Marisol Tanaka-Boateng did not work in IT. She processed permit applications at a county building department, and she had never written a line of code in her life. So when her supervisor mentioned that the department was "adopting AI" and that every employee had a role to play in using it responsibly, Marisol assumed the message was meant for someone else: the data team, the contractors, the people who actually built things. It was not until a colleague pasted a resident's full application, including their home address and social security number, into a free public chatbot to "help write a denial letter faster" that Marisol understood: AI stewardship was not a job for specialists. It was a job for her, and for every person who touched a government system. The colleague meant no harm. That was exactly the point.

This lesson is about your role as an AI steward, the responsibility every government employee carries in an era when AI tools are everywhere and easy to use. You do not need to be technical. You do not need to understand how the models work. You need to understand what responsible use looks like, where the bright lines are, and why your individual choices matter more than any policy document. Stewardship is where understanding becomes action, and it is the connective tissue that holds the rest of a governance framework together.

What stewardship means

A steward is someone entrusted to care for something on behalf of others, something that does not belong to them. A park ranger stewards public land. A records clerk stewards public information. As a government employee in the age of AI, you steward two things at once: the public's data and the public's trust. Both can be damaged in seconds by a well-meaning person using a convenient tool without thinking, and neither can be restored on the same timescale.

The AI systems your agency uses do not belong to your agency. They belong to the public. They are built with taxpayer resources, deployed using government authority, and they affect citizens who have no choice in whether to use them. That last point is where the heightened duty comes from. A customer can take their business elsewhere. A resident applying for a permit, a benefit, or a license cannot. They are required to give you their information and they have no alternative provider. When Marisol's colleague pasted a resident's data into a public chatbot, the resident had no idea, no consent, and no escape. The duty to protect them fell entirely on the employee.

You can have perfect policies written in an office a thousand miles away. You can have brilliant technologists and careful compliance officers. But if the people actually working with AI systems day to day do not understand their role and their responsibility, the best-designed framework in the world will fail at the point where it meets a real person with a deadline. That is the weak link this lesson is trying to eliminate, and the reason stewardship is framed as a habit rather than a title.

Why the stakes are different in government

In the private sector, when an AI system makes a mistake or behaves unexpectedly, the company absorbs the loss. Perhaps a customer is inconvenienced. Perhaps a transaction is reversed. In government the stakes are categorically different, because when government AI fails, it fails on citizens. The loss does not land on a balance sheet. It lands on someone who was required to interact with you and who has no recourse other than the one you provided.

Consider four realistic scenarios. An AI system that prioritizes job training applications and accidentally deprioritizes applicants from underrepresented communities. A chatbot that handles benefit eligibility questions and systematically misguides people about their entitlements. A document processing system that loses important case files because it misclassified them. A predictive analytics tool that guides police resources to neighborhoods that have been historically over-policed. Each of these is not just a technology failure. It is a failure of governance, and ultimately a failure of stewardship.

Government AI adoption will always involve tradeoffs, and there is no such thing as a perfect AI system. There are only systems that have been thoughtfully designed, carefully deployed, and continuously monitored by people who understand the risks and care about the outcomes. Your role as a steward is to be part of that chain of responsibility. Not checking boxes on a compliance form, but actively asking whether the AI systems you work with are serving the public interest, and saying so out loud when the answer is unclear.

The steward mindset

Four characteristics distinguish someone who holds this responsibility from someone who merely operates the software. They are habits of attention rather than technical skills, which is why a permit clerk can have all four and a systems architect can have none.

You understand that AI systems are never neutral. Even if the algorithm is well designed and the data is clean and representative, the system reflects choices: choices about what to measure, what to optimize for, who wins and who loses. A steward recognises that those choices exist and asks whether they are the right ones, rather than treating the output as a fact about the world.

You stay curious and skeptical. You do not accept AI outputs as gospel. You ask how the system arrived at this answer, what data it used, whether it could be wrong, and what would happen if you relied on it. This is not obstruction. It is the ordinary diligence you would apply to a recommendation from a colleague whose track record you did not know.

You take your place in the chain of accountability seriously. You document decisions. You escalate concerns. You do not look the other way when something seems off. You communicate upward and laterally. Institutional transparency and honesty are what hold a government organisation together, and they are made of many small acts of writing things down.

