"The greatest danger in times of turbulence is not the turbulence—it is to act with yesterday's logic." — Peter Drucker
What LLMs reveal about how organisations actually work.
Companies are pouring billions into AI. Yet in many organisations the main output so far is more PowerPoint, more summaries, and more dashboards. AI is accelerating something — just not always the right thing.
Your company is optimised for human limits. That's the problem.
Organisations are optimised for human cognitive limits. How we process information, especially information that we don't like. How we simplify complexity, justify decisions through metrics, reports, categorisation, and the boxes of the organisation charts that surround us.
For all our working life, this architecture (cognitive + organisational) defined efficiency, coordination and control.
This made sense when human intelligence was the only general-purpose system that was available to us to interpret reality and produce organisational knowledge. It doesn't make sense anymore.
This article will end with 7 actions for leaders who want to leverage AI for true return on investment. But they are jump-starters for generating real return on investment. To understand why, read on.
LLMs showed us the scalable power of tokenised intelligence.
You already know the very basics of how LLMs work. These systems process text into individual "tokens" like parts of words, numbers, and categories and they then predict and recombine those pieces based on statistical patterns learned from vast amounts of data. All other things being equal, and as far as we know now, the larger the data set used for training, and the more computational power applied, the more capable they become.
Forget about the philosophical arguments about artificial general intelligence (AGI) or super-intelligence. What LLMs taught us, once they scaled, is that intelligence itself can be classified into tokenised and non-tokenised intelligence.
We can, therefore, refer to human "tokenised intelligence" as a way of processing the world in the same manner. While tokenised intelligence usually refers to machine intelligence, I believe that it is helpful — indeed, necessary — to refer to tokenised intelligence in humans in order to best understand where and how to apply AI to your organisation. And so, in this article I am defining two forms of human intelligence.
Tokenised intelligence.
The way that humans handle complexity, by turning the world into manageable pieces: numbers, categories, metrics, reports, models, slide decks. Human tokenization makes work measurable and comparable, and easy to communicate across an organisation. Tokenization is how companies scale decisions and control operations. Its strength is clarity and efficiency. Its danger is that the simplified picture can quietly replace the underlying reality.
That's why AI hallucinates. Bad news: it's why some of us are finding that lately humans hallucinate just a little bit more than AI.
Non-tokenised intelligence.
This is the way that humans deal with situations that do not fit neatly into metrics or categories. This mode of intelligence is about sensing ambiguity, noticing the weak signals in the noise, judging context, reframing problems. It is how we identify AI slop, a social situation that's off, how we emotionally feel about expert judgment. It's when a CEO ignores the data and goes with her gut — and wins. It's both what the era of industrialization tried to get rid of by moving to "management by numbers," and why we feel the need to define "human in the loop" policies.
The bad news.
Companies have spent the last half-century optimising organisations for tokenised work. A quarter of the way through this century, machines are beginning to excel at exactly that kind of work: tokenized process work. Why is that bad news? Because the operating assumptions of organisations are now colliding with a new cognitive landscape.
It's just like the Wizard of Oz. As the machines expose the mechanics behind the curtain of corporate process, we can see that decision making, coordination and knowledge management are increasingly mediated through tokenised representations whose meaning is statistically mediated through a limited vocabulary: that could be the corporate jargon, or the systematic cascading of standardised and shared objectives derived from tokenised mission and value statements. Or it could be all controlled vocabularies, all performance reporting templates, all standard operating procedures, or even the chart of accounts. It focuses on outputs, artefacts of processes we call records, and it imposes a burden on your organisation to develop even more tokenised processing to manage the tokenised processing: what we call quality management, records management, and many more business management practices that emerged from industrialisation.
The good news.
Tokenisation is fast, compressive, shareable, and auditable because tokens leave persistent artifacts. That is the novel insight from scaled LLMs. Things can be inspected, verified, challenged and defended after the fact. The good news is that with a good grounding and training on your own jargon and institutional simplifications of reality, LLMs are now very very good at taking over that work. They can even be used to demonstrate the absurdity of some of our corporate burden, in the same way as the magic is gone when the magician reveals her trick. Sleight of hand only works if you don't see it.
Here is an example that has become one of the most normalised rituals of organisational life: the timesheet. An employee performs work. Then she performs additional work to describe the work. Then managers perform further work to review, validate, reconcile, audit, and store the descriptions of the work. Entire systems exist to ensure the representations are accurate, compliant, and properly classified. It's just like Dr. Seuss. It is the bee watcher watching the bee watcher watching the bees. You see, timesheets are not some quirky management fad of the last decade, but just a natural consequence of tokenising work.
