Building in the Open: a Whitelabel peer-learning paper

01The short version

Building in the Open01 · The short version

If you read nothing else

Six findings, from enough corners of the room to stop being anecdotes. Each becomes a chapter.

The spine

Comprehension, trust, the dissolving tools, the scoreboard, honesty, and the good-actor test.

The short version.

Six findings came up often enough, and from enough different corners of the room, that they stopped being anecdotes. They are the spine of this paper. Each becomes a chapter.

The first build is usually a misunderstanding. Before anyone ships anything, they have to settle an argument inside their own organization about what AI even is. Comprehension is the binding constraint. Capability rarely is.

Every project hits the same wall, and the wall is trust. Different missions, same two questions. Where does sensitive data live, and what does an AI say to a person who is not okay at 2 a.m. Several teams are solving the second question in parallel without knowing about each other.

The tools are dissolving. The cohort is drifting off packaged software toward bespoke, internally built tools, and discovering it needs a new kind of person to operate them. Close to the mission, fluent with the machine.

"Time saved" is the wrong scoreboard. Efficiency that backfills with more work misses the point. The real question is what the time is for. That question exposes what a leadership team actually values.

You build trust by being honest, not by selling. The organizations getting buy-in are admitting trade-offs out loud and putting the people they serve inside the design process.

The cohort treats "good actor" as a claim that has to be tested. Learning in the open, and lifting the organizations who were never in the room, is how the claim gets earned.

Building in the OpenContents

Contents

Each chapter makes one argument and hands to the next. Read straight through, or open at any chapter and still follow the thread.

How to read

Two minutes for §01. The rest is the evidence.

01The short versionSix things that came up too often to be anecdotes01

02Why this paper existsThe hard part, the method, and its limits03

03Who is in the roomThe cast, and why the sector labels stop mattering05

04The first build is a misunderstandingThe argument you have to win at home first10

05Everyone hits the same wall, and it's trustWhere data lives, and what the system says at 2am12

06The tools are dissolvingOff packaged software, toward the hybrid builder14

07“Time saved” is the wrong scoreboardThe cohort's most uncomfortable thread16

08Build it with them, not for themEarning the right to put a tool in front of someone17

09You cannot outsource your conscienceStanding somewhere while the ground moves19

10Where the work rhymesThe whole case for the network, on one page21

11Where the room still disagreesFive live arguments, recorded not resolved22

12What people are actually buildingSix project snapshots for the case-study series23

13Recommendations · four movesWhat the cohort asked for, as proposals26

14The 2026 cohort, and what happens nextProgress, the living draft, and room in the room28

15AppendicesGlossary, method notes, acknowledgments30

02Why this paper exists

Building in the Open02 · Why this paper exists

Why this exists

Trust is the bottleneck. Culture is. Knowing what the reclaimed time is for is. Almost no funder pays for any of them.

Convened by

Whitelabel, with the AI for Non-Profits Network — a weekly briefing reaching more than 4,000 nonprofit leaders across the US and Canada.

The hard part was never the technology.

Put a postpartum mental-health team, an LGBTQ+ youth charity, a public-health advocacy group, a Navy SEAL transition foundation, and a gender-equity organization in the same room. Ask them what is hard about building with AI. You expect different answers. You get the same five.

That is the finding underneath every other finding in this paper. Trust is the bottleneck. Culture is. Knowing what the reclaimed time is for is. Almost no funder pays for any of those problems directly, which is why a small group of organizations funded by the OpenAI Foundation's People-First AI Fund decided to stop solving them alone.

The People-First AI Fund, in brief

In December 2025 the OpenAI Foundation announced the first wave of its People-First AI Fund: $40.5 million in unrestricted grants to 208 community-based nonprofits across the United States, with a second wave of $9.5 million in Board-directed grants to follow. The organizations in this paper are a small, self-organized subset of that cohort. Participants often refer to it simply as their "OpenAI grant."

Source — OpenAI Foundation, "Announcing the initial People-First AI Fund grantees," 3 December 2025

The network meets every two months. Whitelabel convenes it alongside the AI for Non-Profits Network. Whitelabel sits in the room as a participant too: it builds a postpartum support app with Social Creatures under the same grant program. This paper is not a Whitelabel publication in the marketing sense. It is what comes out when a cohort agrees to compare notes honestly instead of competing quietly.

Building in the Open02 · Why this paper exists

The method

Conversations, not presentations. The format stays loose so the real obstacles surface in place of polished progress reports.

The limit

Three sessions, a self-selecting group. Read the findings as early, directional signals. The network will test them as it grows.

How we know what we know.

The sessions are conversations rather than presentations. Light introductions, a few short project pitches with time-boxed questions, then open discussion of whatever is hard right now. The convener keeps the format loose so the real obstacles surface in place of polished progress reports.

