73 of 82: The USDA AI Tools That Never Got Permission to Operate

Every federal computer system is supposed to get a formal security sign-off before it runs. The Agriculture Department’s inspector general found that 73 of USDA’s 82 operational AI use cases never got one, and that USDA had put AI rollout…

FairByDesign The Receipts Room title card: “73 of 82: The USDA AI Tools That Never Got Permission to Operate”

The Receipts Room

The US Department of Agriculture is not just about farms. Its mission areas, as its inspector general lists them, run from Farm Production and Conservation to Food Safety and Food, Nutrition, and Consumer Services. If you farm, eat, or rely on food assistance, its systems touch your life. So when that department switches on a piece of AI, you have a stake in whether anyone checked it first.

In the US federal government, a computer system is not supposed to switch on until someone senior signs a piece of paper. That paper is called an authorization to operate. It records that the system’s security risks were assessed and that an accountable official accepted them. It is one of the oldest and least glamorous controls in government IT, and it applies to AI just as it applies to a payroll system.

In May 2026, USDA’s own Office of Inspector General reported that most of the department’s AI never got that signature. Of the 82 AI use cases USDA listed as operational, 73 “did not have an ATO prior to operation and were not recorded” in the department’s security tracking system (USDA Office of Inspector General [USDA OIG], 2026, p. 5). Of the nine that did have an authorization, two were missing required documentation.

Here is my proposition: USDA’s AI gap was not a clerical slip. On the inspector general’s evidence, its process put speed ahead of the paperwork that makes someone answerable, and the department cannot be fully sure what it is running. Let’s see whether the report earns that.

The sentence that explains everything

The inspector general’s explanation for how this happened is the most useful sentence in the report. USDA’s process “prioritized AI implementation over cybersecurity and governance controls outlined in federal guidance” (p. 3).

Who benefits from that trade? In the short run, the teams who wanted the tools working. Who carries the risk? Everyone whose information or service passes through a system that no named official has formally signed off.

The inspection, numbered 50801-0018-12, set out to determine whether USDA had put cybersecurity controls in place for its critical AI systems in line with federal standards (p. 2). Inspectors worked from July 2025 to March 2026 and reviewed USDA’s 2024 inventory of AI use cases, the most recent available when they began; the 2025 inventory was not published until January 2026 (p. 9).

USDA’s own rules are clear. Every USDA IT system, including contractor and cloud systems, must have an authorization to operate before it is placed into operation, and the supporting assessment documents must be kept in the department’s Cyber Security Assessment and Management tool, known as CSAM (pp. 3, 5). For AI, the inspectors found, this mostly did not happen.

The receipt: 73 of 82, everywhere you look

The gap was not confined to one corner of USDA. The report’s Exhibit B breaks down active AI use cases by mission area as of 31 December 2025 (p. 12):

USDA mission area or office (as labelled by OIG) No ATO ATO
Natural Resources and Environment 29 3
Research, Education, and Economics 15 2
Farm Production and Conservation 13 1
Marketing and Regulatory Programs 6 1
Departmental Administration and Staff Offices 5 2
Food Safety 3 0
Food, Nutrition, and Consumer Services 2 0
Total 73 9

Source: USDA OIG (2026), Exhibit B, p. 12.

The report does not name the individual tools or say what each one does, so we cannot tell you which systems are involved. What it does say is why this matters: without an authorization, management “does not have assurance cybersecurity controls are in place,” and no named official has formally accepted the risk (p. 5).

Key idea: Authorization to operate

An authorization to operate, or ATO, is a formal decision by a senior official that a system’s security risks have been assessed and are acceptable, so it may run. It names who is responsible, ties the system to a record of its controls, and triggers ongoing monitoring. For the public, it is the difference between “someone checked this and put their name to it” and “it was switched on because it was useful”. When AI tools skip this step, nobody has formally owned the risk of what they can reach or reveal.

Deadlines on the calendar, and missed

The security gap sat alongside a governance gap. The White House Office of Management and Budget’s April 2025 memorandum M-25-21 set agencies a timetable for AI governance.

To be fair to USDA, it met some milestones. The inspector general found it appointed a Chief AI Officer, produced a compliance plan and an AI strategy, and publishes its inventory each year. But it had not updated its agency policies or issued a generative AI policy, both due by 29 December 2025, and it had not put in place the minimum risk management practices for high-impact AI due by 3 April 2026 (pp. 3–4).

When the inspectors asked for USDA’s AI policies and procedures, officials in the Office of the Chief Information Officer said they had not updated or developed any (p. 4).

The inspectors also picked eight AI use cases whose descriptions suggested they might be high-impact under the OMB memo, meaning their output could be a principal basis for decisions affecting things like people’s rights, access to government services, or health and safety. They found that USDA “had not performed or documented its reviews as required” to decide whether they were (pp. 4, 9).

The twist: even 82 might be wrong

Even the number 82 comes with a warning. USDA builds its AI inventory through an annual data call that asks the owners of each use case to report it themselves. The inspectors concluded that the inventory “could not be verified as complete and accurate” because of that sole reliance on self-reporting, and warned that this can “create a false sense of security” (pp. 4–5).

