Thirteen staff members completed the diagnostic between August 3 and 12. This page gives us the full picture: where everyone already agrees, what they want from the tools, the people we think could help carry this work as AI Guides, and a handful of deeper findings worth discussing on Friday.
DRAFT V1LEADERSHIP PRE-READ
The team is further along with AI than we thought. All thirteen already use AI for work, five of them daily. They are not asking whether Echidna should use AI; they are primarily asking for the guardrails, the accounts, and the training to use it properly. The strongest pattern in the data that emerged was that of a team that has already intuited most of the policy and is waiting for the organization to catch up on paper and infrastructure.
Rather than plotting the team by whether they are pro- or anti-AI (almost everyone is somewhere in the cautious middle), the axis that matters more for Echidna, and that predicts whether the AI policy will take hold, is what kind of wrong they think AI does.
| GROUP | N | WHAT THEY ARE | WHAT MOVES THEM |
|---|---|---|---|
| Standard Setters HIGH PRACTICE · RELATIONAL | 4 | Heavy, capable users whose objections are about consent, voice and judgment rather than security. They will not be reassured by an enterprise contract. | Give them authorship. These are the people to put on the policy drafting, because they will hold it to a standard the rest of the team will likely accept. |
| Configure and go HIGH PRACTICE · PROCEDURAL | 3 | The deepest users and the largest live exposure. Between them they run eleven personal accounts. They believe good configuration solves the problem. | Good defaults, fast. They will adopt whatever is provisioned. They are also where the migration off personal accounts either succeeds or does not. |
| Principled and unpersuaded LOWER PRACTICE · RELATIONAL | 5 | Lighter users with the sharpest ethical reservations: creator consent, dehumanisation, environmental cost, authenticity of voice. | Not a demo. Address the objection on its own terms in the policy text, and give them a genuine opt-out on the tasks they consider off-limits. |
| Awaiting the rails LOWER PRACTICE · PROCEDURAL | 1 | Will follow whatever is set up, and says so. | Clear instructions and a default. No persuasion required. |
Depth of practice does not predict permissiveness. Four of the seven deepest users hold relational objections, and one of the two most capable respondents in the set is also the most insistent that grantee data requires consent before it touches a model. The people using AI most are not the people to worry about. The people to worry about are the three who believe configuration is sufficient, because between them they run eleven personal accounts while rating their own judgment about what is safe to share at 8, 8 and 10 out of 10.
The vertical axis is my coding of each respondent's written answers, not a survey item. It is reviewable: every point carries the sentence it was coded from. If any placement looks wrong to you, it is worth ten minutes, because the quadrant assignments drive the recommendations.
Four commitments show up in nearly every response, across Programs, Grants Services, and Operations.
Our leadership conversations left an open question: for grant materials, is a thorough review of AI's output enough, or must the staff read the source materials first? The staff POV is 11:2 for read-first, and the two who would accept an AI first pass both attached heavy conditions on it and cited volume pressure as the reasoning, rather than on principle. This lines up directly with the 2022 Echidna principle about reading and reacting to what grantees share, so the team concerns and the principles agree, which is great. We recommend we put "Human always reads first" into the policy.
“Otherwise there is a massive risk that AI's framing, omissions and biases shape their understanding before they form their own judgment.”
“Only reading bits and pieces or a strong summary misses nuance that will be important to note and respond to as we engage with grantees.”
Accuracy first, unanimously. Then privacy, then fitting it with the Workspace tools they already work in. Affordability barely registered at 2 of 13. Eight of thirteen answered “depends on the task” when asked whether they want AI embedded in existing tools or opened deliberately for specific work, and nobody asked for anything custom-built: the appetite right now is for a small set of governed tools with saved prompts and templates over time. The pathways page picks it up in detail.
All thirteen named drafting and editing. After that: brainstorming and research (8 each), summarizing (6), meeting notes (5), data analysis (3). By function: Grants Services want reporting, grant tracking, and template-driven data entry. This is the clearest, most automatable pipeline with the strongest appetite from staff. Operations want calendar logistics, vendor research, and org-wide search. Programs want board write-ups, synthesis across partner reports, and scoping and diligence support.
