The 2019 AI order’s lasting demand was an inventory of data and models
Executive Order 13859 told every agency to document its data and models so outsiders could find and use them. Federal AI policy changed direction after it; the demand to know and document data, models and AI uses kept widening.
Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.
Two-minute summaryThe dek and the first sentence of each section, quoted from the article
Executive Order 13859 told every agency to document its data and models so outsiders could find and use them. Federal AI policy changed direction after it; the demand to know and document data, models and AI uses kept widening.
- The week On Monday, February 11, 2019, the President signed Executive Order 13859, Maintaining American Leadership in Artificial Intelligence.
- The data inventory: written for a stranger The order’s data instruction pointed outward.
- The model and code inventory: assets without a home The order put models next to data throughout section 5.
- The barrier register: reasons written down Section 5(a)(iv) is where the order turns operational.
- The traceability record: what “fully traceable” demands The order’s objective asked for “high-quality and fully traceable Federal data, models, and computing resources.”
- The use record: the gap the order left Executive Order 13859 asked agencies to document data and models so that others could use them.
- The strongest objection: records nobody reads The case against this argument is that inventories are cost without visible return.
- What an agency would build today The policy record since 2019 suggests a design for agency data and AI records that does not depend on the current administration’s priorities.
- With hindsight OMB published the notice the order asked for in the Federal Register on July 10, 2019.
The week
On Monday, February 11, 2019, the President signed Executive Order 13859, Maintaining American Leadership in Artificial Intelligence. The Federal Register published it on Thursday, February 14, 2019, at 84 FR 3967. The order named a coordinated federal strategy, the American AI Initiative, and set out five principles for it. They covered technological breakthroughs, technical standards, worker skills, public trust and the protection of civil liberties and privacy, and an international environment that supports American AI research while protecting critical AI technologies from strategic competitors and adversarial nations.
The order then listed six strategic objectives for implementing agencies. One of them reads, in full: “Enhance access to high-quality and fully traceable Federal data, models, and computing resources to increase the value of such resources for AI R&D, while maintaining safety, security, privacy, and confidentiality protections consistent with applicable laws and policies.” The others covered investment in research, reducing barriers to the use of AI, technical standards, training, and an action plan under a National Security Presidential Memorandum signed the same day.
Coordination went to the National Science and Technology Council’s Select Committee on Artificial Intelligence. The co-chairs of that committee would decide which agencies counted as implementing agencies: those that conduct foundational AI research, develop and deploy AI applications, provide educational grants, or regulate and give guidance on AI applications.
Section 5 of the order dealt with data and computing resources, and it was addressed to the heads of all agencies, whether or not they were implementing agencies. Each was told to review its federal data and models to find opportunities to increase access and use by the non-federal AI research community, while protecting safety, security, privacy and confidentiality. The order said agencies “shall improve data and model inventory documentation to enable discovery and usability,” and that they “shall prioritize improvements to access and quality of AI data and models based on the AI research community’s user feedback.”
Section 5 also set a series of deadlines, each stated as a period from the date of the order. Within 90 days, the OMB Director was to publish a Federal Register notice inviting the public to identify requests for access to, or quality improvements in, federal data and models. Within 120 days, OMB was to update implementation guidance for Enterprise Data Inventories and Source Code Inventories “to support discovery and usability in AI R&D.” Within 180 days, agencies were to consider ways of improving the quality, usability and access of priority data that the research community had identified, and identify the resources involved.
The order named the barriers agencies had to identify before opening data or models. Working with their Senior Agency Officials for Privacy, the heads of federal statistical entities and program managers, agencies were to list privacy, civil liberty and confidentiality protections; safety and security concerns, including those related to the association or compilation of data and models; documentation and formatting, including interoperable and machine-readable formats; and changes needed for data and system governance.
