Posts

The goal of AI should be to make the next problem easier, not the same problem more expensive.

That idea became the foundation of my new book:
Reusable Intelligence
Beyond Inference: Building AI Systems That Learn When Not to Reason

AI is becoming remarkably good at reasoning.

But as AI moves into agents, enterprise workflows and everyday operations, I believe we need to ask a different question:
Why should a system keep paying to rediscover what it already knows?

Humans do not reason from scratch every time.
We:
Think → Learn → Remember → Reuse → Notice change → Think again

AI systems should be able to mature in the same way.

In the book, I introduce several ideas around this:

• Learning Dividend: experience should make future work easier
• Reasoning Budget: spend inference where uncertainty actually remains
• Reuse Boundary: know not only what can be reused, but where reuse stops
• Selective Reasoning: Known → Reuse. Unknown → Reason. Uncertain → Escalate.

The bigger idea is simple:

The first era of AI scale was about making intelligence available on demand.
The next era will be about making intelligence accumulate.

Because the most advanced AI system may not be the one that reasons the most.
It may be the one that knows when not to reason.

Reusable Intelligence is now available on Amazon.

If you are building AI agents, enterprise AI platforms, or thinking about the economics of AI at scale, I hope this gives you a different lens for what comes next.

My LinkedIn feed is full of one question: “Has AGI arrived?

In the latest edition of Technology Bytes, I ask a different one:
If AI is now intelligent enough for us to debate AGI, can it solve a simpler problem for itself?
Can AI become cheaper to run, lower in compute and energy, and more reusable?
Can it recognize when a problem has already been solved and avoid paying to reason through it again?

Because the next test of intelligence may not be how much intelligence a system can use.
It may be how intelligently it decides when not to use more.

Read the latest edition of Technology Bytes:
AGI Has Arrived. Now Give It a Simpler Problem.

I’m looking forward to speaking at Green IO London 2026 on:

Green Agentic AI: Optimizing Cost, Energy and Carbon

Agentic AI can consume orders of magnitude more compute than a simple chat interaction.

As agents reason, call tools, retrieve context, retry tasks, maintain memory and coordinate workflows, efficiency becomes a design requirement, not an afterthought.

And this is not only an environmental issue.

More efficient AI also means lower operational cost, better resource utilization and more sustainable economics at scale.

In this session, I’ll explore how we can measure and systematically reduce the impact of agentic AI, including:

• Cost, energy, carbon and water per task
• SCI for AI and emerging SEI and SWI metrics
• A six-stage agent lifecycle for identifying optimization opportunities
• Design levers across models, context, tools, retries, memory and scheduling
• How GSF standards can provide a consistent measurement foundation
• A practical Monday-morning checklist teams can start applying immediately

The goal is not simply to build more capable agents.

It is to build agents that deliver more intelligence with less resource consumption and lower operational cost.

📍 Green IO London 2026
📅 September 30th – October 1st, 2026
📍 Convene Sancroft, St. Paul’s, London

Hope to see many of you there.

https://lnkd.in/dWsK9Nq4

#GreenAI #AgenticAI #GreenSoftware #SustainableAI #AI #SoftwareSustainability #SCI Green Software Foundation Green IO 🎙️ Gaël DUEZ

100 episodes. One evolving question: what does the future of intelligent systems really look like?

When I started Agentic AI: The Future of Intelligent Systems, the conversation around AI was very different.

We were focused on models, prompts and copilots.
Then came agents.

Now the discussion is moving again, toward something much bigger:
enterprise intelligence.

That is why Episode 100 is titled:
The Next Era of AI: From Agents to Enterprise Intelligence

And if you have been wondering when we will achieve AGI, I think there is another question worth asking first.
AGI may still be a long way ahead.

Before we get there, most enterprises have not yet fully learned how to use the intelligence we already have.

We are still figuring out how to connect agents across the enterprise.