Twelve of thirteen do at least some Echidna work in personal AI accounts, and six are paying for tools out of their own pockets. We see a version of this at almost every organization at this stage and size. The instincts underneath are mostly the right ones: people steer their most sensitive work toward the accounts they trust most. The gap here is infrastructure and rules.
Three things in this chart need a discussion and decision.
The team rates itself lowest on technical understanding (mean 3.2, nobody above 6) and highest on prompting and evaluating claims (both just above 5). This is a healthy shape: technical depth is the least important of the six components for using these tools safely. We recommend training starts from fundamentals and basics.
Two ratings should drive the training curriculum. Knowing what's safe to share has the widest spread in the set, running the full 1 to 10, which tells us data rules need to be written down and checkable rather than left to individual judgment. We recommend a simple Echidna AI checker tool for staff to check against in the policy. Clarity on which of my tasks AI could help with averages 5.0, and the Leadership wave scored it low too, so the task-mapping workshop is the highest value training we'll likely run.
The red lines from our conversations hold up well against the staff data. Staff extend them in four directions, and three of the four are about people more so than data.
| STAFF ADDITION | CONCERNS IN THEIR WORDS |
|---|---|
| Strategy is built collectively | “Using AI to build our strategy rather than collectively building it.” |
| Performance feedback stays human | “Using AI to deliver performance feedback rather than receiving feedback from my manager personally.” |
| Guardrails are built together | “Using AI without guardrails that we build together as an organization.” |
| Voice and authenticity are protected | “Keep your personal voice, even if it has some grammatical mistakes. Authenticity goes a long way!” |
We previously discussed AI notetakers as a storage and disclosure question. One staff response adds a different take, from direct experience in a previous role: notetaking in sensitive one-to-ones changes what people are willing to say, because they worry about where their words may end up. The control that answers this is a per-meeting opt-out any team participant can use without explaining why, plus a hard carve-out for HR, personnel, and one-to-one conversations. If that feels right, we recommend drafting that clause into policy.
Two people raised the environmental cost of AI as concerns, and one raised how models are trained on creators' work without consent. Neither is something Echidna can fix, but neither should be waved away either. There is a lot of misinformation about environmental costs of AI out there, so it's important to be cautious about what claims are true. Our suggestion: the policy acknowledges both honestly, and usage is optional on tasks where someone objects.
The rollout needs a small group of staff who model good practice, help colleagues, and keep the policy up to date. Sara and I read all thirteen responses closely with that role in mind. Four people stand out, each for a different reason, and together they cover policy, craft, operations, and geography. To be clear on status: we have not approached anyone yet. This is our proposal for your reaction.
Ellen's only person on the team working entirely inside Echidna accounts, and a daily user. She mapped the 2022 principles to AI rules in her own response and her answer on grant assessment reads like a training rubric: anchor the prompt in your own read, check the output against the source, own the final product. She can model the destination really well since she's already working in that space.
Co-draft the policy language in September, alongside us. Be the worked example for the migration, since her setup is already the destination. Host one office-hours session on her read-first workflow once training starts.
Arjun is the strongest self-rated prompter and evaluator in the set, and the only person with peer evidence: a colleague cited his AI-assisted explanation of standard deviations as the thing that most impressed them about AI use. He also holds concerns on consent and environmental cost, which is why cautious colleagues will trust him.
First, whether he wants the role, given how intentionally he uses AI. If yes: teach the challenge-the-model module in training, and act as peer reviewer on high-stakes synthesis work.
Katrina's the broadest daily user on the team, and the one who's concernd with who keeps this alive after the engagement ends. She has carried the coordination from day one, and she thinks in systems: accounts, templates, workflows, cadence. We'd like to put the account estate and the prompt library in her hands.
Own the account estate and admin settings once migration lands. Keep the prompt library current. Carry the policy's review cadence after our handover in October.
Dennis is a daily user across six tools with a clear personal discipline: AI helps me work faster, I still review and make the final decisions. Nairobi-based; without him the guide bench sits entirely in the US and UK, and a rollout led from San Francisco alone sits poorly next to Principle 5. We'd like one conversation about how he actually works day to day before confirming.