This paper synthesizes what was said. Quotations come from the recorded transcripts, lightly cleaned for readability without changing meaning. The paper marks the difference between a signal heard across the room and a single sharp observation. Three sessions with a self-selecting group cannot speak for the sector. Read the findings as early, directional signals. The network will test them as it grows.

A note on neutrality

The paper is written to be useful to a leader who has never heard of Whitelabel. Where a commercial tool is named, it is because a participant named it from experience. The naming is descriptive.

03Who is in the roomScroll to pan ↔

Building in the Open03 · Who is in the room

Look at the build, and the sector lines dissolve.The premise of the network

Four buckets

Health & mental health · education, workforce & mobility · economic resilience & access to capital · civic life & democracy.

Before the arguments, the cast.

US states across the group

Loose theme buckets

4,000+

Leaders reachable for dissemination

The cohort spans fifteen US states and four loose theme buckets: health and mental health; education, workforce and mobility; economic resilience and access to capital; civic life and democracy. Treat the buckets as labels of convenience. Most of the interesting projects sit across two or more.

What the buckets hide is the cross-sector overlap. A postpartum companion and a youth mental-health tool face the same guardrail problem. A veteran-services CRM migration and a years-deep content library face the same messy-data problem. A regional funder and a national advocacy body face the same culture problem. Look at the build, and the sector lines dissolve. That is the premise of the network.

Building in the Open03 · Who is in the room

On the call

A working session of the network — the room this paper describes.

AI compounds at the join

The organizations doing the most interesting work are often intermediaries: bodies that sit between others and broker trust.

In the room.

The OpenAI grant-recipients peer-learning network on a video call: Jordan Gilliam, Brian Wenke, Ralf Alwani, Nick Montoro, Tania Estrada and Jak Spencer, with a monday.com AI notetaker in the grid. — A working session of the network on the call — the room this paper describes.

AI compounds at the join

One pattern recurs across the cohort and should be named early. The organizations doing the most interesting work are often intermediaries: bodies that sit between others and broker trust. The Forbes Funds bridges funders and nonprofits. The AI for Non-Profits Network bridges the sector and the technology. Whitelabel bridges nonprofits and the build. AI seems to compound fastest at exactly that joint, where one trusted actor can hand a person or a resource to the next.

Building in the Open03 · Who is in the room

Who presented

Part one. Sessions 1 and 2 — not the full cohort.

To verify

Names and affiliations require verification before publication.

Who presented — part one

Organization	Represented by	Field	What they described building
Whitelabel	Ralf Alwani; Jak Spencer	Convener / build partner	Convenes the network; builds the postpartum app with Social Creatures; coordinates the AI for Non-Profits Network.
Social Creatures	Rose Perry (Founder/ED)	Social health	“Bonded by Baby” postpartum companion; AI donation and CRM widget on the website.
It Gets Better	Brian Wenke	LGBTQ+ youth, mental health	Conversational layer over the self-guided “imi” web tool; back-office efficiencies.
The Forbes Funds	Jordan Gilliam	Intermediary funder	Networked nonprofit AI agents (with CMU); open-source civic tooling; platform-agnostic evaluation.
Seal Future Foundation	Nick Montoro	Veteran transition	Salesforce to Monday.com migration; agentic workflows; internal community app; AI grant writing.

Building in the Open03 · Who is in the room

Who presented

Part two. Session 2 — May 2026.

Dissemination

Weekly roundup to 4,000+ sector leaders; case-study spotlighting.

Who presented — part two

Organization	Represented by	Field	What they described building
The Women's Building	Tania Estrada (ED); Todd Haas; Diane Santana	Gender equity	Internal education and a staff committee; development use; client-data sensitivity.
Public Health Advocates	Nicole McMorris	Policy and advocacy	State and local policy advocacy; youth wellness programs.
Only7Seconds	Melissa Hosten	Youth social health & connection	Mapping the AI ecosystem around teens — their pain points and highlights with AI — to build a youth-agency response to social health, connection, and digital wellness.
Veterans for All Voters	Jamison Aweau	Civic life & democracy	Nonpartisan veteran-led advocacy for a fairer political system.
SALDEF	Kiran Kaur Gill (ED)	Advocacy / youth	Sikh American advocacy; national youth leadership; digital-literacy uplift.
Springboard to Opportunities	Sarah Strip; Paheadra Robinson	Economic resilience	Centering low-income Southern families through a racial and class lens.
AI for Non-Profits Network	Jak Spencer	Dissemination channel	Weekly roundup to 4,000+ sector leaders; case-study spotlighting.

Building in the Open03 · Who is in the room

Who presented

Part three. Session 3 — July 2026, the first new arrivals from the first cohort.

The range widens

Seniors and refugee communities join youth and families, and the map stretches from upstate New York to the Texas border.