Section 6 asked OMB for a memorandum within 180 days to guide agencies’ regulatory and non-regulatory approaches to AI, with a draft released for public comment first. It also asked NIST, within 180 days, for a plan for federal engagement in AI technical standards. The order defined artificial intelligence as the full extent of federal investments in AI, a definition that included “cyberinfrastructure, data sets, and standards for AI.” The order was to be implemented “subject to the availability of appropriations.”
Executive Order 13859 reads as a research and competitiveness order, and a large share of its text supports that reading. It talks about research budgets, fellowships, standards bodies and the protection of American technology from strategic competitors. Read from an agency CIO’s desk, though, its plainest instruction was about records. Every agency head was told to review the agency’s data and models and to improve the documentation that lets someone find and use them. The order asked for data that was “fully traceable.” It asked OMB to update the guidance for enterprise data inventories and source code inventories. The order made all of it subject to the availability of appropriations, and none of it depended on which AI techniques would prove useful.
Our argument is that this inventory instruction was the part of the order that lasted. Successive orders and OMB memoranda since 2019 have changed the purpose of federal AI policy several times: from promoting research access, to principles for the government’s own use, to risk management, to accelerating adoption. Each change kept, and widened, the demand that agencies know and document what data and what AI they hold. For an agency CIO or chief data officer, the stable investment across those swings has been the inventory and the lineage behind it, not any one policy’s vocabulary.
The analysis has five parts, each a different kind of record the order asked for or implied:
- The data inventory: what an agency holds, documented well enough for an outsider to find and use it.
- The model and code inventory: the analytic assets built on that data, which had a weaker home in federal policy in 2019.
- The barrier register: the privacy, security, documentation and governance reasons a dataset or model cannot simply be released.
- The traceability record: where data came from and what was done to it, the condition behind the order’s phrase “fully traceable.”
- The use record: what the agency itself does with AI, which the 2019 order did not ask for and later policy made central.
The first four follow the order’s own text. The fifth is the gap that a later order, signed in December 2020, began to fill, and that later OMB memoranda widened. Two closing sections weigh the strongest objection and set out what we would build today.
The data inventory: written for a stranger
The order’s data instruction pointed outward. Agencies were to increase access and use by “the greater non-Federal AI research community.” That audience matters for how an inventory is written. An internal data catalog can rely on shared context: staff know which system a table comes from, which program owns it and which fields are unreliable. A researcher outside the agency knows none of that. Documentation that is good enough for internal analysts can fail an outside user who needs to know the collection method, the population covered, the time period, the known gaps and the legal terms of use before deciding whether the data suits a research question.
The order did not invent federal data inventories. It referred to existing Enterprise Data Inventories and to OMB Circular A-130 and memorandum M-13-13, which it cited for its definition of open data: “publicly available data structured in a way that enables the data to be fully discoverable and usable by end users.” What it added was a test of purpose. An inventory entry was to be judged by whether it enabled “discovery and usability” in AI research and testing, and agencies were to set priorities based on “the AI research community’s user feedback.”
That second clause is easy to pass over. It made the inventory a two-way channel. An agency publishes descriptions, outsiders say what they need, and the agency improves the assets that matter to them. The public notice OMB was told to publish under section 5 was built on that model. The questions it later put to the public asked what made datasets or models well suited for AI research, and what the government should consider to increase their utility, naming documentation, provenance and metadata as examples.
For a CIO, the practical consequence was that a data inventory stopped being a compliance list and became a product with users. A product has quality measures: how complete each entry is, whether the description matches the data, whether contact points still work, how quickly a request is answered. None of those measures was set by the order, and an agency that wants them has to define them itself. The order’s value was in naming the user.
There is a governance question underneath. A data inventory is only as current as the process that updates it. Where entries are written once, at publication, and never revisited, the inventory drifts from the systems it describes. The order did not set an update cycle for data inventories. Later policy did set one for AI use cases, as the hindsight section shows, and we think the same discipline belongs on the data side: an owner for each entry, a review date and a trigger that reopens the entry when the source system changes.