• How to give them the right context.
• How to reuse intelligence instead of reasoning from scratch every time.
• How to govern memory, tools, data and decisions.
• How to move from impressive demonstrations to reliable business execution.

So perhaps the next major milestone is not AGI.
It is enterprise intelligence.

Because the next chapter will not be defined by how many agents an enterprise can deploy.

It will be defined by how effectively intelligence can be shared, reused, governed and connected to real business outcomes.

Reaching 100 episodes is a very special milestone for me.

A huge thank you to everyone who has listened, followed, shared an episode, sent feedback, challenged an idea, or simply spent a few minutes thinking about the future of AI with me.

Every listen matters. Every conversation adds something.
The first 100 episodes explored the rise of Agentic AI.
The next 100 will explore what comes after.

From models → agents → enterprise intelligence.
🎙️ You can find the Spotify podcast link below.

If you have been enjoying the conversations, please follow the podcast, share an episode with someone interested in the future of AI, and help more people discover the series.

Thank you for being part of the journey.
The next chapter starts now.

AI leadership is not about chasing the next model. It is about building enterprise intelligence that survives the next model.

A new AI model is released. It is smarter, faster and cheaper. Does your strategy change?

AGI is “near.” Does your strategy change again?

There is a vast difference between using a model and building enterprise capability on top of it.

A new model may improve reasoning, coding, multimodality or cost. But it will not suddenly understand your business, redesign your processes, integrate with your systems, create your governance, or build differentiated intelligence for your enterprise.

That intelligence has to be engineered.

As I wrote in an earlier Technology Bytes article, this is also the difference between an AI-aware leader and an AI-fluent leader.

An AI-aware leader follows what models can do.
An AI-fluent leader understands what those capabilities mean for the business, what needs
to be built around them, and what should remain stable even as the underlying models change.

So after every major model release, the question should not be:

Do we need a new AI strategy?

It should be:

Does this model materially improve a business capability we have already decided matters?

Models will keep changing.

Enterprise intelligence is the asset. And to make that intelligence durable, enterprises will need Reusable Intelligence.

I will explore that idea in an upcoming Technology Bytes.

The internet taught companies to design information for humans.
AI may force companies to redesign information for machines.

For decades, the content pipeline was simple.
A company created information.

A human found it, read it, interpreted it and acted on it.

That shaped everything:
Websites.
Reports.
Dashboards.
Documentation.
Product pages.
Knowledge bases.

But the pipeline is changing.

AI is now used to:
Ideate → Design → Code → Develop → Test → Deploy → Document

Then the product reaches the customer.
And increasingly, the customer does not consume all of that information directly.
Their AI agent does.

The agent:
Searches → Reads → Summarizes → Compares → Recommends → Plans → Acts

Which creates a strange new loop:
AI creates the information.
Another AI consumes the information.
Yet the artifact in the middle is still designed for a human.

AI writes a 30-page report so another AI can summarize it.
AI creates documentation so another agent can extract the relevant steps.
AI builds dashboards so another AI can interpret the underlying metrics.
AI generates product pages so a customer agent can compare the products.

At some point, we should ask:
Why are we still generating the intermediate artifact at all?

The next content architecture may look very different:
Intent → Structured Intelligence → Agent → Action → Outcome → Learning

The unit of information may no longer be the page, report or dashboard.

It may be a machine-consumable intelligence object containing:
Facts.
Claims.
Evidence.
Constraints.
Capabilities.
Actions.
Dependencies.
Freshness.
Something an agent can directly discover, reason over and act upon.

This changes more than content creation.
A website could become an agent interface.
A report could become a decision object.
Documentation could become executable knowledge.
A dashboard could become an intervention layer, surfaced only when something requires attention.

And SEO may eventually be joined by something very different:
Agent Experience Optimization.

Not simply:
Can people find our information?
But:
Can an AI agent correctly discover, understand, trust and act on our intelligence?

We spent the last 30 years building the internet for human attention.
We may spend the next decade rebuilding enterprise information for machine consumption.