Who loves tokens?
This should not prevent us from applying machines to tokenised process work.
Firstly, LLMs naturally "colonise" tokenised work because that work already exists in a form that machines can process (if you let them.) Reporting, classification, reconciliation, template summaries — these are all native terrain for LLMs.
I have been observing the adoption of LLMs in many organisations and here is what I observe. When work is "automated" through LLMs, most of the productivity comes from speeding up the creation of, the manipulation of, and the distribution of artefacts of work instead of the underlying work itself. More images end up in PowerPoint. More reports get produced, more summaries are written, more documents get classified, more data gets tagged, more (excessively polite) emails circulate. It feels like productivity because cycle time is reduced for the production of work about work. But this is also exactly why so few companies have produced a return on investment. It produces what Gartner called "return on staff." And mainly only for employees doing tokenised work about work. It's circular and therefore does not produce a return on investment.
What is accelerating in these cases is the representational metabolism of the organisation. The only thing increasing is the rate at which the organisation was using its people to produce, transform and circulate tokens. More kindly, the organisation does become faster at describing, documenting, analysing and reformatting reality (but only as long as it is a statistically predictable reality.)
Secondly, once machines handle tokens cheaply, a strange mirror appears. Activities that seemed necessary (administration that was mission-critical alongside the core business) begin to reveal their true function. Some of these are necessary. Society demands them. Things like coordination, accountability, legal traceability. That's valid tokenised work. But some are just residue from history. Representations of reality that were created mainly in order to stabilise other representations of reality, none of which actually represent the complexity of reality outside the corporate epistemics.
LLMs do not just automate this layer; they expose it.
The main risk is the use of AI to reinforce the very overhead that constrains them. The main opportunity is to use AI to detect where token production and circulation no longer improve sensing or decision making. Where control is extended in ways that interfere with strategy.
In simple terms, AI forces a distinction that most organisations cannot make explicitly. Some tokens reduce uncertainty about reality. Some tokens help us understand reality better. After more than three decades applying methodology and process management from the Taylor-Deming-System school toward real world applications, I am very familiar with the success rate of organisations applying the systems-school. Institutional habit constrains human tokenised intelligence so as to cloud the vision of reality.
Organisations (wrongly) privilege tokenisation.
Non-tokenised intelligence is the mode where humans engage situations before they stabilise into categories, sensing ambiguity, noticing weak signals, reframing problems, judging what does not yet fit the model. "Thinking outside the box" is about intuition, embodied judgement, aesthetics, creativity, and internal nervous system signalling.
It is easy to understand why most organisations prefer the goal oriented efficiency and fast legibility for process work in their organisation. Human tokenisation is preferred because it travels cleanly across function and hierarchy. It can be measured, compared, stored, audited, rewarded, and justified.
And this has created a very powerful evolutionary pressure within corporate cultures. Representational clarity often becomes synonymous with competence. Complexity, when compressed into portable symbolic forms becomes a markdown language of its own, and fluency in this language is rewarded. Over time, the shared vocabulary begins to feel more stable, and therefore more real, than the underlying phenomena it was intended to manage.
Too often, however, they are self-referential justifications, leading only to an increase in the efficiency of that which is fundamentally non-effective. In AI, we call this overfit. More precisely, when an organisation standardises and simplifies its operating models, it risks overfitting to the internal measurement system rather than to the external reality. Metrics become the targets. Targets become the behaviours. Behaviours reshape internal incentives. Strategy then becomes interpreted through what is measurable.
Representations are introduced to help organisations navigate reality. Gradually, adherence to the representation replaces engagement with the underlying reality. Compliance becomes a proxy for understanding. Nothing irrational is occurring. Tokens are easier to manipulate than the world itself. They change more slowly. They are socially validated. Attention therefore migrates toward what the system can record and away from what must be directly sensed.
I am well aware that running an organisation requires compromises and simplification. All large-scale human systems depend on stabilised representations. Science relies on models, definitions, statistical thresholds, and taxonomies. Business relies on financial abstractions, segmentation schemes, performance metrics. Government operates through classifications, eligibility criteria, compliance regimes, and reporting structures. These are not bureaucratic quirks; they are the infrastructure that makes collective cognition possible at scale.