Priority data, not all data
The order did not ask agencies to open everything. Its 180-day step asked agencies to consider ways of improving “priority data identified by the AI research community” and to identify the resource implications. That is a triage instruction. It accepts that agencies cannot document every holding to research grade at once and asks them to choose by demand.
Triage by outside demand has a weakness that agencies should face directly. Outside users can ask only for holdings they know exist. Holdings that are poorly described, or not described at all, do not attract requests, so a demand-driven process can leave them in the dark. A complete but shallow inventory, listing every holding with a minimal description, is the condition for demand-driven depth to work. The order’s two instructions, improve documentation and set priorities by feedback, work best in that order.
Descriptions a machine can read
Among the barriers the order told agencies to identify was “data documentation and formatting, including the need for interoperable and machine-readable data formats.” The same standard applies to the inventory itself. An inventory published as a document can be read by a person. One published as structured metadata, with consistent fields and identifiers, can be searched, filtered and joined by the tools researchers use to find data.
Structured metadata also serves the agency. If each entry carries the same fields, the agency can count how many assets lack an owner, a description of collection method or a stated access condition, and fix them in order. A free-text inventory hides those gaps until someone reads every entry. For a CDO deciding where documentation effort should go next, the structured version turns a judgment call into a query.
Someone has to answer the requests
A feedback loop needs a person at the other end. The order told agencies to prioritize improvements “based on the AI research community’s user feedback,” but it did not say who in an agency receives that feedback, how fast they answer or who decides what to change. Those are staffing and process decisions, and they fall to the CDO and the CIO by default if no one else takes them.
We would assign three roles. A request owner logs each outside request, acknowledges it and tracks it to an answer. A data steward for each asset decides whether the request can be met and what it would cost. A release authority, with the privacy official involved where personal data is at stake, approves any change in access. The order’s instruction to work with Senior Agency Officials for Privacy, statistical agency heads and program managers describes the same people; it simply did not arrange them into a process.
The request log is also evidence. It shows which assets outsiders want, which requests were refused and why, and how long answers took. An agency that keeps it can show an oversight body that its inventory is used. One that does not can show only that the inventory exists.
The model and code inventory: assets without a home
The order put models next to data throughout section 5. Agencies were to review “their Federal data and models,” improve “data and model inventory documentation,” and identify barriers to access to “data and models.” It also asked OMB to update implementation guidance for Source Code Inventories alongside Enterprise Data Inventories.
In 2019 a federal model was a harder thing to inventory than a dataset. A statistical model at a federal statistical agency, a forecasting model at a science agency and a scoring model inside a benefits or enforcement program are different kinds of asset. Some exist as published methods, some as code, some as a parameter file inside a vendor product, and some as a spreadsheet maintained by one analyst. The order did not define “model,” and OMB’s later request for information told respondents they could define “data,” “data set” and “model” as they wished, so long as they said what they meant.
That freedom was sensible for a request for information. It was a problem for an inventory. An agency that wants to list its models needs a working definition, a minimum set of fields and an owner for each entry. Without them, a model inventory captures whatever the better-organized teams report and misses the rest.
The connection to source code inventories was the order’s practical answer. Code that implements a model can be listed, versioned and, where policy allows, shared. A model that exists only as a configuration inside a commercial product cannot be listed the same way, and the agency’s knowledge of it depends on what the vendor documents. That difference became important later, when OMB’s 2025 memorandum asked agencies to make sure their contracts retained rights to federal data and to improvements made to it, and to keep vendors from using that data to train or improve their commercial offerings without the agency’s express permission.
Fields that serve both purposes
A model inventory written for outside researchers and one written for internal oversight can share a core set of fields. Both need the model’s purpose, the data it was trained or calibrated on, the version in use, the owner, the date of the last validation and the known limitations. The outward version adds terms of access. The internal version adds where the model’s outputs go and who relies on them.
We would build the shared fields once. We think an agency that writes one record per model, with an internal view and a public view drawn from it, spends less over time than one that maintains a research catalog and a governance register separately. The order did not ask for that design. Later policy, which asked for both public inventories and internal oversight of the same AI, made it the efficient one.
