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks”, arXiv:2602.12670, February 2026, revised June 2026. Yuhao Wu, Tung-Ling Li and Hongliang Liu, “Behavioral Integrity Verification for AI Agent Skills”, arXiv:2605.11770, May 2026. Bernama, “LCS Project 7.55 Pct Behind Schedule, National Audit Dept”, October 5, 2026, reporting the Auditor-General’s Report 2/2026.

Keep reading: ACT II · Chapter 9: The Value Control Loop