Is your leadership team the biggest risk to your AI strategy?

Most organizations are investing in AI.
But if leadership understanding is not evolving as fast as the technology, strategy starts falling behind.

That is the Leadership Adoption Gap.

In the latest Technology Bytes:
The Leadership Adoption Gap: Why AI Strategies Fail at the Top

Are your leaders AI-aware, or genuinely AI-fluent?

What if AI, and the data center behind it, had a voice?

What would it say about the jobs we fear losing?
The energy and water it consumes?
The infrastructure we keep building?

And our relentless demand for more intelligence, more speed, and more compute?

I imagined a conversation between a human and a robot.
🎥 Watch till the end.

#AI #ArtificialIntelligence #SustainableAI #ResponsibleAI #DataCenters Green Software Foundation

Building a powerful AI model is one thing. Making an enterprise depend on it is another.

AI research is moving fast, but enterprise delivery is not keeping pace. A model can look exceptional in a benchmark and still be difficult to deploy reliably across real business processes, systems, controls and economics.

That is the AI Delivery Gap.

In the latest edition of Technology Bytes, I explore why better models do not automatically create better AI systems, and why the next competitive advantage may come from how well enterprises turn intelligence into delivery.

Research asks whether AI can do it. Delivery asks whether the enterprise can depend on it.

Read: The AI Delivery Gap: Why Better Models Do Not Automatically Create Better AI Systems

The token economics model is not scalable. AI economics needs to change.

Today, every time AI reasons, the meter runs.
That is acceptable when the problem is new, uncertain or exceptional.

But once the enterprise has already solved the problem, why should it keep paying to reason through the same logic again?

That is the idea behind my latest Technology Bytes:
The Enterprise Intelligence Compiler: From Probabilistic Reasoning to Deterministic Scale

Use AI where intelligence is actually needed: for the novel, unknown and exception paths.

Then take validated reasoning and turn it into executable business artifacts such as rules, workflows, policies, code, decision tables and controls.

That changes the economics:
Variable inference cost → Predictable execution cost
Probabilistic reasoning → Deterministic execution
Repeated inference → Reusable intelligence
Model behavior → Governed business logic
Scaling tokens → Scaling transactions

The goal is not to remove AI.
It is to stop using AI for problems the enterprise has already solved.

Reason until you know. Then execute what you know at scale.
If AI is going to scale across the enterprise, we need to move beyond an economics model where value is tied to how many tokens we consume.

Most enterprises do not need AI to think every time the business runs.

Yet that is increasingly how we are designing enterprise AI.

A large part of business execution may begin with discovery, ambiguity and reasoning. But once the right approach is understood and validated, much of it becomes repetitive: approve, validate, route, calculate, reconcile, provision, enforce policy. The inputs may change, but the underlying logic often does not.

Still, we put an LLM in the runtime path and ask it to reason through the same class of decisions again and again.

And every time it thinks, the meter runs.
More tokens. More inference. More latency. More variability. More cost.

This is where token economics starts to break down at enterprise scale.

Tokens measure what AI consumed. They do not measure what the business accomplished.

Caching helps, but caching only makes repeated inference cheaper.
Codification means AI does not need to solve the same known problem again.

A better model is:
Discover with AI → Validate → Codify → Execute deterministically → Escalate exceptions/new situations to AI → Learn → Codify again

Use AI when something is unknown, ambiguous or genuinely new. Once that reasoning is validated, turn it into an executable artifact: a rule, workflow, API, code, state machine, or a compact capability that can run like a library on any system.

Then execute it repeatedly without invoking an LLM every time.

When a genuinely new situation appears, escalate it back to AI, solve it, validate the outcome, and codify what was learned.

Over time, the deterministic surface of the enterprise grows, while AI focuses on what is actually new.

This changes the economics fundamentally.

Instead of paying for intelligence every time knowledge is used, pay for intelligence when intelligence is actually required.