Who presented — part three

Organization	Represented by	Field	What they described building
Midtown Utica Community Center	Kath Stam; Kay Klo (ED)	Refugee-serving community center	An academia–nonprofit bridge with SUNY Poly's AI center: staff-first training and faster lesson materials for a six-person team.
City Heights Community Development Corp.	Angel Stancer	Affordable housing & economic opportunity	A community AI lab in San Diego: a trusted space to practice AI, aligned to the Department of Labor's AI-literacy framework; staff train first.
VIDA — Valley Initiative for Development & Advancement	Honesty Martinez	Workforce development	Digital-skills training and student support so low-income adult learners in South Texas finish credentials and enter high-demand occupations.
The Boston Project Ministries	Maridena Rojas	Community development & civic life	AI literacy for Dorchester's seniors, and “On the Porch” with Boston University — an app telling a different story of the neighborhood.

Each arrival was tested against the paper the way the network intends to keep testing itself: the same walls, met in new places. Staff-first capacity building, a trusted room to learn in, and communities involved before anything ships — the findings held, in four settings the first sessions never saw.

04The first build is a misunderstandingScroll to pan ↔

Building in the Open04 · The first build is a misunderstanding

Comprehension is the binding constraint. Capability rarely is.Start here — every other problem waits behind this one

The argument at home

Someone expects a chatbot, the team built a prediction model, and a board member asks where the actual AI is.

The first build is usually a misunderstanding.

Start here, because every other problem waits behind this one.

Before most of these organizations built anything, they had to win an argument at home. The argument was about what AI even is, before budget came up at all. Comprehension is the binding constraint. People arrive from different starting points, and the gap shows up first inside their own walls. Someone expects a chatbot, the team has built a prediction model, and a board member asks where the actual AI is.

For some, the first real piece of work was not a product. It was resetting expectations. The Women's Building came in expecting to deploy an agent and walked all the way back to basic education.

We thought we were going to be just adding an agent, and then everyone was like, what is AI? So we had to backtrack and start more like the education component.
Tania Estrada · The Women's Building

The adoption gap, in numbers

By early 2025 about a third of US adults (34%) had ever used ChatGPT — double the 2023 share, but two in three still had not. Session 3 found the number hard to believe in both directions: low against the pace of the field, and low against what some organizations see on the ground, where the families they serve use the tools daily. The gap is comprehension, not availability.

Source — Pew Research Center, "34% of U.S. adults have used ChatGPT," June 2025

Building in the Open04 · The first build is a misunderstanding

Two moves dissolve the fear: transparency, and a reframe.Hold two truths at once

The tell

The cohort reasons by analogy — the internet, the PC, the BlackBerry, social media. They've decided this changes things; now they're working out how to stand on it.

Underneath the comprehension gap sits a fear: that the tool is here to replace the people. The most experienced voices in the cohort were clear that you do not argue someone out of that fear. The move that works is to take time to listen rather than barrel through with "you don't understand," and to hold two truths at once. The concerns are valid. The tool can also help the people the organization serves.

Two moves dissolve the fear in practice. The first is transparency. Leaders who were open with staff and board found that people read the tool as help rather than threat. The second is a reframe that moves the risk off the technology and onto the choice in front of each person.

You only need to be fearful if you're resistant to learning. There's going to be a new baseline. If you don't use the tool, you're kind of self-selecting out.
Nick Montoro · Seal Future Foundation

One smaller pattern runs through the cohort without anyone naming it directly. They reason by analogy. The internet, the personal computer, the BlackBerry, social media. The cohort has stopped asking whether this changes things. They have decided it does, and they are reaching for the last time the ground moved to work out how to stand on it.

05Everyone hits the same wall, and it's trustScroll to pan ↔

Building in the Open05 · Everyone hits the same wall

This is where the sector labels stop mattering completely.Trust is the universal friction

Where the data lives

For this cohort, trust infrastructure is the product requirement, and the platforms that provide it win the work.

Everyone hits the same wall, and the wall is trust.

Picture a new parent awake at two in the morning with a question she cannot ask anyone. One of these apps is being built for exactly that moment. Understanding was the first friction. Trust is the universal one. It came up in every project, in two distinct shapes: where the data physically lives, and what the system says to a person who is not okay.

Where the data lives

This stuff goes to a server somewhere. What is the architecture we need to create a hybrid environment where information lives on actual servers we can physically access, to protect privacy in a way that doesn't just lean on privacy notices and 23andMe-style people?
Jordan Gilliam · The Forbes Funds

For anyone handling regulated or sensitive data, compliance is a gate. Features come after. The postpartum app is built HIPAA-compliant so it can sit inside clinical settings. The veteran foundation treats some of its data as a national-security matter and keeps the most sensitive processing local. Several participants said enterprise platforms extending HIPAA-grade compliance to their newer agentic features changed what they felt able to adopt. One named Monday.com specifically.

Building in the Open05 · Everyone hits the same wall

Same problem, two organizations, no shared toolkit yet.The single most actionable gap in the paper

Machines policing machines

Running several agents from different models so they cross-check one another and cut hallucination. Peer review, but between models.