Validation evidence belongs in the entry
A dataset entry describes what the data is. A model entry also has to describe how well the model works, and for whom. That means the date and method of the last validation, the data used to validate it, the measures reported and the conditions under which performance is known to fall. A model listed without that evidence invites outside users to rely on it more than its owners would.
Versioning is the companion requirement. Models are retrained, recalibrated and replaced. An inventory entry that names a model without a version cannot tell a user which model produced a past result. We would record each version in use, the date it entered use and the date it was retired, with the validation evidence attached to each. That is the minimum an agency needs to answer a question about a decision made months earlier.
The order’s demand for model documentation fell on statistical agencies, science agencies and program offices alike, with no common template. Setting one, with required fields for purpose, data, version, validation and owner, is standards work a CIO and CDO can do for the whole agency.
The barrier register: reasons written down
Section 5(a)(iv) is where the order turns operational. Before increasing access to data and models, agencies were to identify the barriers and requirements, working with privacy officials, statistical agency heads and program managers. The order named four specific kinds, followed by “any other relevant considerations”:
- privacy and civil liberty protections for affected individuals, and confidentiality protections for individuals and other data providers;
- safety and security concerns, “including those related to the association or compilation of data and models”;
- data documentation and formatting, including interoperable and machine-readable formats;
- changes needed for appropriate data and system governance.
Each is a different kind of reason, owned by a different part of an agency. Privacy and confidentiality belong to the privacy office and, for statistical data, to the statistical agency’s confidentiality obligations. Security belongs to the CISO. Documentation and formatting belong to the data owner and the CDO. Governance belongs to whoever decides who may approve a release.
The order’s phrase about the association or compilation of data and models deserves attention. It recognized that two datasets each safe to release can be unsafe together, and that a model trained on sensitive data can reveal something about it. A register that judges each dataset alone will miss that risk. Later policy came back to it: OMB’s 2025 memorandum said agencies should assess the risks of AI models because they may reveal sensitive details of the data used to develop them, and noted that an agency should not assume a model poses the same privacy and confidentiality risks as its training data.
Why the register is worth keeping
A barrier register does two jobs. It explains to an outside requester why an asset is restricted, which supports the trust the order’s fourth principle asked for. And it gives the agency a list of fixable problems. A dataset restricted because of a formatting gap is a funding question. One restricted because of a confidentiality statute is a legal one. One restricted because no one has authority to approve release is a governance one. Mixing them in a single “not available” status hides which are cheap to fix.
The order also told agencies to “explore appropriate controls on access to sensitive or restricted data and models.” That opens a middle path between public release and none: access for approved researchers, in controlled environments, under agreements. For a CIO, each controlled-access path is a system to build and run, with identity, logging and agreement tracking. The barrier register is where the case for that system starts, because it shows how much demand sits behind restrictions that controlled access could resolve.
Compilation risk at release time
The order’s concern about “the association or compilation of data and models” has a direct consequence for how releases are approved. A release review that looks at one dataset at a time asks whether that dataset, alone, exposes anyone or anything sensitive. A compilation review asks what the dataset exposes when joined to what is already public, including the agency’s own earlier releases and those of other agencies.
That second review needs information a single data steward does not hold: a list of what has already been released, with the fields and geographic and time detail of each. The inventory supplies that list, if it records release history as well as current holdings. In our view, the release history is the part of the inventory that protects the agency most directly, because it is the input to every compilation review.
Models raise the same question in a different form. A model trained on restricted data and then released, or exposed through a public interface, can let a user infer something about the training records. The barrier register should therefore treat a model’s release as a release of information about its training data, and route it to the same reviewers.
The traceability record: what “fully traceable” demands
The order’s objective asked for “high-quality and fully traceable Federal data, models, and computing resources.” It did not define traceable. The plain meaning is that a user can follow an asset back to its origin and through each change: where the data was collected, by which system, under which authority, what was cleaned, merged or removed, and which version a model was trained on.
