The AI adoption paradox: why 90% is using it and 10% is getting value
AI adoption is the biggest technology shift the for-purpose sector has seen since the move to cloud computing and the proliferation of the smartphone — and it has landed in a fraction of the time. Two years ago, most CEOs I sat with hadn’t typed a prompt into anything. Today, 67% of Australian NFPs report using generative AI and 90% of Australian digital workers say they use AI at work. The Productivity Commission has modelled AI adding more than $116 billion to the Australian economy over the next decade — the largest single productivity opportunity in front of the country.
And yet, only 10% of those same Australian workers say AI has significantly improved organisational performance. MIT’s NANDA study of 300 enterprise deployments — the most rigorous international dataset we have on this question — found 95% of generative AI pilots delivering no measurable P&L impact, against an estimated $30–40 billion in enterprise spend.
Sit inside a for-purpose team using AI at scale right now and you see why. A lot of the day is spent bot-sitting: re-prompting, fact-checking, rewriting, undoing, and quietly cleaning up after the tool. The staff member who “used AI to draft the funder report” has often spent longer correcting, verifying and reversing the model’s output than they would have spent writing it themselves. The productivity gain shows up in the individual’s self-report. The lost time shows up nowhere — because no one is measuring it.
The assumption behind every board conversation I’ve been part of in the last twelve months has been that personal AI use at scale would quietly convert into organisational productivity. It hasn’t. The gap between adoption and value in the for-purpose sector is now large enough to be a governance problem in its own right — and it is not the technology’s fault. It is that most organisations have deployed AI without redesigning the work around it, and the hidden cost of bot-sitting is eating the promised gain.
The way through is not a bigger tool budget. It is training that builds real capability, a culture that rewards experimentation, and leadership with the courage to fund focused pilots pointed at front-line staff and the people they serve. Three shifts follow.
1. Personal use has not scaled into organisational value — and the assumption it would was wrong
The 90% adoption number is real. Microsoft’s 2026 Work Trend Index has 63% of Australian AI users saying they’re producing work they couldn’t have a year ago, and 68% saying they fear falling behind if they don’t adapt. Individually, this is a productivity story.
Organisationally, it isn’t. Australia doesn’t yet have a large-sample sector benchmark on AI value capture in NFPs, but the US 2026 Nonprofit AI Benchmark of 346 organisations is directionally useful: 92% have adopted AI, but only 7% say it has meaningfully expanded what their team can accomplish. 65% describe their AI use as reactive and individual — one-off prompts and personal experimentation — and 81% report using it ad hoc. The pattern is that AI is helping people get more of their existing work done, not helping organisations do meaningfully different or better work. Infoxchange’s Australian data reinforces the point: high adoption, low policy and cyber coverage, minimal formal integration.
MIT’s GenAI Divide names the same pattern from the corporate side: general-purpose tools have near-universal adoption and “primarily enhance individual productivity, not P&L performance.” The lead author is specific about the cause — not model quality, not regulation, but a learning gap in which tools never adapt to actual workflows and organisations never adapt their workflows to the tools.
This is the assumption boards need to retire in 2026. Individual proficiency does not aggregate into organisational capability by itself. It has to be engineered. That engineering is a training and capability job as much as a technology one — role-specific AI skills, shared prompt patterns, task-level playbooks, and structured time to practise on real workflows. Most for-purpose organisations have done less of this work than almost any other investment in the last decade, and it is now the constraint.
2. Experimentation is a capability the sector has never built — because risk has only ever meant downside
Doing more with less through AI is not a licence problem. It is an experimentation problem. And experimentation is precisely the muscle the for-purpose sector has been least incentivised to build.
The risk lens most Australian boards apply to AI is downside-only: reputational risk, privacy risk, bias risk, funder risk. Those risks are real, and caution about them is not misplaced. The Productivity Commission has explicitly recommended against heavy-handed AI-specific regulation in favour of amending existing law — an Australian policy signal that the regulatory environment is being deliberately shaped to enable adoption, not slow it. But at board level there is still no equivalent conversation about upside risk — which is really the opportunity cost of not moving: the risk of falling further behind on the technology, on systems, on operations, on the refinement loops that let organisations get better every quarter, and on the impact AI can now put within reach.
Named concretely, that means missing productivity gains, missing mission impact, being outcompeted for talent, and being unable to deliver services within tightening funding envelopes. The UK sector is starting to name this directly: the Charity Governance Code refresh in November 2025 now explicitly requires charities to have a technology and AI policy, and NCVO’s position is that charities should “enable responsible use rather than prohibit all use”. The Australian sector has not yet made that move at code level.