What the system says

The harder half is conversational. This is where the parallel building is most striking. Multiple teams, who did not coordinate, are each working out how a chatbot should decline a question it must not answer, and redirect without going cold, for users who are already vulnerable. A flat "I can't answer that, just go here" can feel cold to a community where people already feel exposed.

It Gets Better grounds its assistant in fifteen years of vetted content and draws one hard line: a chatbot must never be taken for a therapist. The postpartum project builds visible guardrails around issues like suicidal ideation and routes them into trusted human resources. Same problem, two organizations, no shared toolkit yet. That gap is the single most actionable thing the network could close in the next six months.

We especially want young people to not think a chatbot is a therapist. We want to avoid that at all costs. How do we offer as much support as possible without going too far?
Justin Tindall · It Gets Better

As you start getting more of these agents, especially from different models, they quality-check one another. It's greatly reducing the amount of hallucination.
Nick Montoro · Seal Future Foundation

06The tools are dissolving

Building in the Open06 · The tools are dissolving

A new role: close to the mission, fluent with the machine.The hybrid / forward-deployed builder

Watch what they're leaving

Whitelabel stopped using Notion this spring. It started building its own pipeline tools instead.

The tools are dissolving, and a new role is appearing.

A thesis ran through both sessions, and here is the blunt version of it. Packaged software is starting to look optional. The convener named the cost in the same breath.

I haven't touched Notion in four or five weeks because we got frustrated, and now we're building our own pipeline tools. Most of our stuff is internal tools. The guys hate it because they wake up and, oh no, it's a bloody new tool.
Ralf Alwani · Whitelabel

The lived experience backs the thesis. Workflows that failed a few months ago now work, faster and more accurately, as the models improve. The room did not oversell it. Anything going to production still needs an engineering sanity check, and a harder question arrives the moment the barrier to building drops.

The idea that we're all becoming builders. The con side is, okay, now we're all building loads of stuff. It's actually how do we choose the right thing to build, and is it solving a problem that we had?
Jak Spencer · Whitelabel / AI for Non-Profits Network

Building in the Open06 · The tools are dissolving

The move that matters

Stop solving the same nonprofit problem twelve times. Open-source the answers.

A free on-ramp

The flood of openly published workflows and repositories is a free on-ramp for non-technical staff. Wiring data into a website is one of the quickest early wins.

That question creates a job. Several leaders described needing someone who sits between mission and machine: close enough to programs and operations to know what matters, fluent enough with agentic tools to build it. It Gets Better, having lost a deputy director, was redrawing the operations role in exactly those terms.

Maybe instead of a traditional operations director, I look for someone who understands our systems and can create these efficiencies for us.
Brian Wenke · It Gets Better

The cohort believes this capability is more reachable than it looks. Practitioners pointed at the flood of openly published workflows and repositories as a free on-ramp for non-technical staff, and at wiring data into a website as one of the quickest early wins. Which sets up the move that matters most for the network. Stop solving the same nonprofit problem twelve times. Open-source the answers.

Session 3 added a counterweight: the organizations doing this well are not hunting for one hybrid builder, they are building capacity across whole teams first. The Women's Building runs tiered staff training, a foundations session for the skeptical and individual coaching for the advanced; City Heights and Midtown Utica both put staff through the tools before anything ships to the community. People first, tools second.

If we're getting to the point where we can build our own things, we can share that open-source code. This is how I dealt with this very specific nonprofit problem, and it might solve something for all of us to build on top of.
Ralf Alwani · Whitelabel

07“Time saved” is the wrong scoreboard

Building in the Open07 · “Time saved” is the wrong scoreboard

The reframe

Not “AI gives you your time back.” “AI forces a choice about what your time is for.”

The edge

The technology removes the excuse for absent leadership. It does not produce humane leadership on its own.

"Time saved" is the wrong scoreboard.

Ask a funder what AI does for a nonprofit and you get two words. Time saved. The cohort thinks those two words are a trap. Once everyone can build, the differentiator is judgment. Access is the easy half. Judgment gets tested first on the question every efficiency pitch dodges. Time saved for what? Backfill the reclaimed hours with more applications and more output, and you have optimized your way back to the start.

We're getting away from the conversation around time saved. It's really a question of what will you do with the time given back to you. You would hope that would let nonprofit leaders be more present in community.
Jordan Gilliam · The Forbes Funds

There is a sharper edge here that the room kept circling. The efficiency tools are also inventing categories of work that did not exist a year ago. Newsletter triage, agent supervision, model evaluation, prompt maintenance. The pace of change itself is a tax on small teams. The honest version of the value story is closer to "AI forces a choice about what your time is for," and that choice quietly reveals what a leadership team actually cares about. If the test is care, the proof is in who an organization builds for, and whether it lets them anywhere near the build.

