What Does Your AI Policy Want?

By Brett Horvath and Daphne Davis Moore

Here is a question we rarely ask of a governance document: what does it want?

Read your AI policy as a statement of what your institution wants, and the answer is usually plain. It wants to avoid a lawsuit or a bad headline. It wants sensitive information to stay where it belongs, staff to stay inside known boundaries, and the board to feel that someone is paying attention. Those are legitimate goals, and nearly all of them are defensive. What rarely appears is any account of what the organization wants to become able to do.

Our colleague Judy Friedman puts the root question plainly: do you want your AI policy to stifle innovation or encourage it? There is no sitting it out. A policy that says nothing about what the organization hopes to do has been answered anyway.

Try reading your own policy for potential first and for risk second. What would much cheaper translation mean for the communities you serve? Could community groups working on the same problem find each other and start working together? Could your own staff become the creators of the next generation of community experiences, building tools your community owns and shapes? Most policies were never asked these questions. A compliance document can protect an institution from some harms. A living AI policy also decides what the institution should become able to do. It is a short set of commitments your board approves, with procedures beneath them that your team audits, adapts and activates whenever a new tool, rule or risk arrives.

Notice, too, the first question people ask when a significant new capability appears. Some organizations start with "How do we keep this out?" Others also ask "Where might this matter?" A serious institution needs both, and the order and weight it gives them reveal what its policy actually serves. When compliance, risk and fear write all the rules, a policy becomes an infrastructure of no, and an infrastructure of no keeps an organization from building anything new.

Where most policies are now

Somewhere in your shared drive there is a document called something like "Responsible AI Use Policy: FINAL (Board Approved)." Maybe a task force wrote it. A consultant may have helped. People debated the hard words, the board approved it, and an all-staff email announced that the organization now had an AI policy.

Say your board approved it this spring. In July, the European Union changed its AI Act deadlines and began enforcing the act. California began requiring large AI providers to mark the images, video and audio their tools generate. Anthropic, Google, Meta and OpenAI all shipped new models in the first week of September, and Anthropic and OpenAI shipped cheaper ones three weeks later. Your policy has no opinion about any of this. It was finished.

If you have no policy at all, you have plenty of company. In September, NTEN, the nonprofit technology network, and The Bridgespan Group published a survey of 917 nonprofit staff and executives. Forty-five percent of respondents use AI daily or more. Among executives, 22 percent said staff training on safe and responsible use was in place, and 39 percent said written guidance was.

In October, one of us asked about 120 people at a gathering in Wichita, most of them from small community foundations, who had an AI policy of any kind. Seven raised their hands. None of the seven had updated theirs in the past 90 days. An organization without a policy still has one: whatever each staff member has worked out alone.

Starting is easier than it sounds. After the show of hands, each round table of eight to ten people played a twenty-minute scenario-planning game on one phone. One person at each table took the role of a foundation’s CEO and chose a response to each scenario while the rest of the table talked through the choice. By the end, each table had a first AI policy built from its choices and a set of questions for building it out.

The gap between policy and practice is usually ordinary. Someone uses a personal account because the approved tool cannot do the job. Someone else stops asking permission after a request disappears into review. A program team finds a valuable new capability, has no way to evaluate it, and files it as a security exception or lets it drop. Most of this is people making reasonable choices inside a system that no longer describes their work.

Why communications teams have the most at stake

At ComNet this year, AI was in every room. If you lead communications, you probably do not own your organization's AI policy — but it shapes your team's work more than almost anyone else's.

Communications teams carry much of what an organization hopes to do: reach more people, bring in new donors and volunteers, and get attention for grantees and partners. They are often the ones working hardest to try something new. They are also the most exposed when something goes wrong: a tool drafts something in the institution's voice that it cannot stand behind, a community member's story ends up somewhere it should not, or AI-made content goes out without a label. For a communications team, a policy works either as a set of chains or as a set of pathways, and the difference shows up in the work every day.

Your team also has information no one else has. It is often the heaviest AI user in the building, so its practice is the clearest picture of where policy and reality have separated. You already own the disclosure question in practice: what gets labeled, what the label says, and what the organization implies when it says nothing. And when the policy is done, the organization hands you the public language, so you inherit any gap between what the document says and what the organization can defend. That gives you standing to ask for a better policy, and the evidence for what it should say.