If AI is going to scale across enterprises, we need to move beyond an architecture where tokens become the unit of business execution.

The goal is not to minimize tokens.
The goal is to maximize business value while minimizing the intelligence required at runtime.

Reason where necessary. Codify what is learned. Execute what is known. Escalate what is new.

Why are organizations paying AI to relearn what they already know?

Much of enterprise AI today follows the same pattern:
A transaction arrives.
An LLM reasons.
An outcome is produced.

Then the next similar transaction arrives, and the LLM reasons all over again.
That makes sense when the problem is genuinely new.

But once the organization has learned the answer, why should the reasoning remain in the execution path?

The term I see emerging for this next stage is Codified AI.

Codified AI converts validated AI reasoning into deterministic, executable artifacts that can be reused without invoking an LLM again.

The progression looks like this:
Novel → Reason → Learn → Bound → Codify → Execute

When something is unknown, use the best intelligence available to discover the answer.
As the problem becomes understood, reduce the intelligence required.
And when the logic becomes explicit and independently testable, convert it into workflows, rules, policies, decision tables, code and tests.
Then run it deterministically.

Zero LLM calls in the validated execution path.
This is the broader idea behind the Intelligence Curve:

Use maximum intelligence to discover. Use minimum intelligence to execute.

The shift is bigger than token savings.

It is a move from consuming intelligence repeatedly to converting intelligence into reusable organizational capital.

One reasoning session can become a workflow executed millions of times.
One policy analysis can become an approval engine.
One design session can become an operational control.

And even if inference becomes nearly free, determinism does not arrive with the discount.
Reproducibility, auditability, bounded behavior, predictable execution and governance still matter.

Perhaps AI maturity should therefore not be measured by how many agents, model calls or tokens an organization consumes.

It should be measured by how effectively the organization converts what AI learns into reusable execution.

I explore this in my new white paper:
The Intelligence Curve: From Frontier Reasoning to Codified Execution

White paper attached.

Excited to announce my new book, AI Economics: Building and Scaling Intelligence.

AI is getting cheaper, more capable and increasingly autonomous.

But there is a question I believe matters far more:

What value are we actually creating with all this intelligence?

Today, much of AI economics is measured at the point of consumption: tokens, API calls, GPU hours, model costs and infrastructure spend.

These measures tell us what AI consumed.

They don’t tell us what AI accomplished.

That is the starting point for AI Economics.

I believe the economic unit of AI needs to move closer to the reason we deploy AI:

Compute → Output → Tasks → Successful Tasks → Outcomes → Value

And as we move from models and copilots toward agents and reusable skills, differentiation moves upward:

Model → Capability → Skill → Domain → Industry → Enterprise

The ability to encode domain knowledge, industry expertise and enterprise context into reusable, executable intelligence could become an important source of competitive advantage.

But AI economics cannot stop at financial cost.

AI has a physical economy. Intelligence ultimately depends on infrastructure, hardware, energy, water and materials.

We therefore need holistic AI unit economics spanning finance, infrastructure and physical sustainability.

Not simply what AI costs, but what outcome it creates, what value that outcome delivers, and what resources were required to deliver it.

This leads us toward measures such as cost per outcome, energy per outcome, carbon per outcome and value per unit of resource, alongside absolute consumption.

There is also one industry where I believe this shift will be particularly profound: IT services.

For decades, much of the industry has scaled through:

People × Utilisation × Rate = Revenue

But what happens when expertise becomes executable?

When agents perform work and skills encode reusable expertise, the delivery model starts moving toward:

People + Agents + Skills + Platforms + Domain Expertise → Outcomes

This has implications for billable hours, offshore delivery, talent pyramids, managed services, pricing, margins, intellectual property and outcome-based contracts.

I dedicate a part of the book to this transition and what an AI-native services firm could look like.

AI isn’t simply changing how IT services firms deliver work. It could change what they sell, how they price it, how they scale expertise, and where they capture value.