The risk they name is “shadow AI”: tools in use without formal approval or oversight, which could mean systems on the network USDA does not know about (pp. 1, 5).

In other words, 73 of 82 is a count of the AI USDA knew about. The inspector general could not say how much it did not.

What USDA promised to fix

The inspector general made four recommendations to USDA’s Office of the Chief Information Officer, and the office agreed with all four (pp. 6–8):

  • Put controls and a department-wide regulation in place so that high-impact assessments are done as OMB requires. USDA said it would adopt the AI impact assessment template developed through OMB’s Chief AI Officers Council, with an estimated completion date of 30 September 2026.
  • Review and update policies and procedures to cover AI, by 31 December 2026.
  • Create a process to keep the AI inventory continuously up to date. USDA proposed an AI checklist for all IT procurements, approved or denied by the Chief AI Officer, by 30 June 2026.
  • Ensure a risk assessment, ATO determination and system impact analysis happen before AI is allowed on the USDA network, by 31 December 2026.

When we checked the inspector general’s website in mid-October 2026, it listed the inventory recommendation as closed and the other three as open, including the impact assessment recommendation whose estimated date had passed (USDA OIG, n.d.).

The best case for USDA

USDA’s response makes a point worth taking seriously. Where AI comes through a service with an existing security authorization, such as a FedRAMP-approved cloud product, a separate risk assessment may not be needed (USDA OIG, 2026, agency response). So some of the 73 may be running inside products that someone, somewhere, has already vetted.

That is a reasonable point, and it cuts both ways. It is one more reason an accurate inventory matters: you cannot rely on someone else’s authorization for a tool you have not recorded.

And the inspector general’s closing argument still stands. Governance and cybersecurity controls need to be “factored in prior to deployment,” because addressing security and privacy afterwards “is not only much more difficult and expensive, but it also exposes USDA to unnecessary risk” (p. 6). USDA told the inspectors it was exploring a third-party AI testing contract, a way to flag AI systems in its security tracking tool, and a draft AI playbook, but none had a final implementation date as of March 2026 (p. 6).

The verdict

Does the evidence earn my proposition? On the trade-off, yes: the “prioritized AI implementation over cybersecurity and governance controls” finding is the inspector general’s own, not my gloss. On “cannot be fully sure what it is running”, also yes, because the inspectors said the inventory could not be verified as complete. Where I have to hold back is on harm, and the limits matter.

The report does not describe any breach, data leak or misuse involving USDA’s AI, and it should not be read as saying one occurred. A missing authorization means the risk was not formally assessed and accepted. It does not prove the tools lack security controls, and some may run inside products that carry their own authorizations.

This is not unique to USDA. Operational AI Readiness puts it bluntly: “If the organization cannot list its consequential AI-enabled workflows cleanly, it is not yet governing AI. It is merely surrounded by it” (Bakos, 2026, Ch. 4).

What we still don’t know

The report does not name the 82 use cases, say what data they touch, or say whether any of the eight sampled cases is in fact high-impact. It also cannot tell us how many AI tools were in use outside the inventory. And the department’s progress since May is visible only through the recommendation status on the inspector general’s website.

Which leaves the question we will keep asking: if one department found 73 unsigned AI tools when it finally looked, what would the count be in agencies where nobody has looked yet?

Questions you can ask

  • If you deal with a government agency that uses AI: agencies publish AI use case inventories. You can ask whether a tool involved in your case is listed, and who is accountable for it.
  • If you work in government: ask whether the AI tool you use has been through your agency’s security authorization, and who signed it off.
  • For security and IT leaders: can you produce, today, a list of every AI use case with its authorization status and the name of the official who accepted the risk?
  • For anyone buying AI: when a product relies on a vendor’s security authorization, record that in your inventory and confirm it covers how you actually use the tool.

Go further. Operational AI Readiness is FairByDesign’s free field manual for AI that is already live. Chapter 4, The Minimum Viable Control Stack, starts with system inventory for exactly the reason this report shows, and Chapter 11, What Auditors, Regulators, and Internal Reviewers Will Actually Look For, explains the records an inspector will ask for. Read it free, or test your own organization with the free readiness assessment, starting with domain 1, inventory and scope.

Next step: we track inspector general and audit reports on government AI and follow up on what agencies said they would fix. To get that analysis as it lands, join Mercury Intelligence for $10 a month (or $100 a year; see what is included).


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. FairByDesign. https://fairbydesign.org/operational-ai-readiness/

USDA Office of Inspector General. (2026, May 12). Cybersecurity of artificial intelligence technology at USDA (Inspection Report 50801-0018-12). U.S. Department of Agriculture. https://www.oig.usda.gov/sites/default/files/reports/2026-07/50801-0018-12_FR_508.pdf

USDA Office of Inspector General. (n.d.). Cybersecurity of artificial intelligence technology at USDA [Report page with recommendation status]. Retrieved October 16, 2026, from https://www.oig.usda.gov/reports/inspection-evaluation/cybersecurity-artificial-intelligence-technology-usda

Leave a Reply