Conscious AI: the same revolution, maturing

I drafted what became the conscious framework between 2018 and 2019. Its purpose was to give some structure to organisations operating in what Klaus Schwab had, a couple of years earlier, labelled the Fourth Industrial Revolution: the broad shift towards hyper-connectivity, distributed knowledge and cyber-physical systems. Seven years on, the world has moved, but the framework has not needed structural revision. AI has turned out to be the technology those underlying forces were always going to produce, and the operating disciplines I described have become more useful as it has matured, not less.

This piece explains why I think that is, what AI actually changes inside each part of the framework, and how I now work with these tools day to day as a CTO.

Why I'm staying with 4IR rather than calling it 5IR

There is a respectable academic argument that we have now entered a Fifth Industrial Revolution, and several camps make it in slightly different ways. The two most prominent are the European Union's “Industry 5.0” framing, which centres on human-centric and sustainability concerns, and a separate (though related) claim that generative AI is a qualitatively distinct break from the connectivity wave Schwab named.

Both have something to them as rhetoric. As analysis, I find neither convincing. The connectivity wave Schwab described in 2016 already implied AI as one of its consequences. Cheap distributed compute, inexpensive inference, very large training datasets and real-time integration were always going to produce something that looked like generative AI. The arrival of broadly capable models in 2023 was not, in my view, the start of a new revolution. It was the point at which the implications of the existing one became impossible to ignore for people who had previously been able to overlook them.

The change in the operating environment is real. AI has dramatically reduced the cost of analysis: a junior analyst with a reasonable laptop can now produce in an afternoon what would have taken a senior consultant about a week in 2020. But this is the connectivity revolution maturing into its obvious consequences, not a separate revolution. The underlying forces are the same. The implications are sharper and the window for response is much shorter.

So I continue to describe the operating environment as 4IR. I would rather be the practitioner who saw it coming and has been refining the response disciplines since than a writer who declares a new revolution every couple of years.

What actually changes

Three shifts in the operating environment materially affect how the framework should be applied.

The cost of analysis has fallen dramatically. What used to be a fortnight of work for a small team (generating ten viable strategy options, modelling a dozen financial scenarios, drafting a competitive analysis) can now be done in an afternoon with competent prompt engineering. The obvious consequence is that analysis is much cheaper. The more important one is that the bottleneck in decision-making has moved. What is now scarce is the ability to look at a confident-sounding output from a model and reliably tell whether it is also correct.

New failure modes have appeared, none of which existed in any meaningful form before 2023. Plausible-sounding answers that turn out to be hollow when pressed. Teams accepting whatever the first model said as the default, in the way they used to defer to the most senior voice in the room. Well-written wrongness that survives internal review because no-one wants to admit they didn't quite follow it. Visible AI tooling deployed to signal modernity rather than to do useful work. None of these are technology problems. They are people-and-process problems wearing AI costumes.

The genuinely human parts of the work have become more important, not less. Spotting when a persuasively phrased answer is wrong. Recognising when the question itself has been framed badly. Operating in domains so novel that no training data could cover them. Building the trust needed to land an unpopular recommendation. Automation has touched none of these, and the rest of an organisation's capability now depends on them more heavily than before.

Each framework, briefly updated

The table is the short version. The sections after it pick out the updates that matter most.

The change most teams have not yet absorbed

If I had to name the one capability that has become more important than any other in the last couple of years, it is the ability to tell a confident answer from a correct one. Before about 2023 this was useful but secondary. Fluent, well-structured, confidently asserted output almost always came from someone who knew what they were talking about, so fluency was a weak but fairly reliable signal of competence.

AI has broken that signal. Modern models produce fluent, structured output on any topic, including topics on which they are wrong in subtle and consequential ways. Telling the two apart is now a primary leadership skill, not a peripheral one, and it is not the same skill as technical expertise.

