From Policy to Practice: An AI Roadmap
AI implementation isn’t a big launch. It’s a series of careful steps that protect people and show clear public benefit. This post offers a roadmap that is practical and people-first for government and non-profit teams: start small, involve staff, unions or worker reps, and community early, set clear boundaries, and keep equity and human oversight at the center, so services improve.
Vision and values in action
Begin each initiative with a short mission-fit statement that explains the public benefit, who is likely to benefit, and how you will know. Share that statement with the people who will be affected inside the organization and, when the tool touches residents or clients, with the community early in the exploration phase. If the work is strictly administrative and behind the scenes, share internally and make it easy for staff to surface concerns. Name where AI is not appropriate, so your boundaries are clear. Examples include tools that issue evacuation or shelter-in-place directives without qualified staff review, tools that make personnel or disciplinary decisions without required human judgment, and unreviewed public-facing chat tools. Define when human review is required, who conducts it, and how people can appeal or escalate concerns. Commit to transparency and content authenticity, including notices when content is AI-assisted and brief “how we built this” summaries for public-facing tools. Keep equity in view from the start by naming likely benefits and risks for different groups and by stating how you will provide accessibility and translations.
People and readiness
Policies do not implement themselves; people implement them. Assign roles or named individuals rather than committees. You need an executive sponsor who clears roadblocks, an AI program lead who coordinates day to day, product owners accountable for specific use cases, and partners for data, security, privacy, procurement, communications, HR or worker representatives, labor relations, and community members, working as one team.
Provide short, role-specific training for non-technical staff so they understand what AI can and cannot do, how to spot issues, and how to escalate them. Build equity literacy into reviewer training so staff recognize biased outputs, accessibility gaps, and language access needs. Offer ongoing coaching and simple feedback loops with response times; close the loop so people see how input shaped the work. Invite community partners early to help narrow pilots to real problems that matter.
A living use-case and tool inventory
Maintain an organization-wide inventory that is easy to update. Organize it into three buckets: Approved, Pilots, and Under Consideration. For each entry, capture the program or service need it is tied to: the problem and goal, affected population, why AI fits now, data and sensitivity, risks and mitigations, human oversight, evaluation plan, and accessibility and language considerations. Record the decision and the rationale in plain language. Keep a decision log of ideas you chose not to pursue and why. This record strengthens transparency and helps prevent reinventing the wheel.
Practical oversight for tools and data
Before you buy or renew a tool, ask the vendor which model(s) it uses, how it was tuned (what adjustments and safety rules were added), how often it changes, what data it stores and for how long, and whether you can opt out of your data being used to train their systems. Confirm it runs in a private, closed environment and that your prompts, files, and outputs remain confidential; if not, do not use it with sensitive information.
Then ask for recent security audits, clear labels when AI helped produce content, and a clean exit path so you can take both your data and your configuration in open, machine-readable formats, not a vendor-only file. Request a plain language summary of safety settings, especially refusal rules, which make the AI decline risky or out-of-scope requests and explain why.
Inside your organization, write down who owns the data, who can see it, what can be shared outside (and what cannot), and how long you keep it before deletion, including backups. For public agencies, include public-records and retention duties; for non-profits, include donor or client confidentiality, funder reporting requirements, and any data-sharing agreements.
Match the pace of review to the level of risk: higher-risk tools get more frequent checks, spot reviews of outputs, and quick safety drills before expanding; lower-risk tools get a lighter, scheduled review. Keep a simple incident plan everyone knows: describe the issue, contain it, notify the right people, fix it, and record what changed, so you are ready when something goes wrong.
Evaluation and impact that matter
Measure what the public or your clients experience and what the workforce feels. Track service quality (accuracy, timeliness, clarity), equity (who benefits or is missed, for example equitable wait times across neighborhoods, impacts of the digital divide, and accessibility), workforce health (time saved, burnout, confidence), and operational outcomes (error and appeal rates, cost to serve). Use honest before-and-after comparisons and run short safety drills to see how the tool behaves in hard cases and to confirm your safeguards work before you expand.
Key terms: safety drills, in plain language
A red-team test is a “try to break it” session. You push the tool into tricky situations to see where it gives wrong, risky, or biased outputs before real people rely on it. The point is to learn its limits, improve prompts or safeguards, and confirm that human review, notices, and escalation paths work.
A failure-mode exercise (a “what-could-go-wrong” tabletop) is a quick walk-through of realistic mistakes the tool might make, who would be affected, how you would catch those issues, and what you would do next. It helps you decide whether to continue, adjust, or pause a pilot, and it assigns clear owners and timelines for fixes.
Equity and ethics checkpoints
Equity is not a principle on the wall; it is a series of checkpoints you pass before expanding scope. Trigger a checkpoint before any public-facing use, before expanding pilots into underserved areas or groups affected by the digital divide, before connecting to sensitive data, after staff or community concerns, during procurement, and again before renewal. At each checkpoint, verify that data represent the people served and limitations are documented; that language and accessibility plans are in place; that potential harms and mitigations have owners and timelines; that staff, HR, labor relations, and community input have been collected and considered; and that appeals and human-override paths work in practice. Define how monitoring will continue after approval, so accountability does not end on launch day.
Working with staff, HR, unions, and community
Partnership begins before a tool is chosen. Share the mission-fit statement, the no-go list, and live demos with staff and worker reps or HR and union leaders early. Be explicit about what AI will not do, such as replacing roles that require human judgment or changing job classifications without bargaining and legal review. With community, run early listening sessions to narrow the problem, show prototypes in accessible formats, support language access, and report back on what you heard and what you changed. For non-profits, brief your board or program committee early so oversight travels with the work.
Thoughts to carry forward
Start with mission fit and clear no-go uses. Keep equity visible at every turn. Name people to roles, train for practical judgment, and invite staff, unions, worker reps, and community in early. Build one living inventory and record not only what you pursue but what you decline. Match guardrails to risk and practice incident response before you need it. Measure outcomes that matter to the public, your clients, and the workforce, and share what you learn in plain language.
AI Disclosure: I knew what I wanted to write about for this post, starting with a detailed set of thoughts that I wanted to incorporate. I asked ChatGPT-5 Thinking to turn the thoughts into a draft post based on the tone and outcomes I was looking for, then GPT-5 Thinking and I worked through a few more iterations. I shared the new version with Karen and incorporated her feedback. Worked another draft with GPT-5 Thinking and made additional edits making sure it was true to my voice. GPT also helped me create the roadmap with compass icon.
Note: This post offers general guidance to help organizations plan AI work. It is not legal advice. Your approach should reflect your organization’s mission, values, applicable policies and laws, labor agreements, procurement rules, privacy and security standards, accessibility needs, and public records obligations. Please work with your counsel and internal teams to tailor what is right for your organization.