Episode 3: Who answers for this system?

Asked why its AI inventory was never published, a US agency could not say: there was no record, and the people responsible had left. Episode 3 is about the difference between an AI title and real ownership, and the three…

FairByDesign Operational AI Readiness title card: “Episode 3: Who answers for this system?”

Operational AI Readiness, the show. Short episodes, one obstacle each: what goes wrong, what good looks like, and one thing you can do this week. Episode 3 of 12.

Podcast and video: coming soon. Subscribe to the free Mercury Brief to hear when this episode is released.

Imagine an auditor asks why your organization never published the list of AI systems it was required to publish. The answer comes back: we don’t know. There is no record of the decision, and the people who were responsible have left. That is close to what US federal auditors were told this year. It is the clearest public example we have found of the obstacle in this episode: AI that everyone uses and no one, in practice, answers for.

What went wrong: an AI list nobody owned

The US Small Business Administration (SBA) supports small businesses with counseling, capital and help with federal contracting (U.S. Government Accountability Office [GAO], 2026, p. 2). Federal agencies have been required to publish annual inventories of their AI uses under rules dating from 2020 onwards. In May 2026, the GAO reported that the SBA issued its first AI use-case inventory in March 2026 (GAO, 2026, p. 10).

For 2021 to 2023, SBA officials said they believed the agency had no publicly reportable AI. The GAO found at least one reportable system in use, and officials attributed the miss to not looking for AI outside the Office of the Chief Information Officer (GAO, 2026, p. 11). According to the GAO, that system was a machine learning tool that assessed fraud risk in Paycheck Protection Program loan-forgiveness applications and automatically cleared those flagged as low risk (GAO, 2026, note 23).

In 2024 the agency had an internal list and said on its website that an inventory would be published by December. It was not. Officials could not explain why, could not say whether the list had gone to the Office of Management and Budget, and told the GAO there was no documentation of key decisions from that time; the staff responsible for the inventory were no longer at the agency. In February 2026, officials were still working out how to fill those roles (GAO, 2026, p. 11). The GAO pointed to basic internal control standards: assign responsibility, delegate authority for key roles, and document them so knowledge does not leave with the people who hold it. It recommended policies that define roles and responsibilities, and the SBA agreed (GAO, 2026, pp. 11–12).

In September 2026, the SBA’s Office of Inspector General added a second view, covering the period to 30 April 2026. The agency had appointed a Chief AI Officer and made most of its AI uses public, listing 34 (Office of Inspector General [OIG], 2026, memo; pp. 4–5). But its AI governance board, due by July 2025, existed only as a draft charter and had not met as of April 2026 (OIG, 2026, p. 3). None of the 34 uses had been designated “high impact.” The inspector general thought two might qualify, including a fraud-detection pilot for pandemic loans that was not on the public inventory; managers said it was not high impact, but provided no documentation for that determination, which federal guidance requires (OIG, 2026, pp. 5–7). SBA management agreed with all six recommendations and planned to complete the main actions by 30 September 2026 (OIG, 2026, pp. 8–9).

The point is not that the SBA lacked people who cared. It had a Chief AI Officer, a council on paper and an internal list. What it lacked, for long stretches, was the unglamorous middle layer: named people, with written authority, doing specific jobs in a way that survives staff leaving.

Key idea: operational ownership

Operational ownership means one named person answers for how an AI workflow behaves in daily use: what it does, what changed, and what happens when it goes wrong. It is different from a job title, a sponsoring department or a committee. It matters to ordinary people because when an AI-influenced decision about your loan, benefit or account goes wrong, you need someone with the authority to look at your case and change the outcome.

Why ownership goes missing

Chapter 12 of Operational AI Readiness argues that ownership is often in the wrong place. AI governance is handed to legal, compliance or ethics teams while the live system is built and changed elsewhere, so the workflow ends up “narratively governed and operationally undercontrolled” (Bakos, 2026, p. 77). Chapter 4 adds the problem of diffusion: when a recruiting team uses a tool, engineering maintains it, a leader sponsors it and a vendor supplies the model, the answer to “who owns this?” becomes a meeting invitation. Named roles “do not eliminate collaboration. They eliminate diffusion” (Bakos, 2026, p. 35). And Chapter 1 notes that a system can have a product manager, a security reviewer, a sponsor and a vendor contact and still no one can say who has operational authority over its high-impact outputs (Bakos, 2026, p. 19).

What good looks like: three names and one conversation

The free readiness assessment turns this into three questions. Is there one named operational owner, “a person, not a team or a meeting”? Are reviewer and approver roles named, with their authority documented? Is there a named person with authority to pause or stop the workflow, “and do they know it?” Its suggested first move: put the owner’s name and the stopper’s name in the inventory record. The move that proves it: write down reviewer and approver authority, then check that the people named agree.

Chapter 12 adds two tests. The owner should be close enough to the live system to know when a prompt, data source or review queue changed. And a team’s definition of “done” should include ownership, review design, logging, approval logic where needed, a record of changes and a way to interrupt the workflow, because once a workflow influences consequential outcomes, those controls are part of the feature (Bakos, 2026, p. 78). Finally, write it down so it survives someone leaving.

One thing to do this week: add two names to the inventory record for one AI tool: an operational owner and a person who can pause it. Then ask each of them whether they agree, and whether they would know how. If either hesitates, you have found your most important gap.

What the record does not show

Neither report says any person or business was harmed by the gaps described. We have not seen whether the SBA completed the actions it planned for 30 September 2026, so we do not say either way. The GAO’s description of the automatic-clearance tool comes from its earlier work, which we did not reopen for this episode. And the reports describe one agency; nothing here suggests that any individual official acted improperly.

Questions you can ask

  • If you are affected: “Who is responsible for the automated tools used in my case, and how can I reach them to ask for a review?”
  • If you are affected: “Can that person change the outcome, or only pass my question on?”
  • If you run AI at work: “For each consequential AI workflow, can we name one owner, the reviewers and their authority, and one person who can stop it today?”
  • If you run AI at work: “If our AI owner left tomorrow, what would their successor find written down?”

Go further. Operational AI Readiness is FairByDesign’s free field manual for AI that is already live. Chapter 12, “Building an Engineering-Owned Control Culture,” explains why ownership has to sit close to the live system; the ownership checklist opens Chapter 4. Read the chapter free. The first 10 readers who send feedback can email their own questions to info@fairbydesign.org. Once a month, Viktoria Bakos also opens one online session of up to 45 minutes for a leader or team working through a real operational AI or assessment problem.

Your next step: take the free readiness assessment on one AI workflow. Domain 2 covers ownership and authority, and the results show your weakest area, three next moves and the chapter to read. Nothing is stored; it is a self-assessment, not legal advice or an audit.


FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.

References

Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live (Free ed.). FairByDesign. https://fairbydesign.org/operational-ai-readiness/

U.S. Government Accountability Office. (2026). Artificial intelligence: Uses and risks for small business contracting and innovation research (GAO-26-107828). https://www.gao.gov/products/gao-26-107828

U.S. Small Business Administration, Office of Inspector General. (2026). SBA’s progress on implementing federal artificial intelligence requirements (Management Advisory Report 26-15). https://www.oversight.gov/sites/default/files/documents/reports/2026-09/SBA%20OIG%20Report%2026-15%20-%20SBA%27s%20Progress%20on%20Implementing%20Federal%20AI%20Requirements.pdf

Leave a Reply