The consequence is that few for-purpose organisations have working experimentation loops. MIT’s data on why the 95% of pilots that fail actually fail is instructive: it is not the models. It is brittle workflows, lack of contextual learning, and misalignment with day-to-day operations. Pilots that integrated into a specific task, with measurable inputs and outputs, succeeded far more often than pilots that tried to do “AI strategy” across the whole organisation. Buying beat building 2:1: external partnerships with learning-capable tools reached deployment about 67% of the time, versus about 33% for internal builds.
For-purpose organisations do not need to run experiments the way a tech company does. They need a defensible way to try something small, measure it against a baseline, decide whether to scale it, and kill it fast if it doesn’t work. That is a culture question before it is a process question. Staff need explicit permission to try, to be creative, and to close things that don’t work without career cost. Without that culture layer, no amount of process will produce experiments worth learning from.
3. Boards want more AI — but haven’t made the budget or the trade-offs it actually requires
This is the change I think matters most, and the one most boards still haven’t made honest with themselves about. There is now near-universal board-level appetite for AI, innovation, and productivity. There is almost no matching board-level willingness to fund it, to redesign around it, or to trade off other priorities to make it possible.
The Microsoft Work Trend Index puts hard numbers on the Australian version of this leadership gap. Only 28% of Australians say their organisation is clearly aligned on AI strategy and policies. Just 13% say reinvention is recognised when results take time. And 51% say it feels safer to focus on current goals than rethink work with AI. This is not a workforce that lacks appetite. It is a workforce whose leaders have not resourced or protected the redesign work.
The Australian charity sector is more exposed than most. Infoxchange’s 2025 report shows 67% of NFPs using AI — but only 14% have an AI policy and only 23% have a cyber security plan. Paying What It Takes (Centre for Social Impact, SVA, Philanthropy Australia) documented the underlying dynamic years ago: most funders won’t fund the indirect costs of leadership, systems, data and evaluation, so charities chronically underinvest in the very things that make AI adoption succeed. Private non-profit R&D spend in Australia sits at $1.69 billion — 0.06% of GDP, unchanged in real terms.
The MIT data is a warning about where the money that does get spent tends to go: more than half of enterprise generative AI spend targets sales and marketing, while the strongest returns sit in back-office automation. Boards that call for AI without directing where the investment lands are effectively delegating the productivity strategy to whoever happens to be experimenting with a chatbot that week.
Appetite without budget is not strategy. It is aspiration. What the sector needs from leaders now is different — the courage to fund training that lifts real capability, to protect the culture that lets staff experiment, and to back focused pilots aimed squarely at front-line workers and the people they serve, rather than diffuse “AI strategy” exercises that never touch the work. And the sector cannot afford aspirational AI in the fiscal environment now closing in around it.
How to close the AI potential gap in 2026
The technology has arrived. The productivity hasn’t. The reason is almost never the model — it is the operating discipline underneath it, and the honesty of the board conversation on top of it.
For senior for-purpose decision-makers, three moves matter now-
1. Invest in training and capability so people can do the work, not just use the tool. The 90% adoption number is really a 90% “I opened the app” number. Real value requires staff who can structure a prompt, judge an output, and integrate AI into a task — role-specific, task-level training, repeated. It also requires named capability owners who make the learning compound. Without paid time to build those skills and embed them, the bot-sitting tax keeps eating the gain.
2. Change the culture so experimentation and creativity are the default, not the exception. Boards need to name upside risk — the opportunity cost of falling further behind on technology, systems, operations, refinement and impact — in the risk register alongside downside risk, and back it with a culture that lets staff act on it. That means explicit permission to run small experiments, to be creative in how AI is applied, and to close things that don’t work without career cost. It also means holding privacy, ethics and governance conversations in the same room as experimentation, not in opposition to it. A “safer to focus on current goals” default is a strategic choice with a price tag attached — boards should be pricing it and choosing consciously.
3. Lead with courage — fund focused pilots aimed at front-line staff and the people they serve. MIT’s finding is unambiguous: the pilots that fail almost never fail on model quality. They fail on workflow, data readiness, governance, and ownership. The leadership move is to pick a small number of focused pilots pointed squarely at front-line staff and beneficiaries — the intake worker, the case manager, the fundraiser, the person on the other end of the service — and to fund the operating layer underneath them properly. Something else has to give. The organisations that will show measurable AI impact in 2027 are the ones whose 2026 boards had the courage to make that trade-off explicit.
The 90% adoption number is a gift and a warning. The gift is that the workforce is already ahead of the leadership. The warning is that the advantage won’t last if boards keep asking for outcomes without funding the conditions. The organisations that close the gap in 2026 will be the ones whose boards invested in the skills, protected the culture, and had the courage to back pilots that put AI to work for front-line staff and the people they serve.
If you’re a senior leader or board director thinking about where to put your next AI, experimentation or capability investment, I’d value the conversation. Send me a note to mike@shackletonlabs.com.
This article was first published here on LinkedIn on 7/7/26.