By Session 3 the cohort was circling a shared fix: a common language for outputs, outcomes, and the measures that sit behind them — a framework a nonprofit can put in front of a board or a funder to say what AI is for, beyond hours saved. It is a leading candidate for the network's next shared artifact.

08Build it with them, not for themScroll to pan ↔

Building in the Open08 · Build it with them, not for them

You ask the teenager first.Earning the right to put a tool in front of someone

The most durable strategy

Honesty about trade-offs. Advocacy for the technology came second, or not at all.

Build it with them, not for them.

You do not put an AI in front of a vulnerable teenager and hope. You ask the teenager first. It Gets Better runs a Youth Voices ambassador program and asked it directly how young people feel about AI, including the uncomfortable framings of it as a cheating aid and a replacement for human thinking. The postpartum project ran a co-design process with around thirty mothers inside a clinical program to reach a working beta, with a path toward a thousand parents.

The most durable adoption strategy in the room was honesty about the trade-offs. People tend to soften once they hear someone is not trying to sell them on AI, and is being human about a morally complex thing. Advocacy for the technology came second, or not at all.

A complementary stance worried less about fixing the technology landscape and more about equipping individuals to move through it safely. Brian Wenke at It Gets Better described it as a deliberate refusal to spend the organization's limited capacity trying to change the platforms. The job is changing how a young person perceives and navigates them.

By Session 3, co-design had also stopped being only a youth story. Public Health Advocates' youth researchers are shaping policy recommendations for their school board; Springboard's families set the agenda from baseline focus groups; The Boston Project Ministries starts with Dorchester's seniors. Across ages and communities the sequence is the same: ask first, build second.

Building in the Open08 · Build it with them, not for them

Somebody has to make sure the tide lifts the boats furthest from shore.The equity question, stated on its own

The working answer

Progression over perfection, paired with a slow, deliberate roll — so real people are involved early without being exposed to a half-finished system.

This is also where the equity question lives, and it deserves to be stated on its own rather than folded into co-design. The same communities most likely to be left behind by this wave are the ones several of these organizations exist to serve. The concern is concrete. An AI that does not sound like the people using it. Bias in its defaults. Assumed digital literacy that a first-generation family may not have been handed.

AI doesn't talk like a Latina. There's bias and representation, the origin of it. So we're changing that mindset first, and then leading toward language justice for the people we work with.
Tania Estrada · The Women's Building

Springboard to Opportunities frames its entire project around centering low-income Southern families through a racial and class lens. SALDEF is closing a digital-literacy gap for first-generation youth. Read together, these are one shared insistence. If the tide is going to rise, somebody has to make sure it lifts the boats furthest from shore. The working answer is progression over perfection, paired with a slow, deliberate roll, so real people are involved early without being exposed to a half-finished system.

09You cannot outsource your conscience

Building in the Open09 · You cannot outsource your conscience

Ethical posture cannot be inherited from a vendor.The ground keeps moving; the cohort stands somewhere anyway

The reality

You choose a platform on a Tuesday for good reasons. By Friday it has been acquired, signed a defense contract, or struck a deal you would never have signed.

You cannot outsource your conscience to a platform.

That is the ground these organizations build on. Platform allegiances move faster than a mission-driven body can re-architect its systems. One participant caught both the entanglement and the futility of pretending the platforms are cleanly separable.

Every one of these systems is connected. Even the ones that are separate are still built off APIs that are connected. What does that mean for how we're thinking about our ethics and our care?
Jordan Gilliam · The Forbes Funds

Two coping stances emerged in the room. They do not contradict each other. One is pragmatic acceptance: the technology will exist regardless, the more good actors engage with it the better, and a useful tool can be turned to good ends even when its maker's every decision is not endorsed. The other widens the lens past AI to a plainer truth. Few of the platforms a youth charity already depends on were built in its users' interest. The job was always navigation. Either way the conclusion is the same. Ethical posture has to be reasoned and communicated. It cannot be inherited from a vendor.

The practical corollary, named plainly in Session 3: living with one vendor is a dangerous space. The cohort's larger builders now deliberately spread their work across model providers and clouds, so that no single platform's Tuesday-to-Friday turn can take the mission down with it.

Building in the Open09 · You cannot outsource your conscience

An identity, not a slogan. A claim to be tested.What actually holds the cohort together

The job itself

Learning in the open, and lifting the organizations who were never in the room.

Which brings the paper to the thing that actually holds this cohort together. Not a sector. Not a tool. Not a grant. An identity. These organizations see themselves as the good actors in the space, and they treat that as a claim that has to keep being tested.

There's a responsibility that a lot of us who consider ourselves good actors, let's test that. Let's see, are we good actors? We have a moment here to prove that.
Jordan Gilliam · The Forbes Funds

Nick Montoro keeps the safety mechanisms at the front of the work as the thing that separates this group from people shipping carelessly. Others say they joined the space precisely because it needed more good players in it. That self-conception is the reason a network like this exists. If being a good actor is a standard the cohort has to keep earning, then learning in the open, and lifting the organizations who were never in the room, is the job itself.

