When AI Stopped Being Magic: What 2025 Taught Us About the Real Work Ahead

What happens when the shiniest new technology stops dazzling us and starts demanding something far more difficult: actual change?

John Hetherington brought the rain but not the cold when he flew in from Canada to London, a fitting metaphor for the conversation that followed. Because whilst 2025 didn't freeze AI innovation in its tracks, it certainly dampened some of the euphoria that characterised the previous three years. And according to John, that's precisely what needed to happen.

"After the whole 2022 release and then 2023 hype and 2024 even more hype, 2025 seemed to hit a bit more of a trough of disillusionment," John explained, invoking the Gartner hype cycle with the weary familiarity of someone who's seen this pattern before. "A lot of people were like, Is it really giving what we expect? Is it giving the value we were promised?"

The question wasn't rhetorical. Across boardrooms in North America and beyond, companies were demanding proof. The grand promises of massive cost savings and transformative productivity gains were meeting the cold reality of spreadsheets and quarterly reviews. 'Show me the numbers,' executives insisted. Prove it works.

The Dirty Secret Nobody Wants to Admit

But here's where the conversation took an unexpected turn, revealing something that rarely makes it into the polished case studies and vendor presentations. People were using AI and then treating it like a guilty secret.

Two stories from earlier in the year perfectly captured this bizarre dichotomy. In one company, an employee was made to feel like they were cheating when they used AI. "I'm never gonna tell anyone again," they resolved, retreating into secrecy. In another case, a CTO of a well-known tech company took the opposite approach. When an employee sheepishly admitted to "maybe using AI a bit," the CTO's response was immediate: "I want you using it. I'm paying you good money. Why would you not?"

The contrast speaks volumes about organisational culture, but it also reveals something more profound about our collective uncertainty. We're still working out what it means to augment human intelligence, still navigating the uncomfortable space between automation and authenticity.

John witnessed this firsthand with a client whose leadership team created board reports using AI, then lied about it when asked. "The reason they said no is that the tone of the company was, We're not there yet with AI," John explained. His advice? Turn the question around. Don't ask whether AI was used. Ask how AI should be used going forward.

It's a subtle but crucial distinction. One question looks backwards with suspicion. The other looks forward with intention.

The Intelligence Illusion

The group identified something that many organisations are only now beginning to grasp: AI was sold to us on the promise of intelligence, and we bought that promise wholesale. The benchmarks seemed extraordinary. PhD-level reasoning. Abstract problem-solving that rivalled human experts. The language models sounded so convincingly human that they passed informal Turing tests in casual conversation.

"When you first started speaking to these things, it was like the Turing test, right?" came the observation. "They were so well-versed, and they sounded so human-like and smarter than the average person. And we were taken away by that."

But then something interesting happened. As people began using these tools within their own domains of expertise, the illusion started to crack. "When you speak to an AI, and you realise, if you're speaking to it inside your specialism, it's not that smart," the conversation continued. "When you speak to it as an outsider to that specialism, it sounds amazing."

Why 95% of AI Projects Failed

This revelation cuts to the heart of why so many AI pilots failed in 2025. An MIT study found that 95 per cent of AI projects weren't delivering on their promise, and the reasons had little to do with the technology itself. The problem wasn't that the models weren't intelligent enough. The problem was that organisations expected intelligence alone to solve business problems that actually required context, culture change, and careful integration into existing workflows.

The conversation brought up a systems-thinking principle that perfectly explains the failure: Ashby's law of requisite variety. Businesses are complex because they solve complex problems, which is why they have governance frameworks, partnerships, departments, and specialisms. "People are rolling out Copilot without embedding it into the business, without the upskilling, without learning workflows," came the explanation. "You haven't embedded AI into the complexity of your business. Therefore, it will fail to solve complex problems."

The Onboarding Problem

John's analogy was spot on: hiring a brilliant person who knows nothing about your business, your industry, or your methods. No matter how smart they are, they'll fail without proper onboarding. "We have onboarding processes," he noted. "A few months in, then you start to see the value coming out. It's the same thing with AI. You need to embed them into the complexity of your business."

The Return on Investment Paradox

Here's where things get really interesting, and where the conversation touched on something that explains why measuring AI's impact has been so frustratingly difficult.

