Intellyx Cortex by Jason English
I find it strange that as a software industry, we put in so much effort to make our systems efficient to design, build, and scale with DevOps practices, cloud native architectures, event-driven functions, optimized databases and storage, and resource-sipping Lean IT — then we threw it all out the window when Big AI came along.
Don’t get me wrong, I’m not an AI luddite. We’ve covered productive use cases for AI – from fraud detection, to medical research, voice and visual recognition, and software testing – all of which existed years before ChatGPT and foundational models burst onto the scene less than 3 years ago.
At the time, I could scarcely believe it when I learned how an LLM actually works, using a brute force method of training multi-billion-parameter probabilistic models on the whole world’s content so they can get really good at guessing the next likely word in a sequence, or mocking human art and real-world images.
Suddenly, every software company had to be an AI-driven, AI-first, AI-native company, or get left behind. Tone-deaf geniuses appeared on podcasts and at tech summits to explain how AI would someday be smart enough to replace all human intellectual labor and usher in a new age of productivity.
Rather than delivering end customer value for AI investments, we see circular financing arrangements – AI software funding for AI chips, chips funding AI data center capacity, data center capacity funding AI development, and so on – fueling unlimited growth. We know this can’t go on forever.
Borrowing from Lean Manufacturing
In the physical supply chain, we had Lean Manufacturing – keeping just the materials on hand needed to make and deliver products in quantities that would match customer demand as closely as possible, so there was less excess inventory cost and unproductive work in process.
And now, the supply chain for AI is beginning to rear its ugly head. Resource scarcity, climate disasters, trade wars, and global conflicts are set to constrain the physical inputs of AI chips and power-hungry datacenters, while negative economic, political, and social externalities are lining up to make big AI unsustainable.
When the big AI bubble bursts, we’ll need something lighter to help us continue to grow our capabilities without the encumbrance. Not ever-bigger data centers and more GPUs. And definitely not a superintelligence that might get out of control and rule the world. Fuck that, and anyone working on it.
I’m talking about AI that gives you only the intelligence and automation you need, at an all-in price the world can afford. Let’s call it Lean AI for now.
How AI shifts bottlenecks in software engineering
It makes perfect sense for AI to generate code. After all, AI is software, therefore it already speaks the language of code, which should be much easier for it than realistically handling a conversation in English, or any other human language. Code is the ground truth, an empirical way to directly provide instructions to a computer.
Today’s agents can generate code at a rate human developers could only dream of, and they also never need a break, or health insurance, or any of those bothersome things. This effectively reduces the human labor input cost of coding, and shifts that expense elsewhere:
Increased token opex. Just like Amazon and Uber acquired initial market share by setting consumer prices below delivery cost, the initial costs of coding agents are being highly subsidized by AI companies who currently enjoy high valuations. Now that we see coding agents being used at scale, the token meter is running, and companies are blowing through budgets early and rethinking tokenmaxxing, often rationing access to top-tier models.
Shifting coding time to review time. When AI generates 10x more code, it also makes more mistakes. Even if a modern coding engine with well-aligned context and guardrails were to produce a 5% error rate, that’s still a lot of mistakes to factor in at each change point, including requirement drift. Putting a “human in the loop” for code review becomes the new bottleneck, as no developer wants to spend all day reviewing someone else’s work, much less an agent’s.
Shifting professional resources to AI maintenance. The cheese has moved for consulting, but there’s still plenty of services work to do cleaning up the messes AI can leave behind. A new phenomenon of “forward deployed engineering” places AI engineers within companies, to help them operationalize changes and troubleshoot performance, scale (AI SRE) and security issues (AI SOC).
Increased system footprint for AI inference and data. The cloud native pattern of separating application logic and workloads from monolithic systems toward elastically scaled services was a good thing for enterprise agility and efficient utilization. Rather than just turning all of these services over to an AI for inference or agentic automation, we need to make decisions about which AI capabilities would provide ROI for our users, or face a hefty infrastructure bill (GPUs, memory, storage, cloud data costs, etc.).
Given the current public sentiment and political pressures against dotting the landscape with more noisy energy-and-water consumptive AI data centers, companies that want a sustainable future for AI shouldn’t wait to start looking for alternatives.