There was a visibility of us pushing an AI narrative, and it did unintended harm to our brand and to people's perceptions. People are circling back now. It takes grace, and it takes human-to-human connection.
Jordan Gilliam · The Forbes Funds

Being early has a price, even for a trusted institution. The cohort's answer is not to advocate harder. It is grace toward the skeptics, and resets built on human relationships rather than another training.

10Where the work rhymes

Building in the Open10 · Where the work rhymes

The case for the network

Each row is a candidate for a shared tool, template, or working session — which is the same as saying each row is a reason this group should keep meeting.

The argument

The argument for a peer network rests on overlap: different projects solving the same underlying problem in parallel.

Where the work rhymes.

Shared problem	Who is hitting it	What the cohort could make of it
A conversational layer over trusted existing content	Bonded by Baby; imi (It Gets Better); Forbes Funds agents; Social Creatures widget	Shared patterns for grounding an assistant in vetted content (RAG); a reusable evaluation approach.
Guardrails and graceful escalation for vulnerable users	It Gets Better; Bonded by Baby; The Women's Building	A common redirection-message library, referral lists, and high-risk topic guardrail prompts.
Large, unstructured data needing cleanup	Seal Future Foundation (CRM migration)	Shared methods for AI-assisted ingestion and cleaning, and an honest account of where it still doesn't save time.
Internal efficiency vs. client-facing use	The Women's Building; Seal Future; most participants	A simple decision framework for which use cases need which level of care, review, and compliance.
Culture change and fear of replacement	The Women's Building; It Gets Better; Seal Future	A staff-engagement playbook built on the “new baseline” framing.
Equity, bias and who gets left behind	The Women's Building; Springboard; SALDEF; Only7Seconds	Shared practice on representative design, bias checks, and reaching low-literacy and non-English-first communities — with an eye toward communities not yet in the conversation.
Revenue and sustainability models	Bonded by Baby; The Forbes Funds	Comparative notes on earned-revenue models that route value back into mission.
Youth and community co-design	It Gets Better; Only7Seconds; SALDEF; Bonded by Baby	A shared co-design protocol for putting end users into governance and design.
Bridging people between organizations	The Forbes Funds; the network itself	A model for inter-organization referral, where one trusted agent passes a person to the next.

11Where the room still disagrees

Building in the Open11 · Where the room still disagrees

Recorded, not resolved

If a peer network is honest, it should report the things it has not yet talked itself into agreeing on.

The point

The cohort treats the absence of resolution as data. The next sessions test which arguments compress and which are permanent.

Where the room still disagrees.

Four live disagreements ran through the first two sessions. Each one will shape what the cohort builds next.

Speed of rollout. Progression over perfection (ship to ten real users, iterate, accept some discomfort) versus a slower, education-first approach that refuses to put a half-finished tool in front of someone vulnerable. The working answer was "fast on internal tools, slow on client-facing ones," but nobody believes that line holds once agentic features mature.

Platform exclusivity. The funder relationship pulls toward consolidating on OpenAI. The build experience pulls toward multi-vendor for resilience, cost, and cross-checking between models. The cohort has not decided whether platform-agnostic is a discipline, a tax, or both.

Whether bias is fixable at the prompt layer. "AI doesn't talk like a Latina" makes a structural claim about training data. Others are confident careful system prompts, RAG grounding, and tone calibration close most of the gap. The room is split on whether the remaining distance is a tuning problem or a generation problem.

What "open-source the answers" obligates. Everyone agreed in principle to share what it builds. Nobody agreed where: GitHub, a network-hosted repo, the funder's site, or organization-specific publication. The legal and brand-protection questions sit underneath, unresolved.

12What people are actually buildingScroll to pan ↔

Building in the Open12 · What people are actually building

Project snapshots

Six project snapshots for the case-study series.

Bonded by Baby

The market over-indexes on the baby and under-serves the parent's mental health.

What people are actually building.

Bonded by Baby — Social Creatures with Whitelabel

Postpartum mental health · HIPAA-compliant · NYC, scaling national. A digital companion extending a clinically-trialled postpartum curriculum beyond the limits of an in-person social-work team. Pairs short- and long-form trusted content in a familiar feed-style interface with a conversational companion and personalized support cards, on a RAG platform with visible guardrails for high-risk issues and a themes-and-insights panel for the clinical team. Co-designed with around thirty mothers. As of July 2026 the app is beta-ready and in the App Store, live with forty parents through a Mount Sinai program and expanding toward two hundred.

imi — It Gets Better

LGBTQ+ youth mental health · free, web-based · fifteen years of vetted content. A self-guided mental-health web tool, with evidence that longer engagement improves how young people feel. The grant work adds a conversational layer to guide users and deepen engagement, grounded in the organization's content library, with a firm boundary against being mistaken for therapy. Governance is participatory: Youth Voices ambassadors advise on design and on how young people actually feel about AI. Deliberately free and browser-based to remove barriers to access and allow easy exit.