A point was raised that should make every productivity-focused executive pause: "What if you're overworked already? You're going to use AI, you're going to get the time savings, and that hidden time saving you've made, you're going to use it to get some of your day back."

In other words, people aren't reporting their productivity gains because admitting they now have spare capacity could mean more work piled onto already full plates. So they quietly use AI to complete tasks faster, then use the time they save to breathe, think, or avoid burnout.

John encountered this resistance firsthand with accounting firms. Month-end close processes that traditionally took anywhere from five to twenty days could be dramatically accelerated with AI. But accountants pushed back: "If I reduce the time spent on reporting, what do I do instead? I don't want to reduce my time because this is my job. I get self-worth from reporting."

The question everyone kept circling back to: what strategic work should people be doing with their newfound time? It's a question that sounds simple but requires genuine thought. "I think that's where we need clarity," came the insistence. "Once we nail that, then people will have a lot fewer reservations about using AI."

Without that clarity, companies are asking people to leap into uncertainty. Use this tool to eliminate your current work, they say, whilst remaining vague about what comes next. It's hardly surprising that people resist.

The CRAFT of Conversation

One of the most practical insights came from John's approach to prompting, which he captured in the acronym CRAFT: Context, Role, Action, Format, and Tone.

It sounds straightforward, but the depth behind it matters. When someone types a single line like "review this legal paper and give me the top five citations" and then complains that AI is rubbish, they've essentially hired a brilliant consultant and given them no information to work with. No context about what they're trying to achieve. No clarity about the role they want the AI to play. No specification of the desired output or the appropriate tone.

"People think it's going to be intelligent, know what I want and just give me what I want immediately," John observed. The reality is far more collaborative. The best results come from treating AI as a coworker who needs proper instruction, not a mind reader who intuits your needs.

The hosts use a similar framework called CRIT (Context, Role, Interview, Task), though they acknowledged John's addition of tone was valuable. The point isn't which acronym you use. The fact is that practical AI usage requires more thought than most people initially realise, and that's become clearer throughout 2025 as the novelty has worn off.

The Learning Curve We Didn't Expect

There's a fascinating dynamic emerging that nobody predicted in those heady early days. When ChatGPT first launched, its accessibility was part of its magic. Anyone could use it. No technical knowledge required. Just type and receive surprisingly coherent responses.

But as the conversation revealed, that's changing. "When this was first released, it was highly accessible. Anyone could use it. But now you get deeper into the wave, and it's like you're using multiple different tools to do multiple different things."

John demonstrated this perfectly when describing his current workflow. For general tasks, he uses ChatGPT 5.2 because it automatically switches between different modes. For conversational output, he prefers Claude. For deep research, he turns to Manus, which can perform tasks that ChatGPT claims to do but doesn't, such as downloading and transcribing YouTube videos. For design presentations, Figma.

Then comes the exciting bit: he stacks them. Create content in Manus, refine it in ChatGPT, and further refine it in Claude. It's a sophisticated workflow that requires understanding the strengths and limitations of each tool, knowing which performs best for specific tasks, and having the patience to move content between platforms.

This is simultaneously powerful and problematic. Powerful because those who invest time in learning these workflows gain significant advantages. Difficult because it creates a steeper learning curve, just as AI was supposed to democratise capability. ChatGPT's deep research feature, which launched in February, remains unknown to most users because it requires an extra click. That single click of friction meant 90 per cent of users never discovered a transformative capability sitting right in front of them.

The Copyright Wake-Up Call

One story from 2025 that crystallised the gap between hype and reality was the Deloitte incident. The consultancy had to pay the Australian government back £250,000 after AI-generated content in a report contained fabricated citations. The recommendations were apparently solid, but some of the supporting citations were invented wholesale.

The group was quick to point out that this wasn't a technology failure. It was a human failure. Someone didn't verify the output. "That's the same risk as if you let an intern do it," came the observation, cutting through the tendency to blame AI for mistakes that ultimately result from inadequate oversight.

This is where transparency becomes crucial. The Royal Institution of Chartered Surveyors released guidelines in 2025 that John found encouraging, particularly regarding accountability and the need to declare when AI is used. "If you're serving clients or you're delivering something to someone, you need to know how you're using AI to create that result," he insisted.