The cost of a defensive policy: the capture gap

We call it the capture gap: the distance between what AI could help an institution do and what that institution is equipped to evaluate and use responsibly. Most new products deserve to be turned down. The capture gap measures the possibilities that never get a serious hearing, because there is no legitimate way to consider them.

A communications team serving a Spanish-speaking community could go from translating three documents a year to translating everything. The only review process available is a security questionnaire built for procurement, and nobody owns the question of how good a translation has to be before a benefits notice goes out. A research team could review a decade of interview transcripts in a week. The conversation ends there, because the transcripts hold community members' stories and nobody has the authority to decide what may be done with them.

Nobody tracks these missed chances, so nobody counts what they cost: slower programs, administrative work a tool could do, services some people cannot use, and strategic options that never become visible enough to accept or reject.

AI policy is a strategy question that never sleeps

When an organization leaves open what it wants from AI, someone else fills the space: vendors through their prices and defaults, peers through what they adopt, and whichever staff member is under the most deadline pressure. Every software company you already pay is shipping AI into products you already use, each launch with a demo, a set of defaults and a renewal date.

Strategy, policy and technology always end up defining one another, and whichever goes first sets the terms. When the technology goes first, it quietly writes the strategy: the tool's defaults decide what AI is for, and the policy shrinks to checks applied afterward. When strategy and policy go first, they shape each other in the open. Strategy asks where saying yes to a new capability would matter for the mission. Policy sets how the organization can say yes, and what it protects while it does. A strategy without a policy is the latest shiny object; a policy without a strategy is a list of prohibitions.

When strategy and policy come before the technology, tools become interchangeable. Staff can use ChatGPT, Claude, Microsoft Copilot, the assistant built into your donor database or whatever ships next month, each under your rules, in accounts you control. A small team with a clear policy and strategy can often evaluate a new tool on Monday and put it to work by Friday.

The order also decides where the learning goes, and that may matter most. Every time staff use AI, the organization learns something: which instructions work, what a good translation review looks like, what went wrong and why. Kept with your people and in files you own, that knowledge compounds and travels with you to the next tool. Collected inside one vendor's product, it lives in that product's settings and history, and much of it stays behind when you leave. Decide in your policy, before you pick the tools, that what your people learn about AI stays with your organization.

So a living AI policy starts from a different goal: this organization will become more capable as AI changes, in service of its mission, without giving up its commitments or its judgment. It takes on new tools at a pace the organization can evaluate, and it keeps real choices in the organization's hands, so the organization makes them on purpose.

What makes a policy living

Organizations that get the goals right still reach first for a better document: more categories, a bigger risk matrix, a revised list of approved tools. Whose regular job is it to keep your policy current? Picture how AI news reaches your policy today. A few people see it, one of them thinks it matters and mentions it in a meeting, and whether it ever reaches the policy depends on who they happen to tell.

A living AI policy replaces that luck with a routine, built in two layers. The top layer is a page or two your board approves and rarely changes: what AI is for in your mission, whose trust you carry, what you will always protect, and who decides. Beneath it, your team keeps a working layer it rewrites as fast as the tools change: how staff try a new tool, who reviews it, and what gets reported. One named person watches for changes and brings them to the team. And the top layer stays short enough that a program officer can find it in a minute and act on it that afternoon, because a policy too heavy to follow protects no one.

Many organizations already run routines like this for investments, and some for security incidents. Few have one for AI. In the NTEN and Bridgespan survey, 30 percent of executives said a process to approve AI tools and vendors was in place, and 42 percent had named a person or group responsible for AI oversight. Each organization builds its routine from what only it knows: its mission, its data and its relationships. The routine itself has a trigger and three moves: something changes, and the team audits, adapts and activates.

The trigger: something new arrives or changes

Keep your scheduled reviews, but don't wait for them. A new tool becomes available, vendors change terms, staff habits shift, or a theoretical risk becomes real. None of it follows the board calendar. Ask: what could happen on Tuesday that would make us revisit this policy on Wednesday? If nobody can answer, the organization has a review schedule and no way to respond to change.

Give someone, or an AI assistant set to scan the news each week, the job of flagging changes that touch your policy, your tools and your data, along with new capabilities that could help a staff member or a team do better work and advance more of the mission this month, or even this week.

Audit: treat what people do as information

Audit what is true: which tools are in use, what staff are doing, what has changed since you wrote the policy. The distance between the document and daily practice is one of the most useful sources of information an organization has. Workarounds show where approved tools fall short; repeated exception requests show where the policy misreads a need. In the same survey, 78 percent of nonprofit staff said they learn about AI by using the tools themselves.