Building in the Open12 · What people are actually building

Project snapshots

An intermediary funder networking agents across a region, and youth researching AI's effect on their own mental health.

The Forbes Funds

Agents are networked — around 700 by mid-2026 — so a resident can be handed between organizations in-chat.

Networked nonprofit agents — The Forbes Funds, with CMU

Intermediary funder · open-source · Pittsburgh / Appalachian region. A platform where an organization can stand up an AI agent from its public website, with for-profit usage subsidizing nonprofit access. Agents are networked, around 700 by mid-2026, so a resident seeking help can be handed between organizations in-chat, compressing a multi-day referral search into minutes. Includes an insights view for board reports and grant writing. The Forbes Funds positions itself as platform-agnostic, evaluating tools built on OpenAI, Anthropic and others, and is building open-source civic tooling for release on GitHub.

Youth researching AI and mental health — Public Health Advocates

Policy and advocacy · San Joaquin County, CA · youth-led. Twenty-eight youth, ages 13 to 19 and primarily BIPOC, working as research leads, facilitators, and peer mentors on one question: how AI strengthens rather than undermines mental health. Monthly AI-foundations sessions pair with wellness programming, peer focus groups follow in the fall, and the year closes with a youth-led symposium and policy recommendations to their school board.

Building in the Open12 · What people are actually building

Project snapshots

Bringing capability in-house, and an education-led, people-first approach.

The Women's Building

Progression over perfection — but a slow, deliberate roll before any client exposure.

Operations modernization — Seal Future Foundation

Veteran transition · data-sensitive · agentic workflows. Using the grant to bring capability in-house: AI-assisted grant research and writing; a Salesforce-to-Monday.com migration of tens of thousands of records, owned by one person rather than outsourced; and an internal community app built from white-labelled data. Notable signals include enterprise HIPAA-grade compliance extended to agentic features, and experiments with multi-agent cross-checking to cut hallucination. Sensitive data is handled locally by design.

Internal capability and education — The Women's Building

Gender equity · San Francisco Mission District · since the 1970s. A deliberately people-first, education-led approach. A standing staff committee meets regularly to build literacy and agree where AI is appropriate, distinguishing admin use from client-facing use, with explicit attention to client-data sensitivity. Early focus is development and fundraising. Live staff concerns include impact on critical thinking and creativity, bias and representation, and environmental impact, raised alongside the organization's move to solar. Next phase is hands-on training and language-justice considerations before any client exposure.

13Recommendations · four moves

Building in the Open13 · Recommendations · four moves

What the cohort asked for

Four moves, framed as proposals.

First deliverable

We have helped build ‘Guardian’, built by Whitelabel — now seeking organizations to test and build with.

Four moves.

Build the shared safety and escalation resource first. The artifact with the broadest demand and highest stakes: redirection-message patterns, a vetted referral list, guardrail prompts for high-risk topics, and a privacy-policy template small teams can adapt. It Gets Better, Bonded by Baby, and The Women's Building each solve a version in parallel. This is the network's first concrete deliverable.

Move the value conversation off “time saved.” Retire "hours saved" as the headline metric. A better question is "hours redeployed," and the discipline is to say in advance what they will be redeployed to. This lands on funders as much as grantees: a funder who asks "what is the time for" gets better answers than one who asks "how much faster." Treat it as evaluation reform.

Building in the Open13 · Recommendations · four moves

Learning in the open

Lightweight intake before each session, a between-session channel for handoffs, and a case-study series at three, six and nine months.

The collective ask

Continued convening, university evaluation, shared open-source tooling, dissemination through the AI for Non-Profits Network.

Stand up shared infrastructure for learning in the open. Three pieces would let the cohort actually publish what works and what doesn't. A lightweight intake before each session (stage, tools, top friction, one thing to share, one thing needed); a between-session channel for handoffs; and a case-study series that follows each project at three, six and nine months. Together they turn anecdote into longitudinal data.

Take a collective shape to the next funder ask. The OpenAI Foundation's 2026 People-First AI Fund opened while this paper was being prepared, with grantees to be announced by October 2026. Alongside individual asks, the network should explore a single collective proposition: continued convening, evaluation in partnership with a university, shared open-source tooling, and dissemination through the AI for Non-Profits Network. A repeatable model the Foundation could fund as a model, not a one-off grant.

Learn in the open with us

14The 2026 cohort, and what happens next

Building in the Open14 · The 2026 cohort, and what happens next

A living document

This paper is a working draft on purpose. The network revises it as the evidence grows, in the open.

Since the sessions

The recommendations in §13 are moving from asks to workplans, starting with the shared safety and escalation resource.

A draft on purpose.