Clients increasingly expect their service providers to use AI. But they also expect transparency about how it's being deployed and, critically, how it affects pricing. If a law firm completes in minutes what used to take hours, should they still charge by the hour? The business model question is forcing uncomfortable conversations across professional services.

What 2026 Holds

As the conversation shifted to predictions, a clear theme emerged: 2026 will be about moving beyond experimentation and into genuine integration. The critical thinking, communication, and fundamentally human skills will matter more, not less. AI will handle the computationally heavy lifting, but someone still needs to ask the right questions, interpret the results, and decide what to do with the time saved.

The Blurring Lines Between Tech and Consulting

John predicts we'll see the lines between consultancies and tech companies blur significantly. Accenture's £390 million acquisition of Consigli wasn't just a one-off deal; it's likely the beginning of a trend. Companies are making expensive bets on capabilities they can't build fast enough themselves. The question is whether they can successfully integrate these acquisitions or whether we'll see costly write-downs in a few years when the expected synergies fail to materialise.

The group debated whether we might see the reverse as well: tech companies acquiring consultancies to gain change management expertise and human-centric implementation skills. It's a fascinating inversion that speaks to a more profound truth about what AI actually requires to succeed.

Governance Gets Real

On the regulatory front, governance will become critical. Not the kind of theoretical governance that looks good in policy documents, but practical governance with someone actually empowered to orchestrate AI usage across an organisation.

John's three M model offers a framework: monitor, measure, and manage. Organisations often excel at the first two but fall on the third. The solution? Chief AI officers or champions who actually have the authority to guide implementation, not just document it.

The AGI Question

The conversation took an intriguing turn when discussing artificial general intelligence. John doesn't think we'll hit AGI in 2026, arguing that current large language models are fundamentally limited by their architecture. They work by understanding patterns, not by actually learning in the way humans do. His prediction: true AGI is still two to three years away, probably arriving around 2028 or 2029.

But a more controversial prediction emerged: humanoid robots doing ordinary tasks by the end of 2026. Not in controlled factory environments, but out in the world, serving people and performing services. John was sceptical, pointing to the vast gap between a robot that can perform one specific task very well and an autonomous humanoid that can navigate the messy complexity of human environments. We'll see who's right.

The Real Work Begins

What made this conversation valuable wasn't the predictions about what technology will emerge next year. It was the honest reckoning with what 2025 actually taught us: that the hard work isn't building the technology, it's integrating it into human systems. It's culture change, not software deployment. It's about answering uncomfortable questions about what people should do with their time, how business models need to evolve, and what we actually value about human work.

The trough of disillusionment that characterised 2025 wasn't a failure. It was necessary. It's when we stopped being dazzled by the magic trick and started asking how the trick actually works, what it costs, and whether it's really helping us solve the problems we need to solve.

2026 will be the year we find out whether organisations have learned those lessons or are still chasing the illusion of intelligence without doing the hard work of integration.

Why Should You Listen to the Full Episode of Project Flux?

Catch the deeper debate about whether reinforcement learning is actually constraining AI's ability to discover truly novel solutions to problems like the Riemann hypothesis or cancer research. 

John and the hosts explore the unsettling question of what happens when AI solves problems using methods we can't understand or verify. They also discuss why embodied AI and continual learning might be the missing pieces for achieving AGI, share war stories about accountants who resist productivity gains because reporting gives them self-worth, and reveal John's sophisticated workflow for stacking multiple AI models to produce better results than any single platform can deliver. 

Plus, hear the fascinating discussion about remote-controlled humanoid robots where real people operate the machines in your home, the future of robo-taxis in the UK following Waymo's 2027 license, and why business models across professional services are heading for a pricing revolution. 

The conversation goes far beyond what we've covered here, diving into the psychology of AI adoption, the emerging risks we haven't yet considered, and practical strategies for citizen innovation. 

This is the conversation about AI's real impact on Project Flux that cuts through the hype and gets to what actually matters: the messy, human work of integrating intelligence into organisations still figuring out what they value. You’ll definitely find this worth your time!