Emily and Dana anchored the call, Sara joined from Nairobi, and Kim dialed in from vacation for the middle stretch. We walked the diagnostic findings together. Nothing in the results surprised the room, which is itself a finding. The real work happened in four places: the failure-modes conversation, the red lines, the tool question, and the shape of ongoing support. Decisions first, discussion below.
Three stances were on the table: discourage AI, stay passive with guardrails, or actively encourage. The room chose the middle deliberately, and framed it as a position to revisit rather than a permanent stance. Nobody is required to use AI; those who do follow the policy. If the productivity gap between users and non-users grows, that conversation reopens.
The champions idea became AI Guides (Sherpas stays on the table as the alternative name): 3 to 4 people, representative across regions and levels, opt-in first. The rename is substantive. A champion cheers; a Guide shows you the route, including telling you where a route is a bad idea. Guides get explicit license to call out poor or lazy AI use, backed by an external retainer for the harder questions. The room preferred this over an accelerated hire: the feeling was a hire waits about a year, until we know where the real gains are. The staffing memo reflects this now.
We tabled the single-tool assumption and walked the alternative: a small set of approved tools inside one policy, one data standard, and a shared sandbox. Instincts in the room ran both ways: openness to letting people make a case for a tool that meets the policy, tempered by back-end management, compliance across products, and cost. Cost got a hard flag: enterprise AI spend can balloon far past what was budgeted, so whichever shape we pick, spend gets monitored and capped from day one. Worth knowing: today's enterprise tools have no real usage dashboard, so oversight stays human. The staff diagnostic will show whether tool preference is broad or concentrated, and the August 21 session decides.
Four edits agreed: the scale moves from 1–5 to 1–10 (the room felt 1–5 flattened nuance, and a live re-rating on the wider scale proved the point), the no-wrong-answer framing comes out, scenario 3 splits into two sub-questions (one where the person read the proposal first, one where they didn't), and the Echidna Principles get woven in. New grant applicants becomes new grant proposals. Scenario 4 stays as is: even where the honest answer is we don't know, the question plants the right seed.
The week of September 7 is board week, so it's dead for this work. New sequence: policy drafting the week of September 1, senior staff review of the draft async by email, with a live conversation added only if the draft surfaces misalignment, then sign-off late the week of September 7 or the week after. The Board stays out of the detail: the room's read is that the Board set the sandbox and doesn't need another pass unless the policy strays beyond that guidance. The schedule page gets updated once dates firm up.
The starter list from the findings page held. The room sharpened it in four places:
On policy style, the steer was to sit between two bad extremes: a policy that just says use your judgment, and one so specific nobody reads it. Illustrative examples over exhaustive rules.
This was the deepest thinking of the session, and it will shape both the policy and how we measure success.
The erosion worry is about slow drift. The image used in the room: a frog in boiling water. The risk to watch is people shifting from initiating work to merely editing AI output, where judgment doesn't disappear but engagement quietly does. For program work specifically: if people stop reading what grantees submit, the damage shows up later as eroding trust and slipping grantee perception scores, and by then it's been happening for a while.
Newer staff carry a specific risk. Someone who reaches for AI before they know what good looks like compounds the problem of not knowing what they don't know. Onboarding and the Guide model both need to account for this.
People process differently, and the policy should respect that. For someone who thinks by writing, handing the writing to AI shortcuts their own learning. Someone who processes verbally might genuinely sharpen their thinking by talking a problem through with AI before writing a word. Blunt rules can't capture this, which is one more argument for teaching judgment through examples rather than legislating every case.
The room wants early detection, and named some instruments: a staff satisfaction pulse roughly every six months, watching the grantee perception survey, possibly leaning further into in-person site visits, and quicker lightweight grantee check-ins between the big three-year survey cycles. There's also a genuinely useful question in the mix for the policy: what do humans need to keep doing enough of, themselves, to stay capable of checking AI's output?
Past instances of clumsy AI use were handled by quiet redirection, and the room named the tension directly: come down hard and people hide their AI use, which is the worst outcome. The direction agreed: senior staff model transparency by sharing how they use AI including where it fails, staff share the AI output alongside their own process, and audits stay collaborative. The Guide role is built for this, guiding rather than enforcing.
One idea from the room to carry into the build: teach the enterprise tool the policy itself, so it can flag in the moment when a request brushes a red line. Doable with today's tooling, needs stress-testing before we rely on it, and worth prototyping during setup.
The framing the room settled on: use AI where it increases impact, and keep it away from where it diminishes quality. The clearest internal candidates named: grant planning and forecasting tools, data dashboards (the current tools are clunky), strengthening writing, and private, disclosed use of AI as a coaching sounding board.
Two threads stay open deliberately. Grantee-facing AI education is off the priority list for now without being off the list forever. And there's real ambition in the room about AI for grantmaking itself: pattern recognition across the portfolio, meta-analysis of what's worked in education philanthropy, anticipating unintended consequences of the interventions Echidna funds. The Board is unlikely to fund external AI-in-education work, and the room wants the internal analytics option preserved. Both belong on the 2.0 list.
Prepared by SA Advisory from the July 29 working session. Discussion points are recorded without individual attribution as our standard practice for session notes. Please shout if anything looks off. Next on the calendar: tool decision session, Friday August 21.