You understand that stewardship is collective. You cannot do this alone. You need colleagues who share the same values, leaders who enforce them, and systems that make it easy to raise concerns without fear. A lone steward in an unsupportive organisation is a burnout waiting to happen, which is a problem to solve rather than a fate to accept.

The bright lines every employee must know

Stewardship begins with a small set of non-negotiable rules. You do not need to memorize a manual; you need to internalize these four.

Never put sensitive or personal information into a public AI tool. Free, public chatbots are not approved government systems. Anything you type into them may be stored, used to train future models, or exposed. Personally identifiable information, meaning names paired with addresses, social security numbers, case details, and health or financial data, must never go into an unapproved tool. This single rule would have prevented Marisol's colleague's mistake entirely.

Treat AI output as a draft, never as a decision. AI tools generate fluent, confident text that can be wrong. A model may invent a regulation, misstate a deadline, or fabricate a citation, a failure mode called a hallucination. As a steward, you are accountable for anything that goes out under your agency's name. AI can help you draft; it cannot decide, and it cannot be blamed. The human signs the letter.

Use only approved tools for government work. Your agency designates which AI systems are vetted for which kinds of data. If you do not know whether a tool is approved, the answer is to ask, not to assume. "I did not know" is not a defense when a resident's data ends up somewhere it should not be.

Disclose when AI was meaningfully involved. Transparency is part of trust. If an AI tool drafted a substantive communication or informed a decision affecting a resident, the public and your supervisors should be able to know that. Hidden automation is how trust erodes, and it erodes fastest when it is discovered by someone outside the agency.

The AI Stewardship Pledge

Many agencies formalise this commitment with a stewardship pledge. It is not a legal document. It is a professional commitment that embeds the steward mindset into your practice, and its five clauses each carry a specific meaning worth reading slowly rather than skimming.

"I commit to the responsible use and oversight of AI systems in government." Saying this makes a public commitment within your organisation that you will take the role seriously. You are signalling to colleagues that you believe responsible AI matters, and you are creating accountability for yourself in front of witnesses. "I will question AI outputs, examine underlying assumptions, and escalate concerns promptly." This is your permission to be skeptical and your licence to ask for explainability. It is also your responsibility to validate outputs rather than merely accept them, especially where those outputs will shape decisions that affect citizens.

"I will prioritize fairness, transparency, and the public interest above convenience or efficiency." This is the values statement, and it exists because of a predictable pressure. There will be moments when using AI as is is easier than building safeguards, and moments when accepting the answer gets the work done faster than questioning it. The pledge is your reminder that those are exactly the moments to slow down.

"I will advocate for inclusive governance, equitable data practices, and continuous improvement in AI systems." This commits you to being an advocate rather than a passive user. When you see gaps in how AI is governed, you speak up. When you notice that certain communities are underrepresented in training data, you flag it. When you see better practices at other agencies, you propose them.

"I will contribute to building a culture where responsible AI is everyone's responsibility." Finally, you commit to spreading the mindset. You mentor colleagues, share what you learn, normalise conversations about AI risks, and help make thoughtfulness about AI the default rather than the exception.

Departments often write their own short form of the same commitment. Marisol's department adopted one after the chatbot incident: a commitment to protect the personal information of the people they serve, to use only approved AI tools for government work, to treat AI output as a draft they are responsible for verifying, to be transparent about when and how AI informs their work, and to speak up when they see AI used in a way that could harm the public. The value of a pledge is not the signature. It is that it converts a vague organisational value into a personal commitment a specific person made, and the conversation around signing it changed behaviour in ways the incident report never did.

Practical steps for becoming an effective steward

Being a steward is not abstract. Six specific things are available to you starting today, and none of them requires a budget line or a new job title.