Traceability is likely the costliest of the five records to build, because it is hard to add after the fact. A dataset’s lineage has to be captured as the data moves, by the pipelines that move it. An agency that copies extracts by hand between systems, or that receives data from vendors and states without provenance metadata, can describe its holdings but cannot trace them. For AI work, the gap matters more than for traditional reporting, because a model’s behavior depends on exactly which records it saw.
OMB gave the term a working definition in 2025. Its memorandum on federal use of AI said that “traceability refers to an agency’s ability to track and internally audit datasets used for AI, and where relevant, key metadata.” It added that clear documentation reflecting the process for model-driven development is “a significant enabler of traceability.” The definition is internal and audit-focused, where the 2019 order’s use was outward and research-focused. The underlying record is the same.
Traceability across the data supply chain
Federal programs depend on data they do not create: records reported by states, grantees, contractors, regulated entities and other agencies. Traceability for that data depends on what the supplier records and passes on. An agency can set requirements in agreements and data standards, but it cannot trace what its suppliers did not document.
That is why the 2025 memorandum’s procurement guidance matters to this argument. It said agencies should treat relevant data, and improvements to it such as cleaning and labeling, as a critical asset, and should keep sufficient rights to federal data and its improvements in AI contracts. An agency that does not hold the cleaned and labeled version of its own data cannot trace a model trained on it. We would draft data rights clauses and lineage requirements together, with the CDO and the contracting officer, rather than one after the other.
Versions as well as files
Lineage for AI work has a requirement that ordinary reporting lineage can skip: a fixed snapshot. A monthly report can be rerun against the current data and produce an acceptable answer. A model cannot be explained, retrained or audited unless the exact training and test data can be retrieved as they were. That means retaining or being able to reconstruct each version of a training set, with the transformations applied to it.
Retention of that kind has costs and legal limits. Records schedules, privacy commitments and storage budgets all bear on how long a snapshot can be kept. An agency that decides those questions in advance, for each AI use case, avoids deciding them during an audit or a dispute. We would record the retention decision and its basis alongside the lineage record, so that a later reviewer can see whether a snapshot is missing by design or by accident.
The 2025 memorandum’s procurement guidance also encouraged agencies, where practicable, to document the provenance of data used to train, fine-tune or operate procured AI. For a purchased system, provenance is something the agency has to obtain from the vendor. Writing that duty into the contract is easier than asking for it after award.
The use record: the gap the order left
Executive Order 13859 asked agencies to document data and models so that others could use them. It did not ask agencies to document their own use of AI. The OMB memorandum that the order required, issued in November 2020 as M-21-06, made the boundary explicit. It addressed regulatory and non-regulatory approaches to AI applications developed and deployed outside the federal government, and it said that “government use of AI is outside the scope of this Memorandum.”
The gap closed under the same administration. Executive Order 13960 of December 3, 2020, Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government, began by citing Executive Order 13859. It set nine principles for agencies’ own design, development, acquisition and use of AI, including that AI be “responsible and traceable” and that relevant inputs and outputs “should be well documented and traceable, as appropriate and to the extent practicable.” And it created a new record: an annual inventory of each agency’s non-classified and non-sensitive AI use cases, including current and planned uses, built to criteria the CIO Council was to publish.
The use inventory is a different object from the 2019 data and model inventory. It lists applications, not assets. One dataset can feed several use cases, and one use case can draw on several datasets and models. For a CIO, the two inventories gain value when they reference each other: each use case points to the data and models it relies on, and each data and model entry shows which use cases depend on it. Without that link, a quality problem found in a dataset cannot be traced forward to the decisions it affected, and a problem found in a use case cannot be traced back to its source.
