Our approach to applied AI in government

In April 2025, the Office of Management and Budget issued Memorandum M-25-21, directing federal agencies to appoint Chief AI Officers, establish governance boards, publish AI strategies, and apply defined risk practices to high impact systems. It followed Executive Order 14179, issued in January of that year, which directed agencies to remove barriers to adoption.

Federal policy has therefore resolved the question of whether artificial intelligence belongs in government operations. It has provided considerably less direction on how agencies should build it within the constraints under which they operate: accreditation boundaries, controlled unclassified information handling, records retention obligations, and an obligation to explain adverse determinations to the public.

Conditions in a vendor demonstration differ materially from conditions in an operating mission environment. Most program disappointment in this domain originates in that difference, and it is rarely attributable to model performance.

The following principles govern how Sierron approaches the work.

Begin with the task rather than the technology

The productive question is not where artificial intelligence could be applied. It is which task recurs at volume, follows documented procedure, and currently consumes disproportionate staff capacity.

Tasks meeting those criteria are frequently administrative: correspondence triage, records classification, eligibility pre screening, backlog reduction. They produce measurable results with consistency.

Programs beginning with the technology tend to produce capability without a corresponding justification, and they are typically discontinued when initial support diminishes.

Treat data readiness as the substantive work

In most engagements, the artificial intelligence component constitutes a minority of the effort. The majority establishes what data exists, which system holds authority, how it is classified, who may access it, and whether it may lawfully be used for the intended purpose.

Programs that proceed without resolving these questions do not fail at artificial intelligence. They fail at retrieval, and attribute the failure to model performance.

Govern capability already in use

By the time an agency completes an artificial intelligence policy, staff have generally been using such tools for an extended period. Governance addressing only future authorized deployments leaves existing exposure unexamined.

The NIST AI Risk Management Framework organizes into four functions, Govern, Map, Measure, and Manage, which serve effectively as an operating structure. Alignment to that structure also simplifies M-25-21 compliance at a later stage.

Design for accreditation from the initial architecture decision

Artificial intelligence capability that cannot be accredited cannot be deployed.

Boundary definition, data residency, the FedRAMP posture of each component, logging sufficient to reconstruct a decision, and control inheritance constitute early stage architecture constraints rather than late stage compliance activities. Treating them accordingly determines whether the path to production is measured in months or remains indeterminate.

Instrument outcomes and report them accurately

Interaction counts are not outcomes. An assistant that answers ten thousand questions has produced a volume figure rather than a mission result.

The relevant measures are those the mission already tracks: completion rates, cycle time, backlog, cost per transaction, and error rates under review. Establishing a baseline before deployment is essential. Programs that omit this step typically identify the omission at the point they must justify continued funding.

Application in practice

Sierron maintains a narrow position on where artificial intelligence belongs. A meaningful proportion of our recommendations conclude that a given problem requires a corrected integration, a resolved data source, or a redesigned form rather than a model.

That conclusion is less compelling to present. It is frequently the one that advances the mission.

This view was formed through repeated observation that the technically interesting problem is rarely the binding constraint. A disproportionate share of engagement time addresses which system owns a record, who is permitted to view a field, and what occurs when the answer is incorrect. Considerably less time addresses model selection than initial program discussions anticipate. I previously regarded that groundwork as preparatory to the substantive work. I now regard it as the substantive work, with artificial intelligence representing what becomes achievable once it is complete.

Applied artificial intelligence in government is a mission delivery problem with a technology component rather than a technology problem with a compliance requirement. Programs succeed when they are engineered on that basis.