Ask almost any program office in a CFO Act agency about enterprise artificial intelligence and you’ll hear a version of the same exchange. The mission side is ready to put AI to work on real operations. The chief information officer organization is buying tools, standing up secure environments and fielding copilots faster than anyone expected two years ago. But when the program office asks about strategy, governance, data readiness or integration, the honest answer is often “We’re not there yet.”
That answer is usually accurate. It’s also an open question about who will do the work of getting there. My argument is simple: Most of what determines whether AI works in your operation is knowledge you already hold. Readiness isn’t something you wait for. It’s something you build.
Why pilots stall
Industry numbers vary by study and definition, but they point the same direction. A preliminary 2025 report from the Massachusetts Institute of Technology’s Project NANDA found only about 5% of integrated AI pilots were producing substantial financial value. S&P Global found 42% of surveyed organizations had abandoned most of their AI initiatives before production. Those aren’t federal failure rates. The pattern is what matters: the model usually works. What breaks is everything around it. Nobody defined “production-ready.” Nobody set a baseline, so nobody could prove value. Employees didn’t trust tools that were never embedded in their workflows. Governance was either an afterthought or a bottleneck, and staff routed around it.
One point matters more in government than almost anywhere else — AI amplifies whatever data environment it lands in. A model reading three conflicting records for one case doesn’t clean them up. It can reproduce the conflict at machine speed, and when it writes back to a benefits system, a case management platform or personnel records, the blast radius is bigger than any one person’s mistake. Buying the tool and being ready for it are different things.
What the business side actually experiences
Technical roadmaps describe what’s being deployed. What the business owner experiences is a steady stream of technology change without the organizational change management needed to absorb it.
The stack shifts underneath us. Models are swapped or deprecated on a cadence measured in weeks, and a workflow validated against one model behaves differently against the next. Vendors ship AI into products we already own as routine releases, with marketing but no impact analysis, no revised procedures, no training and often no clear statement of what data the feature touches. Nobody owns the job of translating “Version 4.2 adds AI summarization” into what a claims examiner or human resources specialist should now do differently. And a product assessed without AI can quietly become an AI system, in the Office of Management and Budget sense, mid-contract.
None of this is a knock on any CIO shop. Acquisition and security answer “Can we use it?” What is often missing is someone accountable for how the organization actually absorbs that change, and it’s the same gap every prior platform shift eventually had to close. (Lift and shift to cloud?)
We’ve been here before
Economic historians, most famously Paul David, showed that factory electrification paid off only after owners stopped bolting motors onto steam-era layouts and redesigned the floor, the jobs and the supervision. The enterprise resource planning (ERP) wave of the 1990s produced working software and spectacular implementation failures. Federal IT’s own history of failed acquisitions helped drive the Clinger-Cohen Act, which created agency CIOs and capital planning by statute. When PCs, spreadsheets and later cloud outran governance, the fix was standing structure: IT Information Library (ITIL)-style change management and, federally, FedRAMP, which turned an unmanageable stream of vendor offerings into a permanent intake pipeline.
Different technologies, different eras, but the lesson is remarkably consistent: organizations captured value only after they built repeatable ways to turn new technology into changed work and accountability. Organizations captured value only after they built repeatable mechanisms for turning new technology into changed work and changed accountability. At AI’s pace, that mechanism has to be permanent.
This is already a mandate
Inside the executive branch, the question of whether an agency needs an enterprise AI strategy is settled. OMB’s M-25-21 requires every CFO Act agency to name a chief AI officer, convene an AI governance board, publish an AI strategy and compliance plan and maintain a public use-case inventory. It defines “high-impact” AI, where the output is a principal basis for decisions affecting rights, benefits, safety or access to services, and sets minimum practices for it: pre-deployment testing, impact assessments, ongoing monitoring, human oversight and the ability to shut a use down. The deadline to document those practices was April 3, subject to exceptions and waivers. The obligation is in effect now. M-25-22 applies parallel discipline to acquisition, including performance tracking and protection against vendor lock-in.