AI Economics brings these ideas together by connecting intelligence, skills, outcomes, resources and value into one economic system.

As intelligence becomes increasingly abundant, the question becomes:

What should we use intelligence for, and what value can we create with it?

📘 AI Economics: Building and Scaling Intelligence is now available in paperback on Amazon worldwide. Link in the comments section.

Do you think AI is actually reducing work?

It is certainly reducing the cost of producing work. Analysis that took days can now be generated in minutes. Code appears in seconds. Ten scenarios can be created where previously there might have been two.

But something interesting happens next.

Organizations rarely take that productivity gain and simply stop. They ask for more analysis, more variants, more scenarios, more recommendations, more output.

The bottleneck does not disappear. It moves.

Before AI, scarcity often sat with the people producing the work. Increasingly, it sits with the people who must review it, prioritize it, make decisions, and act on it.

That is the AI Work Paradox.

AI can make every producer more productive while simultaneously increasing the amount of work flowing toward managers, reviewers, architects, editors, and decision-makers.

Which leads to a very different AI productivity question:
Not: How much more can we produce with AI?
But: How much additional intelligence can the organization actually absorb and convert into value?

In the latest edition of Technology Bytes, I explore why AI may be shifting the real constraint from production capacity to absorption capacity, why “hours saved” tells only half the story, and where those saved hours actually go.

The AI Work Paradox: The Real Constraint Is Absorption

Thank you to our member organizations, and especially to the leads and teams from Amadeus, AVEVA, Schneider Electric, and Siemens, for applying the SCI for AI specification to real-world AI use cases.

This is exactly how a specification becomes truly useful: through community participation, practical implementation, shared learning, and continuous refinement. Together, we can turn measurement into action and make sustainable AI the norm. Green Software Foundation

AI can make your organization more intelligent while making it less knowledgeable.

That sounds contradictory.

But as AI starts doing more of the analysis, reasoning, and decision support, something subtle begins to shift:

The capability remains.

The understanding behind that capability starts moving elsewhere.

We have spent decades learning how to manage technical debt.
Now enterprises may need to manage a new kind of debt:
AI Intelligence Debt.

The issue is not whether we should outsource intelligence to AI. We should, where it creates value.

The real question is:
Which intelligence can we safely outsource, and which intelligence must the organization continue to own?

In the latest edition of Technology Bytes, I explore:

• how intelligence debt accumulates
• why knowledge, reasoning, and judgment are different
• the idea of an Intelligence Balance Sheet and why AI strategy may soon become a question of intelligence ownership

Beyond Technical Debt: The Rise of AI Intelligence Debt

Day 25 of Green, Efficient AI is live: The Interface Sets the Bill.

A user creates intent. The interface translates that intent into inference demand.

The model produces the bill. Two products can wrap the same underlying model and produce inference bills that differ by orders of magnitude, purely because the interfaces around them ask for different things.

A regenerate button that made sense against a weaker model. A model picker that defaults to the largest tier. A chat surface shipping the full session history on every turn. Dashboard summaries generated on cards nobody views.

Measure. Intent-align. Defer. Revisit. The ladder.

Read the full issue ↓

Green Software Foundation #greenai #efficentai #aieconomics #leanagenticai #ai

Before we build the next AI data center, we should answer 8 questions.

Not after construction.
Not in the next sustainability report.

Before the shovel goes into the ground.

1️⃣ How much water will it use — and from where?
2️⃣ How much electricity will it require — and who pays for the grid upgrades?
3️⃣ Can it turn down when the grid is under stress?
4️⃣ Whose land is being used — and who is affected?
5️⃣ What’s next door — a river, aquifer, forest, or neighborhood?
6️⃣ What are the public incentives worth per job actually created?
7️⃣ Who checks the commitments — and how often?
8️⃣ What happens if the data center closes? Who pays for cleanup?

These aren’t complicated questions.

