In September 2026, a group of NGO CEOs in health, community and human services sat down for a private off-the-record breakfast conversation. The room ran under the Chatham House Rule. The topic was AI, and what is happening with it inside their organisations.
Everyone in the room is working with funding that will not stretch to meet demand, so anything that frees up time gets a serious look. What comes with it is a compliance, privacy and security load that everyone is still working out. Some are further ahead and others are playing catch-up.
Most have policies that date rapidly
Almost everyone in the room has done policy work. For some it started two years ago, for others more recently, and one or two are still working with something informal. Several have a register setting out which roles can use which platforms and for what kinds of data. Boards have had the governance discussion, and some are having it on an ongoing basis.
Nobody described their position as settled. The people furthest along were the ones most conscious of needing to go back and update what they’ve done. Agents arrived and changed what policies need to cover.
Policies are setting out which tools people are allowed to open, but a common gap CEOs are experiencing is the quality of people’s thinking while using the tools.
Judgement is as important as ever
One organisation ran a hackathon. Pairs were given the same scenario and each pair used a different gen AI tool. The answers came back quite different from the same starting brief. The discussion afterwards was about what led to the variation: was it the tool, the pair’s input, or something else. Several organisations are running prompting training, but the evolution in tools means that what was good practice six months ago may not be today.
The missing thinking is showing up in documents
Several CEOs described the same thing arriving from their teams. A long document, well formatted and grammatically correct, with the deeper thinking missing.
One CEO tells their team there is no point employing them if what arrives came straight out of gen AI without their judgement across it. Another set a rule that the effort put into a document should match the effort being asked of the reader. Thirty pages produced in a minute and passed on takes more of the reader’s time than it saved.
Tenders are showing the same pattern
One organisation that commissions services described a recent tender. Thirty-seven applications came in and thirty of them read much the same. The three that made the shortlist showed gen AI had been used to interrogate the problem rather than to write the answer. Submissions are read by panels that include people with lived experience, and what those panels are looking for is the value an organisation can show it will add.
Several organisations have set rules for grant writing. In one, no application may be written by gen AI. In another, gen AI drafting stays inside the organisation’s closed system and a person writes the final version. That keeps the text out of external tools that could later flag it as AI generated.
Staff surveys provide valuable insights after implementation
Several CEOs said they would take this back to their organisations. One organisation surveyed staff about six to seven months after rollout of gen AI tools and had close to an 80 per cent response. Sixty-seven per cent of those using the tools found them very helpful and a further 26 per cent helpful. Two in five use them up to two hours a week, just under a third for two to five hours, and one in five for five to ten. Time saved was the biggest single response, ahead of accuracy and less repetition. What other CEOs wanted from the same exercise was a picture of who is using what, and why.
Closing access to platforms does not remove the exposure
Staff will find their own tools if the organisation does not provide them. Two approaches to access came up. Some have closed access to platforms so staff cannot reach anything other than approved tools. Others have left the tools open and given firm guidance about what must never be typed into an external system. Nobody described either one as solving the problem.
The exposure people were least comfortable with is the one they cannot see. AI is now built into a lot of third-party products, and an audit done six months ago is already out of date. From 10 December, organisations covered by the Australian Privacy Principles have to set out in their privacy policy where computer programs make, or substantially help make, decisions that could significantly affect someone’s rights or interests, and what personal information those programs use. Organisations need to be asking suppliers what has been built into their products.
The risks are landing on people
Phishing risks are increasing, and sharing information on LinkedIn looks like a trigger. One organisation shared that new starters are receiving messages within a day or two of announcing new roles on LinkedIn. The announcement on LinkedIn is connected with other publicly available information to develop a picture of the chain of management and potential weak links. CEOs reported gift card scams, urgent payment requests, and staff being groomed through a short email exchange before anything is asked of them. One organisation traced a breach that began when an attacker took over a legitimate mailbox at another service, so the request arrived from a real address. Cyber insurers came up as an under-used resource. Some will fund prevention, training and testing.
Inside organisations, some staff are taking workplace grievances to gen AI instead of to their manager. The problem is the gen AI lacks perspective and agrees with the framing it is given, so people come away more certain than when they started, and something that would once have been a conversation arrives as a formal claim. One CEO reported losing long-serving staff because the technology grew more complicated faster than the support they received to keep up with it.
Safe Work Australia published guidance in July on AI and digital technologies. It names work intensification, surveillance and unfair or discriminatory decision making as psychosocial hazards employers have to manage and consult on.
Accountability cannot be handed over
An organisation still answers for what it publishes and what it decides, and the context has to be put back over whatever the tool produces.
That includes the voice of the people organisations serve. One CEO is seeing people with lived experience reach for gen AI because they believe it will say it better than they can. Government then hears a polished version without the experience behind it.
One CEO shared that funders are pressing providers to deliver more online when one in four Australians does not have regular access to technology or bandwidth.
The core issue isn’t new
In 1966 Joseph Weizenbaum built Eliza, the first program people talked to and believed understood them. He spent much of the rest of his life arguing about it. Ten years later he since we do not now have any ways of making computers wise, we ought not now give wrote: computers tasks that demand wisdom. Critical thinkers are needed more than ever, in teams, in leaders and in suppliers.
The Peer Room is an invitation-led breakfast for chief executives of NGOs in health, community and human services. The Wider Lens is a version of it for the senior leaders below them. Email susanne@lebout.com.au if you would like an invitation to either, or to put one of your senior leaders forward.