Tension emerges when environmental volatility accelerates. Representations freeze assumptions about the world. These representations inevitably age, and in the 21st century they do so increasingly quickly. But people are stuck in their weights. Signals shift before categories shift and yet these anomalies are normalised. Surprises are explained away. The organisation becomes progressively more fluent in its internal language while progressively less sensitive to external change.
The danger is not immediate error but epistemic drift. Reality evolves continuously, while representations update discretely and often politically. Inside the organisation everything appears coherent: dashboards align, reports reconcile, indicators improve.
Outside, anomalies accumulate, surprises increase, decisions lose calibration. The system becomes wrong in a stable, internally consistent way.
But let's now take stock of what AI's tokenised cognition allows us to shift in the 21st century.
Re-establishing human role and human agency.
So, let's make clear that humans are not displaced because intelligence becomes tokenised. But it should be clear that machines are, since November 2025, better at playing the roles and tasks associated with tokenised intelligence. They do so with high efficiency within representational systems (systems that use symbols and structures to simplify the management of reality.)
Since the November 2025 releases of general purpose LLMs from all the main frontier labs, and for published technical reasons beyond the scope of this article, my observations point to machine superiority in all of the main categories of quality that worry us: hallucination, drift, and bias. You may disagree with the timeline, but the monthly updates now clearly tell us that we are at the pivot point in tokenised-intelligence superiority.
Our confusion about its superiority has come from the mix of tokenised-intelligence and non-tokenised intelligence that is active in our corporate discourse. That is because narrative and text output from humans betray the layer of non-tokenised thinking where we excel.
AI can provide true ROI
If we want to leverage the power of AI, today, for a real return on investment, we need to rethink where and how to apply human cognition to the enterprise. It is not about what roles, job profiles or competencies are to be "outsourced" to AI. It is about digging deep into the task-layer of an organisation and decomposing those tasks into the tokenised and non-tokenised components. Because as token production becomes abundant, scarcity shifts.
Human contribution must shift toward what resists clean tokenisation. This is all about judgment under ambiguity, problem framing, anomaly detection, contextual interpretation, ethical arbitration, deciding when representations no longer track reality.
Since more than a decade, all organisations have been calling for innovation. Not just for unicorns, or technology or marketing innovation dependent firms. Agility, innovation, and speed are now the mantra of every organisation, including the most conservative if we listen to the board room. No human organisation is able to ignore the imperatives of the 21st century.
And as humans are pressured out of token-dominated roles because machines are simply better and more efficient within representational systems, then the strategic risk is not automation itself, but misallocation: continuing to optimise humans for the cognitive mode machines increasingly dominate.
TL;DR: 7 CEO Actions You Can Start Monday
No CEO, governance, or senior leader can dismantle representational infrastructures overnight. Standards, reporting systems, performance frameworks, compliance mechanisms carry institutional momentum and contain, at their core, a valid purpose.
You want to get ROI from AI investments. You see that we are at the start of a revolution that cannot be missed. The secret is that the ROI is not found by investing in LLMs to automate your tokenised work, but rather in leveraging the scale of tokenised machine thinking to let people do what a machine can never do: truly feel other humans and think out of the box.
Rebalancing token work in our organisations is the key area for action to leverage this new technology revolution starting up. Real ROI from AI requires putting humans into non-tokenised work. They don't necessarily want to go there. And this conflicts with the rigid culture of symbolic representation that we built in the 20th century.
And so, my practical recommendations are all about taking the first steps toward a fundamental change in how organisations operate. You need to start removing and calibrating obsolete processes that depend on human tokenised intelligence and putting humans into uncomfortable context where they outperform machines in non-tokenised intelligence. That messy intuition. The innovation that comes from boredom. That sixth sense from the nervous system that signals shifts in the underlying reality.
From experience, I'm not advising massive projects, but small agile steps that can be implemented right away to start getting real return on investment from AI. They work because they are small enough to not receive too much backlash, but important enough that they start to retrain the human minds within the organisation. And you will note two main themes to get ROI from AI:
1. If the process is wrong, AI automation will likely produce more time for coffee, but no improvement to the top and bottom line of your financial statement. LEAN first, AI second.
2. If people are not forced out of tokenised thinking, they will not be able to do what humans were designed for: collaboration, creativity, intuition, and innovation.