Scope rules that shape the record
Executive Order 13960 set scope rules that still affect what agencies record. It excluded AI used in defense or national security systems, AI embedded within common commercial products such as word processors or map navigation systems, and AI research and development activities. It applied to existing and new uses, to stand-alone AI and AI embedded in other systems, and to AI developed by the agency or by third parties on its behalf, including relevant data inputs used to train AI and outputs used in support of decision making.
Those rules put judgment calls in the hands of agency staff. Whether a feature inside a commercial product is “common” or mission-specific, and where research ends and use begins, are questions an inventory owner has to answer for each case. An agency that records the reasoning for each exclusion, as well as the result, can defend its inventory when the rules change, and they did change, as the hindsight section shows.
Scope changes, and records that survive them
The definition of who must inventory AI has itself moved. Executive Order 13960 applied to agencies as defined in 44 U.S.C. § 3502(1), excluding the independent regulatory agencies described in § 3502(5). OMB’s 2024 memorandum applied to all agencies defined in § 3502(1), a definition it noted includes “any independent regulatory agency,” and its 2025 memorandum stated that independent regulatory agencies, “which were not included in the definitions of ‘agency’ in Executive Order 13960, are covered by this memorandum.” An independent regulator that kept no AI use record under the 2020 order would have had to build one once the memoranda reached it.
Changes of that kind are where linked records pay off. An agency that already held a register of data assets and models, with notes on which ones fed automated decisions, could produce a use case inventory by working from what it knew. An agency starting from nothing had to survey program offices, ask vendors what their products contained and reconstruct the history of each use. The 2019 order’s data and model inventory, built well, was the head start.
The strongest objection: records nobody reads
The case against this argument is that inventories are cost without visible return. Staff time spent describing datasets and models is time not spent analyzing them. Public inventories can sit unread. An agency budget officer can fairly ask what an inventory produced, and a CIO who answers “compliance” has not answered.
The objection is strongest against inventories built for a single deadline. A list compiled for one report, by surveying offices once, decays as soon as it is filed, and its next version starts again from nothing. That pattern is costly, and with new inventory demands arriving under successive orders and memoranda since 2019, a single-deadline approach would repeat the cost each time. A register kept by the people who own the assets, and reviewed on a schedule, avoids that restart and carries its value from one policy to the next.
Our answer is that the return comes from reuse across questions. A maintained register answers a request for information from OMB, a records request, an inspector general’s inquiry, a congressional question and a security review with the same data. It shortens incident response when a dataset is found to be wrong, because the use cases it feeds are already known. And it supports release decisions, because it holds the release history that compilation review needs. None of those returns appears in the inventory’s own budget line, which is why the case for it has to be made at the level of the agency’s whole data and AI program.
What an agency would build today
The policy record since 2019 suggests a design for agency data and AI records that does not depend on the current administration’s priorities. None of what follows is a requirement unless a current order or memorandum says so; agencies should confirm the current requirements with OMB guidance and their counsel. It is what we think holds up.
- One register, several views. Keep data assets, models and code, and AI use cases in linked records with shared identifiers. Publish the views policy requires: a data inventory, a source code inventory, a use case inventory.
- Owners and review dates on every entry. An entry without an accountable owner and a next review date is a historical note, not an inventory.
- Barriers as structured fields. Record why an asset is restricted using the four categories the 2019 order named, so the restrictions can be counted and the fixable ones funded.
- Lineage captured by pipelines. Invest in data movement that records provenance automatically, starting with the datasets that feed AI use cases.
- Contract terms that keep the record whole. Data rights, improvements, documentation and audit support in AI contracts, drafted with the CDO.
The common thread is that the inventory is infrastructure. Each policy since 2019 has asked a different question of it: what can researchers use, what does the agency use, which uses are high impact, which are ready to scale. An agency that built the records once, with links between them, could answer each new question by changing a query. One that built a separate list for each order rebuilt it each time.
Sequencing the work
An agency starting late does not need to build every record at once. We would begin with the AI use cases already in production, because those are the ones current policy asks about and the ones whose failures reach the public. For each, record the data assets and models it depends on, with owners. That produces a partial data and model inventory shaped by use, which is a better starting point than a survey of every holding.
