TK
OMB Memorandum M-24-10 and the NIST AI Risk Management Framework both assume federal agencies operate with AI-ready data.[1][2] The agencies do not. The federal AI procurement cycles in market right now are landing on data estates that have never been comprehensively audited for quality, classified for retrieval, or instrumented for lineage. The gap between policy assumption and operational reality is wider than most agency leadership recognizes, and the cost of closing it is one of the largest hidden line items in federal AI program economics — paid either upfront, before the pilot launches, or in remediation after the pilot fails to ship.
What "AI-ready data" actually means
The phrase "AI-ready data" gets used loosely across federal AI conversations. Stripped of the marketing layer, it has a specific operational meaning. Federal data is AI-ready when it satisfies five dimensions simultaneously, against the workloads the agency intends to run.
Federal data is AI-ready only when all five dimensions hold simultaneously.
High scores on four of five do not average. The weakest dimension is the binding constraint.
The five dimensions are not new — federal records officers, database administrators, and data stewards have been measuring them for decades. What is new is the requirement that they hold simultaneously. A federal records environment that scores 90% on completeness, 85% on accuracy, 40% on classification, 30% on lineage, and 50% on accessibility is unusable for federal AI retrieval workloads, even though three of its five scores look fine in isolation. AI requires the weakest dimension to be strong enough. The composite score is what determines whether the pilot ships.
Each dimension fails in its own way when AI workloads land. Incomplete data produces incomplete retrieval results, which the model fills in with confident hallucination. Inaccurate data produces accurate-sounding wrong answers. Poorly classified data produces queries the agent can't target. Untracked lineage makes the audit trail unverifiable. Inaccessible data makes the whole stack inert. The model handles none of these problems; the data handles all of them.
Where federal agencies actually are
The federal data estate is not where the policy framework assumes it is. Across federal agencies with AI deployments in market or in pilot, the distribution of overall data readiness skews heavily toward the lower end of the maturity range.
The bulk of the federal estate sits in the lower-middle maturity bands. Almost nothing is fully ready.
The policy assumes the right side of this chart. The operational reality is mostly the left.
The shape of the distribution matters more than the precise percentages. The bulk of the federal estate sits in the lower-middle bands — agencies that have begun cataloging their data but have not yet measured it against the five dimensions, or that have measured one or two dimensions but not the full set. A small minority of agencies sit in the upper bands. Almost none are fully AI-ready by the operational definition.
This is not because federal data is uniquely bad. It is because federal data was built over the past two decades against operational workflows that did not require AI-ready quality. Case management, FOIA processing, contract administration, scientific data submission — all of these tolerated data quality issues that an automated retrieval system cannot. The data is fit for its original purpose. It is not fit for the new purpose AI workloads are creating, and nobody planned the transition.
"Federal AI policy assumes the data is ready. The procurement cycles assume the data is ready. The model vendors assume the data is ready. The data is not ready. The cost of closing that gap is real, and the only choice is whether to pay it before the pilot or after."
The sequencing problem
Federal modernization budgets typically fund infrastructure first, applications second, data quality third. The sequence is almost always wrong for AI.
Federal procurement economics push agencies toward the wrong sequence. The right sequence costs less and ships sooner.
Same components, different order. The order is the variable.
The standard sequence — build the data lake, deploy the AI workload, discover the data quality problems in production, fund remediation reactively — costs meaningfully more than the inverted sequence over the lifetime of the program. The reasons are economic, not technical. Remediation done while AI workloads are in production has to be done with the lights on, with active users, with audit obligations already accumulated. The same remediation done before deployment is straightforward project work with clear scope and a defined end.
Federal procurement economics push agencies toward the standard sequence anyway. Data quality work is hard to scope, hard to demonstrate value for, and easy to defer when budget pressure hits. Infrastructure spend is visible, easy to procure, and politically defensible. AI deployment spend has executive sponsorship and external visibility. Data quality spend sits between them — necessary, invisible, and structurally underfunded.
The agencies that invert the sequence — fund and complete a data quality remediation cycle before launching an AI workload — pay less in total program cost, ship pilots that work, and avoid the remediation cycle that catches the agencies that did not. The inversion is rare. It should not be.
The maturity model
Federal data readiness for AI is not binary. It moves through a defined sequence of stages, each with characteristic capabilities and characteristic gaps.
Federal data readiness moves through four stages. Most agencies sit in Stages 1 and 2.
Stage 2 to Stage 3 is the realistic next move for most agencies — and the one this budget cycle should fund.
The four-stage model is descriptive, not prescriptive. Most federal agencies sit in Stage 1 or Stage 2 today. Stage 3 is reachable inside a 12–18 month remediation program with focused effort and budget. Stage 4 — full operational AI-readiness — takes longer and requires sustained governance investment, not a one-time cleanup project. Each transition between stages has predictable costs and predictable durations; the variability is mostly in how much the agency chooses to fund.
The agencies that will be deploying federal AI successfully over the next three years are the agencies that move from Stage 2 to Stage 3 in this budget cycle. The agencies that defer that transition will discover the gap when their AI workloads hit production, and the remediation will then happen with worse cost economics.
What this rules in and out
Four strategic conditions reshape what federal CIOs should be funding through FY26 and into the next budget cycle:
- Data quality is a procurement prerequisite, not a downstream concern. Every federal AI procurement should include a data-readiness assessment scoped before the model selection, with remediation work funded as part of the program — not as a separate later project. Programs that defer the data work end up paying for it twice and shipping later than planned.
- The five dimensions are non-substitutable. Federal data that scores high on four of five dimensions is not 80% AI-ready; it is unready against the dimension that scores low. The weakest score is the binding constraint. Procurement evaluations that average scores across dimensions are measuring the wrong number.
- Stage 2 to Stage 3 is the realistic next move for most agencies. Federal CIOs trying to leapfrog directly from Stage 1 or Stage 2 to Stage 4 will overcommit budget on capabilities the agency cannot yet operate. The deliberate path moves through stages; the budget cycles align to stage transitions.
- Sustained data governance investment matters more than one-time cleanup. Data quality decays without governance. Agencies that fund a one-time remediation project without standing up the governance function to maintain quality will discover the same gap reappearing within two years. The cleanup is the start of the work, not the end.
The decision
Federal data is not AI-ready. The policy framework assumes otherwise, the procurement cycles assume otherwise, and the model vendors assume otherwise — but the data itself is what it is, and the AI workloads will land on it as it is. The decision for federal CIOs is whether to fund the data readiness work as a deliberate prerequisite to the AI program, paid upfront with clear scope, or as a forced remediation after the pilot has failed to ship, paid in production with broken operational conditions and accumulated audit obligations. The cost is similar either way. The timing changes whether the program delivers.[4]