Notice where high-impact AI lands: benefits adjudication, eligibility, candidate screening, fraud detection and workforce analytics touching pay or promotion. Those are exactly the use cases where AI promises the most value. The Equal Employment Opportunity Commission (EEOC) and the Justice Department have already warned that algorithmic hiring tools remain subject to disability-discrimination protections. The places where the value is greatest are the places where the governance questions are sharpest. That’s why the business side belongs at the governance table.
It also reframes the internal conversation. When a program office asks its CIO about enterprise strategy, it isn’t asking for something extra. It’s offering to help meet an obligation the agency already carries.
What the business side can do now
The most productive thing a business unit can say to its CIO and CAIO is: “We’ll carry part of this.”
The business unit should bring an aggregated, quantified demand signal instead of scattered tool requests: use cases with volumes, cycle times and expected outcomes. That’s the business case the CIO needs for their own budget defense. Frame every ask around M-25-21 and M-25-22 rather than unit preferences. Volunteer as a disciplined first pilot customer, with a baseline already in hand, written success criteria and a real willingness to shut the pilot down. Ask for a seat on the governance board, not a delivery date the enterprise team can’t promise.
Then put one concrete proposal on the table: a standing AI change-management function, jointly owned by CIO/CAIO and the business, running four pipelines. A single register of every AI-relevant change entering the environment. A lightweight impact triage: which workflows it touches, whether it alters risk posture, whether it’s high-impact. A channel that translates release notes into role-specific guidance before the change lands. And vendor notification written into contracts, using the leverage M-25-22 gives acquisition teams. The point isn’t to create another governance layer. It is to create a repeatable way to see AI-related changes coming, assess their impact, and prepare the workforce before those changes land.
Here are eight steps to keep in mind as you accelerate your AI journey:
Readiness is business work
Almost everything that decides whether AI succeeds in your operation is knowledge your unit already holds. The work is making it explicit before the technology arrives.
Know where AI could help and where it shouldn’t. Every unit has a mental map of its pain points. Turn it into a shared, prioritized list. When the agency asks “Where should we start?”, the units with specifics shape the agenda. The rest get someone else’s answer.
Know your data. You don’t need to become data engineers. You need to be the people who can say, with confidence, which system tells the truth about what and where the trouble spots are. Nobody else in the agency can say it for you.
Describe your processes. You can’t automate what you can’t describe. Writing down how the work actually flows surfaces improvements worth making with or without AI, and it exposes which steps actually require judgment and which are simply rules we have historically relied on people to interpret.
Name your red lines. Some decisions touch people’s benefits, employment or rights. Say where a human must stay in the loop, and make that review real: a named reviewer, enough time and information to challenge the output, and the authority to reject it. A signature alone isn’t oversight.
Capture a baseline. If you don’t know today’s cycle times, volumes, error rates and backlogs, you will have a hard time proving next year’s improvement or defending the investment that produced it. Keep adoption numbers separate from mission results.
Bring your people along. If the workforce doesn’t trust the tool, adoption will stall regardless of how capable the technology is. Staff are watching AI arrive with curiosity and anxiety about accuracy, accountability and their own futures. Meeting that honestly is leadership work, not IT work.
Ask vendors better questions. What data trained this? How is bias tested, and can we see the results? What happens when the model is wrong? How will we be told when it changes? What rights do we keep in our data?
None of this needs a new tool, a new budget line or a technical team. It needs the decision to treat readiness as part of running the business, within the authority you already have.
“Not there yet” is an invitation
The first chapter of federal AI was experimentation. The next one is execution. The units that succeed will look the way the research keeps describing: governed data, clear objectives, baselines set before deployment. A business unit that does this work becomes the partner its CIO most needs, a disciplined first customer that proves value, surfaces risk honestly and leaves a template the rest of the agency can follow. If the enterprise isn’t there yet, the business side does not have to wait. Much of the work required to get ready is already within its control.
Don Bauer is a retired federal senior executive and human capital technology leader who most recently served as AI ambassador at Office of Personnel Management Human Resources Solutions. He co-founded QuickHire, later acquired by Monster, and has led major federal HR IT modernization efforts at Customs and Border Protection and the State Department. He consults, writes and speaks on AI and government transformation.
Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.