What a 30-person foundation learned about AI, the middle manager problem, and a startup making AI you can trust

Here’s what landed for me this week and what I feel is worth your time:

The Annenberg Foundation spent six months doing something well that most organizations jump past when they think about AI: they asked their staff and grantees what they actually needed before buying any tools.

We know that AI adoption is largely being driven from the top down. Annenberg took a different approach and ran listening sessions with every employee starting from "what are your pain points," not "where could AI fit."

They found that 63% of staff were already experimenting with ChatGPT, Copilot, and Canva's AI features — mostly ad hoc, no guidance or governance. Nice to see the "at-bats", but not so great to see this happening outside the purview of the org’s basline cyber security or existing data surfaces.

While that discovery validates the importance of getting (sanctioned) tools in the hands of staff for exposure and learning, what I love about the approach is that it spends the bulk of early time and resources identifying the clearest opportunities for value. We know that one of the critical tensions in introducing AI technologies to an org is the apprehension of individual contributors. A posture of "AI is here to complement and augment our expertise, not to replace" feels genuine when the conversation starts with "how can you best be helped?"

The most telling finding: when they surveyed their grantees, AI for fundraising, outcomes prediction, and information analysis was classified as urgent needs — much more urgent than the foundation's own staff felt. I think this is natural; people who contribute want their contributions to go as far and as fast as possible in creating impact. The people who do the work know that any process change takes time to be safe, effective, scalable, predictable.

That tension is a real story in the sector right now. The Annenberg process is detailed in TAG's Member Spotlight Series (https://www.tagtech.org/ai-resources-for-philanthropy).

Harvard Business Review published a piece this week that named the bottleneck between the two dominant ends of the implementation conversation: end-user value and strategic direction.

Middle managers are the ones tasked with turning executive AI ambitions into operational reality, and they're getting tepid support. No training. No budget. No license to say "this use case doesn't make sense for us." They're caught between a board that wants AI transformation and a team that's just trying to get through the week.

For NFPs this hits harder than for most orgs because there's only a thin middle management cushion in the best of cases. Your program directors, your ops leads, your development managers — they're doing the work and absorbing the AI learning curve.

The HBR takeaway: if you're rolling out AI without budget or space for manager support, workflow redesign, and permission to push back on bad ideas, you're setting them up to fail.

Hallucinations are one of the biggest reasons NFPs can't deploy AI in high-stakes contexts — benefits eligibility, grant decisions, client referrals, health information. If you can't trust the output every time, you can't use it. LLMs are inherently non-deterministic. Agent workflows can be configured to try and add more validation and predictability by breaking up tasks, but this can only go so far.

Probably raised $9M to build toward "deterministic-grade reliability," which is exactly what the sector needs before most mission-critical use cases become viable. The funding is small relative to the problem, but the category validation matters. Investors betting on "reliability-first, speed-second" AI signals that the industry recognizes reliability as the real deployment bottleneck, not raw capability.

For now: test every AI output before it reaches a beneficiary or donor. But the category is moving.

That's it for this week. See you next Tuesday!

If you got something out of this, the best thing you can do to support the project is to forward this to a colleague who is figuring this out too. Many thanks to everyone who has got in touch over the last few weeks! Your kind words and stories are so appreciated.