One conversation about his day-to-day workflows first. If it fits on both sides: anchor training for the Nairobi side of the team and be the first point of contact for colleagues in the region.
A few people fit better as named contributors than as guides, and we'd rather use their strengths precisely.
The most technically literate sceptic on the team, with the clearest grasp of how these models are often sycophantic and can quickly backtrack. We'd ask her to pressure-test the policy before it ships and to draft the HR and one-to-one carve-outs.
She wrote the best line across all responses: "AI can be approved for the task without being approved for the data." That is a two-gate structure we'd build the policy on, and we'd want her in the drafting review.
He mapped the Grants Services pipeline more clearly than anyone and asks sharp questions about vendors. We'd bring him into the tool evaluation and the workflow design for grants, if we choose that pathway for live builds.
On process: our next step is two or three short conversations (Dennis, Amanda and Jake, possibly one more) to fill the gaps, then we bring you a confirmed guide team. If any of these placements doesn't match your read of the person, it would be great to hear your take.
Each one shapes how we'd sequence the work, and one of them pushes on a decision we've already made together.
Twelve of thirteen working in personal accounts looks alarming on a first read. But since Echidna hasn't yet chosen or provisioned its tools, that is the gap this engagement should close. Faced with real work and no organizational option, the team did the sensible thing: they found the tools themselves, six are paying from their own pockets, and many will hopefully have already switched model training off in their settings. The instincts in the team are good ones. This is a finding about missing infrastructure, and the fix is the tool decision on Friday, followed by training and migration.
The closest analogy is running Echidna business from a personal Gmail. The person may be careful, the inbox may even be well secured, but it still can't be what the organization relies on, because Echidna can't see it, can't require anything of it, and has no agreement with anyone about what happens to the contents when they leave. The same gaps apply to personal AI accounts, and none of them close with just a settings toggle:
Which means account migration is a week-one move under either pathway: enterprise spaces for everyone, personal accounts retired for work, on the same logic that puts work email on @echidnagiving.org.
All thirteen condemned the scenario, but they split cleanly on the reasoning. Roughly half read it as a security problem that an enterprise account with training switched off would fix. The other half read it as a consent problem that an enterprise account tier wouldn't fix, because the data belongs to the grantee, who was not asked. The July red line we built solves the first group's concern fully but not the second group's. If the policy ships with only the enterprise-account rule, about a third of the team may follow a policy they believe misses the point, and that is the most common situation within which policies stop being followed.
“Would my answer change with an enterprise tool? Yes... appropriate privacy and security controls.”
“My answer wouldn't change if they used an enterprise tool instead. The issue is sharing another company's private information.”
The policy needs an explicit position on when grantee material can go to a model at all, what grantees are told, and what de-identification happens first. The team has already handed us the sentence to build it on: AI can be approved for the task without being approved for the data. Tool approval is one gate; data approval is a second, separate gate. It maps cleanly onto Principles 2 and 3, and it came from your own team, which is a fantastic spine to begin with.
One respondent raised a risk we hadn't heard yet: Echidna's methodology, values, and grantee details becoming visible to actors who would penalize gender-equity work on ideological grounds. This might be a line we add in the risk framework plus a few practical vendor questions on data residency, retention, and government requests, which the pathways page already carries into the comparison.
The team's most-cited fear, over-reliance, currently has no instrument to judge it is or is not happening. When we asked how anyone would notice their own thinking going generic, eight of thirteen said they couldn't. The five who answered offered cheap mechanisms: form your own view first and compare, be able to explain a conclusion without the transcript, and lightweight peer spot-checks. These may be worth folding into the risk framework when we build it out.
The session is the tool decision, and this page plus the pathways pre-read are the inputs. The agenda:
Read the guide proposals in Part 5 with your knowledge of each person in mind, and come with your reaction: anyone miscast, anyone missing? The survey shows us how people think about AI but you know things about workload, tenure, and team dynamics that we don't. Please shout if anything on this page looks off.
Based on 13 staff diagnostic responses received August 3–12, 2026, read alongside the 3 leadership responses from July 22–27. Quotes are verbatim. Analysis by SA Advisory (Shabnam + Sara).