But today, they aren’t always asked together, consistently, and early enough.
And that’s a problem.

AI data centers are becoming some of the largest infrastructure projects of the AI era.
They consume enormous amounts of electricity.
They require cooling.
They can consume significant amounts of water.
They occupy land.
And they become part of the communities and infrastructure around them.

This isn’t an argument against building them.
We need the infrastructure to power the future of AI.

But we need a better contract between AI infrastructure and the communities that host it.
That’s why in my latest episode of Agentic AI — the future of intelligent systems, I propose an AI Siting Ledger — a shared eight-question framework that puts the numbers, commitments, and accountability on the table before construction begins.

Because sustainability reporting tells us what happened.

The AI Siting Ledger asks what will happen — before we approve it.

And as agentic AI becomes more autonomous and persistent, the demand for compute will only grow.

The future of intelligence isn’t just a software problem.
It is a physical infrastructure problem.

🎙️ I’ve explored this in my latest podcast episode:
The AI Data Center Problem — What Does Intelligence Take?

👉 Spotify podcast link in the description below.

What would you add to the eight questions?

#AgenticAI #AIInfrastructure #DataCenters #GreenAI #Sustainability #AI Green Software Foundation

If you’ve ever felt like the world is moving faster than you can think —
Keep reading.

If you’ve sat in a meeting about “AI transformation” while quietly wondering what it all actually means for you —
Keep reading.

If you’ve built a life that looks impressive on LinkedIn but feels restless at 5 AM —
Keep reading.

If you’ve ever wanted to stop running, sit on a bench, and ask yourself one honest question —
What am I actually running toward?

Then this story was written for you.
6:17 AM.
Two strangers. One bench. A park by the sea in Mumbai.

Every morning at exactly 6:17, a man runs past. He builds AI systems for a living. His life is fast, optimized, and impressive on paper.

Every morning at exactly 6:17, a woman is already sitting there. She photographs sunrises for a living. Her life is slow, uncertain, and honest.

One morning, she looks up and asks him a question nobody at work ever has.
That conversation changes everything.

No plot twists.
No villains.
No algorithms.

Just two people figuring out what kind of life is actually worth building — in a world that won’t stop telling them to move faster.

I wrote this novel because I believe the most important question of the AI age isn’t what machines can do.
It’s what humans choose to do with the time, space, and freedom that technology creates.

6:17 AM is available for FREE on Amazon for the next few days.
It’s short. You can finish it in one sitting.

Maybe on your morning commute.
Maybe before your first meeting.
Or maybe tomorrow at 6:17 AM.

Link to the book - https://amzn.to/4ggnFbD

Link in comments 👇

The token model doesn’t tell us what AI accomplished.

It tells us what AI consumed.
Tokens.
Compute.
Inference.
Context.

But consumption isn’t the same as value.

An AI system can consume 100,000 tokens and still produce nothing useful.
Another can consume 10,000 tokens and successfully complete the same task.
Yet our AI economics often starts and ends with:

How much did it cost?

That’s the Meter Problem.
Think about a taxi.
The driver takes a wrong turn.
Then another.
Then a third.
The meter keeps running.

You wouldn’t accept paying extra because the driver kept taking wrong turns.

Yet that’s effectively what happens with AI.
You refine the prompt.
Add more context.
Retry.
Ask it to reason again.
The meter keeps running.

And with Agentic AI, the problem can multiply:
Reason → Tool call → Fail → Retry → Delegate → Reason again
Every step consumes resources.

So the question shouldn’t just be:

How much does a token cost?

It should be:

How much useful work are we getting from the tokens we consume?

Because:
More tokens ≠ More value.
More reasoning ≠ Better outcomes.
More iterations ≠ Greater efficiency.

The next evolution of AI economics isn’t just cheaper tokens.
It’s more useful outcomes per unit of compute.

That’s the thinking behind Lean Agentic AI. (https://leanagenticai.com/)

Intelligence measured by output.
Efficiency built by design.