Some findings need a response. If someone pastes a community member's story, a grantee's unpublished financials or client identifying information into a consumer tool, treat it as a data incident: find out what was shared, where, and who needs to be told. Then ask why the approved path failed them, and fix that too.

Expect some experiments to fail. Keep a simple learning log: one line per use worth remembering, with what you tried, for which part of the mission, what happened, what you learned and whose trust it touched. Read it every month. Share what didn't work with your peers; it is often more useful to them than what did.

Adapt: make the mission case before the risk case

Adapt the policy where the document and reality have separated. An adaptation can loosen a rule as easily as tighten one. Start with what a capability could do for the mission. Then test that specific proposal against privacy, security, accuracy, labor and dependency. A risk review with nothing specific to test tends to follow whatever worries people most.

Take the translation example. The mission case: benefits notices and crisis messages reach people in their first language within hours, not months. The risk review then has specific questions to ask. Who bears the cost when a translation gets an eligibility rule wrong? Does a community that institutions have served badly want machine translation from you? The answer may be no. Either way, you know what you are protecting and what you are giving up.

Activate: put the change to work, inside your organization and beyond it

Activate the change so it reaches the decisions and relationships the policy governs. This is the step institutions skip. Decide who can propose a change, who can approve an urgent one between board meetings, how disagreements get settled, and how a decision travels through procurement, data, staffing, grantmaking, vendors and programs. The test is what your organization does differently because the policy exists.

If you fund other organizations, activation reaches people who had no part in writing your policy. It becomes language in grant agreements, reporting templates and due diligence. A disclosure rule that costs your communications team one approval step can cost a four-person grantee a week. Requiring grantees to certify how they used AI moves the cost of your uncertainty onto organizations with less capacity to absorb it, and produces statements nobody believes. Give grantees a way to tell you the policy is failing them, and act when they do. The field's most useful knowledge about AI is building up in the organizations doing the work, and a funder's policy decides whether that knowledge flows in or stays out.

Where to start

Begin with the policy you have, if you have one. Open it this month and check the date. Read it first as a strategy document: what does it want the organization to be able to do? Then read it as a protection document.

Bring five questions to your next leadership meeting:

  1. What could happen on Tuesday that would make us revisit this policy on Wednesday?

  2. What are our people already doing that this document does not describe?

  3. Could our staff find this policy today, and do they know what it says and means?

  4. When our people learn something about using AI well, where does that learning live, and would we keep it if we changed tools?

  5. What are we requiring of grantees and partners that we have not tested on ourselves?

Then decide who keeps the routine and the rhythms around it: who can change the procedures, how often you will check them, and who writes the first line in the learning log.

And rehearse one change this month. Pick one thing that changed in AI this week; a new feature in a tool you already pay for will do. Take it through audit, adapt and activate in thirty minutes with the people it would affect, and write down what you decided and what you learned. That half hour is a living AI policy in miniature.

Some of the references in this piece will be out of date within a quarter. That is the argument. Notice where your policy stops matching what your people do, what they have built around it, and which decisions have nowhere to go. Then tell a peer organization what you found. The field's best knowledge about AI is gathering in exactly those gaps, and every organization that shares what it finds makes the next one's decisions easier.

Every day is a new, strange test. We might as well be sharing our answers in real time.

About This Series

This is the third in a series examining what becomes possible when the social sector treats AI not as a tool to adopt but as a force already reshaping the landscape it operates in. The first article, "The Ground Is Shifting. Do More Than Take Notes," established the gap between adoption and transformation. The second article “The First Reader Isn’t Human” explores one dimension of that gap, the communications dimension, and what closing it looks like.

Brett Horvath is a strategist working at the intersection of AI, information warfare, and political and economic risk. He advises elected officials, policymakers, Fortune 100 leaders, and government teams on navigating engineered volatility and rapidly shifting information environments.

Daphne Davis Moore is a communications strategist who has shaped public narratives for more than three decades, guiding presidential campaigns, nonprofits, and Fortune 500 companies through pivotal moments of change. She is a trusted counsel to leaders who understand that the right message at the right moment can fundamentally alter the trajectory of a cause, a career, or an organization.

Next
Next

The Communications Network Welcomes New Board Leadership