• Understand your specific role. Are you working with AI systems or building them? Are you responsible for policy, implementation, or oversight? Your agency probably has a governance structure. Find out where you fit in it, know your decision rights and your responsibilities, and if you are unclear, ask rather than infer.
• Learn about the specific systems in your sphere. If you work with an AI system, really understand it. Ask for technical documentation. Understand what data it uses, what it is optimizing for, and what its known limitations are. If the vendor or your technical team cannot explain it, that is a red flag rather than a sign that you are not clever enough.
• Build relationships across roles. Data engineers, compliance officers, ethicists, technical staff, and business analysts each hold a piece of the puzzle. Knowing these people and understanding what they care about makes it far easier to coordinate when a problem emerges, because you will not be starting from introductions in the middle of an incident.
• Establish validation practices in your own work. Before you act on an AI system's output, run a checklist. Does this pass basic sanity checks? Is the underlying data fresh and relevant? Have we tested this on edge cases? What would the impact be if we were wrong? These practices do not need to be elaborate. They need to be consistent.
• Escalate systematically and early. When you have a concern, whether technical, fairness-related, or about governance, escalate it. Know the escalation path before you need it, document the concern clearly, do not try to solve it alone, and do not wait until there is a crisis. Early concerns are cheap. Late ones are expensive and public.
• Participate actively in governance forums. If your agency has an AI steering committee, a review board, or governance meetings, attend them, participate, ask questions, and help shape decisions. These forums are where stewardship shows up at the organisational level rather than the individual one.

Three jobs, three stewardship moments

The eligibility determination officer. You work for a benefits agency and your team uses an AI system to screen eligibility applications. It has been in place for two years and seems to be working well, processing applications quickly with reasonable appeal rates. Then you notice something: applications from certain zip codes have notably higher rejection rates. The easy path is to assume the system works as designed, trust the algorithm, decide that your job is to process applications rather than question the system, and move on.

The steward path is more work and it is what the job actually is. You document what you noticed. You pull data on approval rates by zip code and find the pattern holds. You reach out to your supervisor and to the data team and ask whether this has been analysed before. You request a fairness audit looking specifically at whether zip code is acting as a proxy for protected class status. You ask about the historical data the model was trained on. You suggest a temporary increase in human review for borderline cases from those zip codes. You communicate findings to leadership and propose a monitoring dashboard that flags the metric monthly. That sequence catches a potential proxy discrimination problem before it becomes entrenched.

The policy analyst drafting guidance. You are asked to draft guidance on how your agency will use an emerging AI technology, and leadership wants it quickly. You could write something generic: "agencies should follow applicable law and best practices." That is a stewardship failure dressed as a deliverable, because it transfers every hard question to whoever reads it next. Instead, you research what other agencies have done, ask hard questions about what problem the AI actually solves, and request data on how the system performs across different demographic groups.

Then you make the guidance load-bearing. You propose specific approval authorities and review thresholds. You recommend monitoring metrics and escalation procedures. You build in a sunset clause requiring review after six months. You create templates for impact assessments. It is more work, and the result is a guardrail that keeps the technology aligned with the organisation's values instead of a paragraph that will be cited by everyone and constrain no one.

The IT staff member implementing a system. You are implementing an AI-powered internal operations system, perhaps for meeting scheduling, email prioritization, or resource allocation. The vendor supplies sample data to test with and the implementation team wants to move. As a steward you examine the sample data for diversity and representativeness, test the system against edge cases such as unusual schedules, non-standard requests, and minority languages, and document any failures or unexpected behaviours. You push back where you see potential issues rather than simply standing up what was requested. You create monitoring so you will know if performance changes over time, and you maintain a contact path for escalating issues discovered in production.

Building a culture of responsible AI use

Pledges are powerful symbolically, but culture change requires structure and sustained effort. Six practices do most of the work, and all six are available to someone with no authority over anybody.

• Lead by example. The most effective way to build culture is to model the behaviour you want. When you question AI outputs, you show colleagues it is acceptable. When you admit uncertainty or change your mind on new information, you normalise intellectual honesty.
• Make responsible AI visible. Do not let AI decisions happen behind closed doors. Surface them in team meetings and discuss them in writing where colleagues can see. When a decision was made thoughtfully, talk about it. Visibility makes responsibility tangible.
• Create forums for dialogue. Work with colleagues and leadership to create spaces where concerns can be raised without fear. A dedicated chat channel, a monthly AI book club, a quarterly lunch-and-learn, or informal coffee meetings all work. The format matters less than consistency and psychological safety.
• Document and share learning. When a system had a problem, or you caught a mistake before it happened, write it down and share it. Create organisational memory so the same problem is not rediscovered by someone else five years later.
• Celebrate responsible practices. When someone escalates a fairness concern, spends extra time validating results, or proposes better data governance, acknowledge it. Make it clear that doing the right thing is valued rather than merely tolerated.
• Connect AI to mission. Responsible AI is not about compliance. A health agency that is thoughtful about AI serves public health better. A benefits agency careful about fairness serves eligible populations better. Tie the practice to the mission and people understand why it matters.