Two things should be said plainly about where this paper sits. It is early, and it is unfinished, and both are deliberate.

The six findings were drawn from the first sessions of a young network. They are directional signals, not settled doctrine, and the group treats them that way: each new session is a chance to confirm a finding, sharpen it, or strike it. The version number on the cover is the promise. When the evidence moves, the paper moves, and it moves in public.

Progress since the sessions has followed the recommendations in §13. The network has kept convening, the first shared artifacts are in motion, beginning with the safety and escalation resource that three organizations were building in parallel, and the case-study series is being scoped so that anecdote can become longitudinal data. None of this is finished. All of it is visible.

How revision works

This paper is co-designed. Every version folds in feedback, comments, and references from across the network, and the structure improves along with the substance. The findings evolve the way they were formed: through the lived experience of delivering AI projects, and sharing what they teach.

Building in the Open14 · The 2026 cohort, and what happens next

The door was never the grant. The door is the work.Welcoming the 2026 cohort

Dissemination

Through the AI for Non-Profits Network weekly briefing to 4,000+ sector leaders, and each member's own ecosystem.

Room in the room: the 2026 cohort.

The 2026 People-First AI Fund, in brief

In 2026 the OpenAI Foundation is committing $50 million in unrestricted grants, up to 10% of an organization's annual budget, to US community nonprofits across community support services, arts and culture, and community journalism and media. Applications closed July 15, 2026, with all applicants notified by October 2026, and the Foundation expects primarily to fund organizations that have not previously received support.

Source — OpenAI Foundation, "2026 People-First AI Fund"

In 2026 the OpenAI Foundation launches the second year of its funding commitment to nonprofits. When the new grantees are announced in the fall, the network's intention is straightforward: reach out to them, welcome them in, and widen both the cohort's learnings and the opportunity of the network itself, growing year on year as the landscape and the grantees evolve.

The principles stay simple. This is a self-organized group of grantees, there is no gatekeeper and no application, and new winners arrive as peers, not students. What they inherit is a head start: three sessions of hard-won findings, a set of artifacts in progress, and a standing invitation. What they are asked for is the same thing the first cohort asked of itself: bring your builds and your failures to the table, test the findings against your own experience, and help this paper evolve.

Tools built together are shared with the wider sector, and what the network learns travels through the AI for Non-Profits Network's weekly briefing to more than 4,000 sector leaders, and outward again through each member's own ecosystem.

Joining the 2026 cohort? Learn in the open with us

15Appendices

Building in the Open15 · Appendices

Glossary

Plain definitions of the terms used in this paper.

Note

Glossary, method notes, and acknowledgments.

Glossary.

Term	What it means here
LLM	Large language model. The technology behind conversational tools like ChatGPT or Claude that generate text from a prompt.
Generative vs. predictive AI	Generative AI produces new content. Predictive or machine-learning models score or forecast from existing data.
RAG	Retrieval-augmented generation. Grounding an AI's answers in a specific, trusted body of content, rather than the model's general training.
Agent / agentic AI	AI that can take multi-step actions toward a goal, not just answer but do, often chaining tools together.
Multi-agent cross-checking	Using several agents, ideally from different models, to review one another's output and reduce errors and hallucination.
Hallucination	When an AI produces confident but false information. A critical risk in high-stakes, vulnerable-user contexts.
Guardrails	Constraints that stop an AI responding in unsafe ways, and route certain questions to human or vetted resources instead.
MCP	A standard that lets AI tools plug into other systems, such as a CRM, to read and act on data.
Hybrid / forward-deployed builder	An emerging role. Someone close to the mission and operations who can build bespoke internal tools with AI, reducing reliance on external engineering.
HIPAA-compliant	Meeting US health-data privacy rules. A prerequisite for tools handling medical information.

Building in the Open15 · Appendices

Method & limits

Three recorded virtual sessions with a self-selecting, voluntarily attending group. Not a representative sample.

With thanks

To the organizations in Sessions 1 and 2, and to the OpenAI Foundation's People-First AI Fund.

B. Method notes and limitations

This paper rests on three recorded virtual sessions with a self-selecting, voluntarily attending group. It is not a representative sample of OpenAI Foundation grantees or of the sector. Quotations are from computer-generated transcripts that may contain errors and have been lightly edited for readability. Names and affiliations require verification before publication. Read the findings as early, directional signals. The network will test them as it grows.

C. Acknowledgments

With thanks to the organizations represented in Sessions 1–3, listed at the front of the paper. Convened by Whitelabel with the AI for Non-Profits Network. This work would not exist without the OpenAI Foundation's People-First AI Fund, which funded every organization represented here and set the people-first, community-first tone this network has tried to live up to. The paper gratefully acknowledges the Fund's principles of unrestricted, community-led, front-line support, and its stated commitment to learning side-by-side with grantees. The views expressed in this paper are the network's own.

Learn in the open with us