1. Cancel one internal report for 30 days. Pick a recurring internal report (weekly status report, dashboard, committee briefing). Stop producing it for 30 days. Ask managers to record what decisions actually suffered. If none did, the report was token overhead. AI Corollary: Until you eliminate all such reporting, do not approve any AI use cases to automatically produce reporting and dashboards. Improve AI portfolio ROI by avoiding the investment in the first place.
2. Run a "token audit" on one business process. Take a process like hiring, procurement, customer onboarding, or compliance review. Count every document, form, approval, and report generated along the way. Then ask one question: Which of these tokens actually change a decision? Anything that doesn't can be removed. AI Corollary: Do not deploy AI to automate a process until you first identify which documents and approvals affect decisions. Otherwise, AI will simply accelerate bureaucratic overhead. Once you know which tokens drive decisions, invest in AI models that propose decisions. Next, transition the management team that must decide from the AI proposals. But use no PowerPoints, no briefings, no tokens in the discussion. This is profoundly uncomfortable and can lead to conflict. But it is how AI ROI is obtained, by forcing humans to do the "out-of-the-box" thinking that AI is incapable of doing by its design.
3. Send managers to the territory once a quarter. Require key managers leading this new wave of non-tokenised intelligence to spend one day per quarter immersed in the real environment their reports describe: customers, frontline operations, suppliers, regulators. No slides. No dashboards. AI Corollary: Update model weights, or system prompts to align AI models to the observed reality. Then leverage the humans, in small innovation teams, to narrate their stories, in their own language to decision makers. No PowerPoints, no documents, just human story-telling.
4. Add a mandatory "What are we missing?" agenda item to leadership meetings. Every executive meeting should include a short discussion of signals that do not appear in dashboards: unusual customer behaviour, operational anomalies, rumours and complaints, unexpected industry moves. These signals often precede strategic shifts. AI Corollary: Do not rely exclusively on AI-generated dashboards for executive decisions. If leadership cannot discuss signals outside the data, AI will reinforce blind spots rather than improve insight. When in doubt, table the doubt for non-tokenised thinking. Flip it into innovation thinking.
5. Make "challenge the dashboard" a formal role. In every major review meeting, assign one participant the explicit role of questioning the metrics. Their responsibility is to ask: What would have to be true for these numbers to be misleading? AI Corollary: Every AI-generated analysis must have a designated human responsible for challenging its assumptions. AI should propose conclusions; humans must test whether the model still reflects reality.
6. Reward people who delete processes. Most organisations reward people who create systems. Almost none reward people who remove them. Introduce a visible metric for eliminating unnecessary internal processes. AI Corollary: Do not approve AI projects that automate a process until someone has first asked whether the process should exist at all.
7. Create a task-level map of your organisation. Goodness, there must be some practical sense to all of those standard operating procedures and operations manuals. Form a taskforce to inventory all enterprise tasks and to classify them. Some activities are structurally tokenizable: repeatable, high-volume, rule-bound, audit-sensitive. These are natural entry points for LLMs and agentic systems. Other activities depend on reality-first cognition: strategy, crisis interpretation, early-stage problem definition, ethics, weak-signal detection. Over-automation here introduces brittleness rather than efficiency. Then use this task map classification to "open-source" your organisational evolution. You don't want this to become a consultancy project. Start with one department pilot. And learn from it. Direct ROI on AI investments will only come from tasks that are cleanly tokenised and valuable. If you don't separate the two modes of thinking, the ROI will never come.
Conclusion
The challenge is not abandoning representations. No modern organisation can operate without dashboards, reports, models, and classifications. These are the mechanisms that allow complex institutions to coordinate decisions at scale.
But AI has changed the balance.
Machines are now exceptionally good at producing, transforming, and circulating tokens. In many organisations, what AI is accelerating is not decision quality but the representational metabolism of the organisation — the rate at which reports, dashboards, summaries, and classifications are created and exchanged.
That can feel like productivity. But faster tokens do not automatically produce better understanding.
The leadership challenge in the age of AI is therefore not simply automation. It is deciding which tokens still help the organisation understand reality, and which merely help it document activity.
As machines take over tokenised work, the human advantage shifts elsewhere: sensing weak signals, recognising anomalies, reframing problems, and deciding when the model no longer matches the world it was meant to describe.
Organisations that win with AI will not be those that automate the most processes. They will be those that are disciplined enough to remove unnecessary ones and courageous enough to rely on human judgment where models inevitably fail.
Because in the end, AI will scale tokens. But humans must scale judgement.
That's what they mean by innovation. And that's how you get ROI from AI.