Recognizing stewardship failures

Sometimes, despite your best efforts, stewardship breaks down. Five signs tell you it is happening, and each is visible from an ordinary desk without special access.

• Lack of transparency. Decisions about AI systems are made behind closed doors with no documentation and no communication to relevant stakeholders.
• Absence of risk management. Nobody can articulate what could go wrong with a system, what its failure modes are, or what the mitigations would be.
• Suppression of concerns. People who raise questions about AI systems are punished or sidelined rather than listened to.
• Equity blindness. There is no process for examining whether a system affects different populations differently, and no commitment to investigating fairness.
• Governance bypass. Systems are deployed without going through required review, or review boards are powerless to actually stop anything.

If you observe these signs, your responsibility is to act, in a specific order. First, document what you are seeing. Write it down, be specific, include dates, decisions, and people involved, because that creates an institutional record rather than a memory. Second, communicate within the appropriate channels, trying to resolve the issue at the lowest level first. Speak to your supervisor, and if that does not work, use your agency's escalation process. Most agencies have ethics hotlines, compliance offices, or Inspector General channels.

Third, offer solutions and not just criticism. It is easy to point out problems, and stewards go further by proposing how to fix them. Fourth, if internal channels fail, know your external options. If your agency is not responding to stewardship concerns appropriately, the Inspector General office, the Government Accountability Office, or Congressional committees might be the appropriate escalation path. This is rare. It is also the ultimate expression of stewardship, putting the public interest above organisational loyalty.

The hardest part: speaking up

The most demanding element of stewardship is the willingness to raise a concern in the moment. Marisol's defining moment was not signing a pledge; it was deciding what to do when she saw her colleague paste the resident's data into the chatbot. Saying something to a peer is uncomfortable. Saying nothing is how small mistakes become front-page failures, and the discomfort is the whole cost of the intervention.

Good stewardship cultures make speaking up easy and safe. That means knowing the path to report a concern, whether a supervisor, an AI governance contact, or an ethics or privacy office, and knowing that raising a good-faith concern is protected rather than punished. Marisol's department added a single line to its intranet: "Saw something about AI that worried you? Here is who to tell." Reports went up, which was the goal. A concern raised early is a near miss. A concern swallowed is a future incident.

Marisol's own first act of stewardship was quiet. She did not report her colleague punitively. She walked over, explained why the public chatbot was a problem, and showed him the approved drafting tool the department actually had. The resident's data was already exposed and that could not be undone, but the next resident's was not. That is what stewardship looks like in practice: not heroics, but a series of small, correct choices made by ordinary people who understood that the trust was theirs to keep.

How your stewardship connects to the frameworks

You are not operating in isolation, and it helps to know which formal structure each of your habits feeds. NIST AI RMF. The GOVERN function explicitly includes roles and responsibilities, so your stewardship practices should align with what it expects. When you ask questions about data you are supporting the Map function; when you escalate concerns about fairness you are contributing to Measure and Manage.

OMB M-24-10. This memorandum sets requirements for AI governance in federal agencies, and your practices help fulfil them. When you participate in impact assessments, help document AI use cases, or support monitoring, you are helping your agency meet its obligations under it rather than doing something extra alongside them.

Agency policy. Most agencies now have AI governance policies. Know what yours says. If they do not exist yet, advocate for creating them. If they exist but are not being followed, be the person who calls that out. The pledge connects you to this larger structure: you are not a lone voice for responsibility, you are part of a governance ecosystem that only works when people inside it actually use it.