In practice, teams need explicit verification habits that go well beyond a casual “does this look reasonable?”. The more productive question is: “what would I expect to see if this answer were quietly wrong, and is any of that present?” Senior leaders need to be visibly comfortable saying “I don't fully understand this, give me half an hour to work through it”, in a culture where AI makes pretending to understand the easiest default it has ever been. Junior staff need deliberate coaching out of treating model output as authoritative before they have the knowledge to push back on it.

The smelting stage in Idea Mining, Phase 4 in Decision Mining and the “Does it Matter?” check in Conscious Agility are where this judgement is exercised. The framework named them as the human-only steps from the start. They are now also where the value of the whole pipeline concentrates.

How I now work with AI day to day

Over the last couple of years a handful of practical habits have become my defaults.

AI agents do the scouting. Named humans decide what propagates. Agents are good at watching high-volume sources: news feeds, regulatory filings, technical literature, code repositories. They are not good at judging which of the things they find matter. So each agent reports to a named person with the authority to decide what enters the wider organisational conversation. The agent extends the sensing layer. It does not replace any of the human nodes within it.

AI in the mining stage of Idea Mining, never in the smelting stage. When I run an Idea Mining cycle, AI tools are welcome to help generate options at the front end. They are not welcome in the room when the team decides which options to refine. The boundary is deliberate and firm: humans don't add new ore during smelting, and AI doesn't enter the smelting room at all. I keep the line crisp because smelting is exactly the work that needs the judgement described in the previous section.

AI-generated and human-generated options on the same page. In Phase 3 of Decision Mining I now run both side by side, against the same scoring matrix, with everything visible to the room. The interesting cases are usually where the AI options cluster around a shape the human options don't. Sometimes the model has noticed something the team missed. Sometimes it is biased towards patterns from its training data, which in practice means patterns from the past. Either way, the comparison is useful.

Every AI investment proposal comes with a people budget. When a vendor pitches AI tooling, I expect the proposal to include an explicit budget for the human capability the tool needs to operate well: verification, exception handling and judgement on outputs. My rough working ratio is around 1:3. For every pound of AI procurement I tend to allocate something like three pounds of investment in the people who will run it, over a comparable period. The ratio is sometimes lower. It is essentially never zero. A procurement meeting that ends without anyone asking “what are we spending on the people who will operate this?” is usually a meeting where the project is already quietly losing money.

Verification capacity is finite. Spend it deliberately. A wrong answer from an AI tool burns more team trust than a wrong answer from a colleague. After one or two embarrassments a team will often over-correct into distrusting everything the tool says, and the tool stops being useful. So I treat verification capacity as a budget, allocated where the cost of being wrong is highest, rather than spent evenly on every output because the model sounds enthusiastic.

The same disciplines, more useful

The conscious framework has always been about people working together under uncertainty. AI is the most powerful technology those people now have. The disciplines the framework names (structured collaboration, deliberate sensing, conscious decision-making, and the habit of asking what kind of problem you are looking at before reaching for a tool) are the same disciplines that decide whether an organisation gets real value from AI or simply produces faster confusion.

The years from 2019 to 2023 were the calibration period for these ideas. The years since have been the test. I still use the framework routinely, not as a frozen 2019 artefact but as a working set of disciplines that AI has, on the whole, made more useful rather than less.

Four things worth taking seriously

For boards: when an AI investment comes for approval, ask what is being spent on the people who will operate it. A proposal with no people budget is already losing money.

For engineering leaders: build explicit verification habits. Ask what you would expect to see if an answer were quietly wrong, and treat verification capacity as a budget to spend where being wrong costs most.

For anyone running idea or decision processes: let AI help generate options, but keep it out of the judgement steps. Smelting, consensus and the “Does it Matter?” check stay human.

For senior leaders: be seen saying “I don't fully understand this yet”. In a culture where AI makes pretending easy, that is what gives everyone else permission to check.

I would be interested to hear where in your organisation fluent AI output has been mistaken for correct output, and what caught it.

Related: Conscious Collaboration · Cynefin · Idea Mining · Decision Mining · Conscious Agility · Lean is dead · Don’t invest in collab tech · all writing