The second step is lineage for the data behind those use cases, beginning with the ones that inform decisions about individuals. The third is the barrier register and release history, which can grow as outside requests arrive. The order’s own sequence, review holdings first and let outside demand set priorities, suited an agency opening data for research. An agency whose first question is about its own AI use can reverse the order and still end with the same linked records.
Measures that show the records work
An inspector general, an appropriations committee or OMB will judge an agency’s records by whether they answer questions quickly and correctly. Measures we would track include the share of AI use cases whose data sources are linked to data inventory entries, the share of entries reviewed within their stated cycle, the time from an outside data request to an answer, and the number of restrictions by barrier category. These are our suggestions, not federal metrics. They give a CIO evidence that the records are maintained rather than produced once for a deadline.
With hindsight
OMB published the notice the order asked for in the Federal Register on July 10, 2019. It invited the public to identify needs for additional access to, or improvements in the quality of, federal data and models that would improve AI research and testing. Its questions asked which federal data and models had technical issues inhibiting access, which were restricted, what characteristics made datasets or models well suited for AI research, and what the government should consider to increase a dataset’s or model’s utility, naming documentation, provenance and metadata as examples. It also asked what data ownership, intellectual property and data sharing terms belonged in federally funded contracts and grants that produce research data.
NIST announced its plan for federal engagement in AI technical standards on August 12, 2019. The plan called on agencies to strengthen AI standards knowledge, leadership and coordination, to promote research on the trustworthiness of AI systems, to support and expand public-private partnerships, and to engage with international parties.
The regulatory memorandum the order required came on November 17, 2020, as OMB M-21-06, Guidance for Regulation of Artificial Intelligence Applications. It addressed AI developed and deployed outside the federal government and placed government use of AI out of scope. Among its examples of ways to reduce barriers to AI, it said agencies should also follow OMB guidance issued under section 5 of Executive Order 13859 on discovery and usability of federal data and models. It asked agencies with regulatory authority to submit plans for consistency with the memorandum by May 17, 2021, and to post them publicly.
On December 3, 2020, Executive Order 13960 extended the policy to agencies’ own AI and created the annual AI use case inventory. On March 28, 2024, OMB memorandum M-24-10, issued consistent with the AI in Government Act of 2020, the Advancing American AI Act and Executive Order 14110, required each agency, except the Department of Defense and the Intelligence Community, to inventory each of its AI use cases at least annually, submit the inventory to OMB and post a public version on its website.
Executive Order 14179 of January 23, 2025, Removing Barriers to American Leadership in Artificial Intelligence, referred to Executive Order 14110 as revoked and directed a review of actions taken under it. It told the OMB Director to revise M-24-10 and M-24-18 as necessary within 60 days. On April 3, 2025, OMB issued M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, which rescinded and replaced M-24-10. M-25-21 cited Executive Order 13960 alongside Executive Order 14179 and kept the annual use case inventory, with the same exceptions for the Department of Defense and the Intelligence Community. It made each agency’s Chief AI Officer responsible for maintaining the inventory and for making sure custom-developed AI code and the data used to develop and test AI are appropriately inventoried, shared and released in agency code and data repositories where appropriate. It said data used to develop and test AI may be a data asset under 44 U.S.C. § 3502(17), to be included in comprehensive data inventories where the OPEN Government Data Act and OMB guidance require. It required each CFO Act agency’s AI strategy to address access to quality data for AI and data traceability.
As checked on November 15, 2026, M-25-21 was posted on the White House website, and Executive Order 14179 refers to Executive Order 14110 as revoked and does not mention Executive Order 13859. Across these documents the stated purpose of federal AI policy changed, from research access to trustworthy government use, to risk management, to accelerating adoption. The demand that agencies know and document their data, models and AI uses was kept and widened at each step.
Spotted an error? Tell us and we will check it and correct the article where needed.