Anti-patterns to watch for

• Stewardship theater. The organisation adopts the language and even runs the pledge ceremony, but behaviour does not change. Systems are still deployed without real scrutiny and concerns are still suppressed. It happens because adopting the symbols is easier than doing the work, and leadership can check the responsible-AI box without making hard choices. The result is a compliance veneer, people who take the pledge and are then pressured to ignore their doubts, and cynicism that undermines genuine efforts. Avoid it by being brutally honest about what stewardship requires: budget, slower timelines for proper governance, and hard conversations about whether particular AI uses are appropriate at all. If the organisation will not make those investments, that is important information. Push for them, and escalate if they are consistently refused.
• Signing as a substitute for doing. Treating a signed pledge as evidence that stewardship is in place. A signature records an intention on one day. It does not tell you whether anyone questioned an output, escalated a concern, or changed a deployment this quarter, and those are the only observations that would.
• Stewardship without support. Individuals try to do the right thing but escalations are ignored, concerns are dismissed, and people who ask tough questions do not get promoted. Swimming against the current burns people out fast. Your best stewards leave, others stop raising concerns, and the organisation loses its institutional knowledge of what might go wrong. Avoid it by building the support: start with a peer group of trusted colleagues, then documentation that makes the case, then concrete examples where stewardship prevented a problem.
• Stewardship as individual responsibility only. The commitment is framed as personal, but without systemic supports. You are responsible for catching problems while the organisation has no processes to help you catch them, because it is cheaper to expect self-policing than to build governance infrastructure. The result is reliance on individual heroics, problems slipping through when the person who would have caught them was busy, and good stewards blamed for not catching everything. Insist on systemic supports: governance structures, documentation requirements, monitoring tools, escalation processes, and training.
• Stewardship perfectionism. The drive for responsibility becomes so intense that nothing ships. Every system is blocked for more testing, more review, more documentation. Your agency falls behind, systems that would genuinely help citizens never get built, and people lose confidence in the process because everything is delayed. Stewardship is responsible risk management rather than perfection. You can deploy a system that has risks as long as those risks are understood, documented, and actively managed. The question is not "is this risk-free?" but "are these risks acceptable, and do we have a plan to monitor and manage them?"

Practice prompts

• Map your stewardship sphere. Take 10 minutes and map the AI systems relevant to your role. Which do you interact with daily or weekly? What decisions do they influence? Who do they affect? For each, identify one specific stewardship responsibility you hold: validation, escalation, monitoring, or feedback. That document becomes your personal stewardship checklist.
• Identify your escalation path. Do not wait for a crisis to work out how to raise a concern. Map your organisation's channels. Who is your supervisor? Is there an ethics hotline, a compliance office, a data governance committee? Know these before you need them. If they do not exist, that is valuable information about your organisation's maturity.
• Find your stewardship colleagues. Who else seems to care about responsible AI? Find them, have a coffee, share what you are thinking about. Build the peer network that will sustain you when stewardship gets hard, and that will back you up when you escalate.
• Take your first stewardship action. Name one concrete thing you can do this week. Ask a question in a meeting. Request documentation about a system. Propose a fairness check. Voice support for a colleague raising a concern. Choose something achievable, do it, then reflect on what happened and what you would do differently.
• Rehearse a stewardship crisis. You discover that a widely used AI system in your agency is systematically disadvantaging a particular population. You have documented it and raised it through appropriate channels. Leadership wants to keep the system running while they study the issue. What do you do? What does each step of escalation look like? What is your line in the sand?

Reflection

Take two minutes, and find a quiet space if you can. What does it mean to you, personally, to be a steward of government AI? Not in abstract terms, but in your specific role, with your specific responsibilities. What decisions will you face, what systems will you work with, and which populations will be affected? Who else needs to be a steward with you, and which relationships do you need to build to create the collective accountability that makes this sustainable? And what is your line in the sand: what would it take for you to escalate a concern to leadership, and what would it take to escalate beyond your organisation? Thinking about that now, in calm reflection, is what lets you act clearly when the moment arrives.

Glossary

• Stewardship. The responsibility of caring for something you do not own, on behalf of those who will use it. In government AI, taking responsibility for ensuring systems serve the public interest.
• Proxy discrimination. When a system does not directly discriminate on a protected characteristic but does so indirectly through a correlated variable, such as zip code standing in for race.
• Governance structure. The formal and informal mechanisms through which an organisation makes decisions, assigns responsibility, and enforces standards. For AI this includes committees, review boards, policies, and escalation processes.
• Impact assessment. A systematic evaluation of how an AI system might affect different populations and how the various risks should be managed.
• Escalation. The process of raising concerns to a higher level of authority when they cannot be resolved at the current level.
• Fairness audit. A systematic review of whether an AI system produces different outcomes for different demographic groups, and whether those differences are justified.
• Institutional memory. The collective knowledge an organisation has accumulated over time, stored in documentation, processes, and experienced staff.
• Hallucination. A failure mode in which a model produces fluent, confident output that is wrong, such as an invented regulation, a misstated deadline, or a fabricated citation.
• Personally identifiable information. Information that identifies a specific person, such as a name paired with an address, a social security number, case details, or health and financial data.

Related lessons

• PII and AI: The Bright Red Lines goes deeper into exactly what must never enter an unapproved tool and why.
• Your Agency's Approved AI Tools covers how tools get vetted, so "is this approved?" becomes a question you can answer yourself.
• Reporting AI Concerns in Your Agency covers the escalation paths named here in practical detail.
• Algorithmic Fairness in Government covers the proxy discrimination problem the eligibility officer above uncovered.
• NIST AI RMF: The GOVERN Function covers the roles and responsibilities your stewardship practices are meant to align with.

Closing

Stewardship is the commitment that ties everything else together. Fairness principles are meaningless unless someone is actually implementing them. Risks need someone to monitor them. Governance frameworks need stewards who will use them, ask the hard questions, and care about outcomes. Carry the mindset with you regardless of where your career goes next: it will make you more effective, help you identify problems before they become crises, and make you the kind of colleague who builds trust inside the organisation and outside it. Government is at a moment of institutional choice about how seriously to take governance and whether to actually listen to concerns. You are part of making that choice.

Key Takeaways

• Stewardship is everyone's job, not just IT's. Every employee who works with or around government AI systems stewards the public's data and the public's trust. You cannot opt out; you can only choose how seriously to take it.
• The people you serve did not opt in. Residents must give you their information and have no alternative provider, which is the source of the heightened duty.
• Never put PII into a public AI tool. Unapproved chatbots may store, train on, or expose anything you enter. This single rule prevents the most common and most damaging mistake.
• AI output is a draft you are accountable for. Models hallucinate confident, wrong answers. The human verifies and signs; the AI cannot be blamed.
• Use only approved tools, and ask when unsure. "I did not know" is not a defense when a resident's data is exposed.
• The pledge is a commitment to values, not a compliance checkbox. It converts an organisational value into a choice a specific person made, and a signature with no change in behaviour is stewardship theater.
• Culture change requires visible leadership at every level. You do not need to be an executive. Every time you question an output or escalate a concern, you model stewardship for colleagues.
• Stewardship is collective, not individual. You need colleagues who share your values, leaders who support you, and systemic supports that make doing the right thing easy. If those are missing, build them rather than absorbing the gap.
• Documentation and transparency are your tools. Writing down decisions and escalations creates institutional memory and accountability.
• Responsible AI means managing risk, not eliminating it. Perfect is the enemy of good; deploy with risks that are understood, documented, and actively managed.
• Speaking up is the hardest and most important habit. A concern raised early is a near miss; a concern swallowed is a future incident. Know the safe path before you need it, including the rare external one.

Frequently Asked Questions

I am not technical. Why is this my responsibility?

Because the failures that matter most rarely require technical skill to cause or to catch. Marisol's colleague did not need to understand a model to expose a resident's social security number, and Marisol did not need to understand one to stop it happening again. Stewardship is a set of habits: question outputs, use approved tools only, document what you notice, escalate early. None of that requires reading code.

My colleague did something risky with a public chatbot. Do I report them?

Start with the smallest intervention that fixes the problem. Marisol walked over, explained why the public chatbot was a problem, and showed him the approved tool the department already had. That is often enough, and it builds the culture rather than the fear. Where the exposure is serious or the behaviour continues, your agency's reporting path exists for exactly that, and a concern raised early is a near miss.

We ran a pledge ceremony. Does that mean we have stewardship?

Not on its own. Stewardship theater is the anti-pattern where an organisation adopts the language and the ceremony without changing behaviour, and it produces cynicism that makes real efforts harder. The evidence that stewardship exists is different: outputs questioned, concerns escalated and answered, deployments slowed for governance, budget spent. Look for those, and notice if none of them happened this quarter.

If I keep raising concerns, will I just block everything?

That is a real failure mode with a name: stewardship perfectionism. Every system carries risk, and demanding that all of it be eliminated means nothing ships and citizens do not get help they could have had. The right question is not "is this risk-free?" but "are these risks acceptable, and do we have a plan to monitor and manage them?"

What if I escalate internally and nothing happens?

Work the order. Document what you saw with dates, decisions, and people. Try the lowest level first, then your agency's formal escalation process, which usually includes ethics hotlines, compliance offices, or Inspector General channels. Offer solutions rather than only criticism. If internal channels genuinely fail, the Inspector General office, the Government Accountability Office, or Congressional committees might be the appropriate path. This is rare, and it is the ultimate expression of stewardship.