Lean AI is a movement, not a technology
Since we don’t like bucketing things here at Intellyx, I look at Lean AI as a movement rather than a technology category.
Lean AI is a “thing you can do” rather than a “thing you can buy,” just lIke DevOps combined agile development and disciplined automation with cross-team collaboration and continuous feedback for any company that delivers software.
Here’s the bright spot – we are already seeing lots of innovative vendors and open source projects that can contribute value to Lean AI efforts, such as:
- Minimized infrastructure deployment options that eliminate AI configuration mistakes and overprovisioning, while reducing the attack surface.
- Specialized AI models and agents that are trained only for specific processes, functions or industry scenarios, ideal for assembling into multi-agent workflows.
- Concierge-type orchestration agents that can decide to use simpler deterministic services and small agent models, while routing the most critical and complex workloads to large model inference systems.
- Data agents that apply semantic understanding to data in motion and at rest, informing observability and security work at lower latency and lower ingress/egress and storage costs.
- Routines that convert machine learning into reusable skills that agents can leverage for faster, more reliable output.
- Massively parallel event-based code review, testing and security agent fleets that are activated by a software change or system anomaly, to conduct investigations and inform an engineer only when attention is required within their work process, then deactivated.
- Edge-deployed inference models that can run on minimized endpoints and devices, where compute, power, and connectivity are severely limited.
- Neoclouds that provide pre-optimized AI environments that maximize GPU utilization, and sometimes even run low-traffic models on commodity CPUs and VPUs.
- Companies are repatriating AI like they repatriated cloud for cost and data sovereignty concerns, often customizing open models and local agents. Or, running AI on the mainframe — it’s already there, highly available, performant, and secure…
- Continuous cost and usage monitoring and controls to eliminate redundancies and token overruns, with continuous feedback from deployed AI resources.
As you can see, there are many avenues to Lean AI, but ultimately, each approach should lower AI cost and impact on the environment, while increasing human agency for continuously improving the customer value of our applications.
Reorienting AI as a human utility, not an authority
Aside from technical and process innovation, Lean AI encourages us to take a harder look at the second-order societal impacts of each use of AI. Is AI serving humanity, or replacing human agency?
It’s hard not to fear obsolescence when we see Super Bowl ads flippantly suggesting we outsource our professional communication, our understanding of the world, our memory — even advice on how we should talk to our loved ones – to a chatbot on the phone.
Sure, the ads can say they are unleashing human potential with Big AI, but in practice, plenty of companies have already accelerated layoffs or stopped hiring – often through decisions that were algorithmically determined. Was the AI initiative already so successful, that it drove those layoffs? More often than not, it was just an excuse for trimming labor costs.
We can’t settle for a pat answer like “just give everyone UBI or something and keep adding more infrastructure and it’ll all work out.”
Why am I even writing this article? Why should a company pay an analyst to write thought leadership, when AI generates content that seems very researched and complete?
AI content can be very useful for documentation, summarization, and research to support a decision, so long as the reader doesn’t automatically trust everything they read. We use a local open model here to summarize our own thoughts from thousands of human-written articles on a variety of topics.
AI-generated content lacks human insight, and though it can steal our words, it’s not relatable. We like how an AI can read us a recipe, or give us instructions on how to fix a bug, but nobody wants to read it. Even AI content searches favor authoritative sources over AI slop.
My colleague Jason Bloomberg wrote about rehumanizing AI in the last Cortex, and I can’t help but agree:
“… the central theme is that AI is dehumanizing by stripping away the uniquely human aspects of life and work. Until society recognizes this dehumanization and actively mitigates it, the world risks a dystopian future – a billionaire’s wet dream of a fully automated world.
While AI has become a permanent fixture in the modern economy, we must work to rehumanize AI systems before its dehumanizing effects become irreversible.”
The Intellyx Take
Like many folks, I’ve struggled with my weight over the years, so I’m the last person to shame anyone for getting dessert. But I will shame your big fat AI if it’s not churning away for a good reason.
For AI to achieve long-term sustainability it will need to provide more value to customers, business, and society as a whole than the resources it consumes. AI needs to put humans in the lead, not a human in the loop.
I’m calling it Lean AI for now, but you can call it whatever you like. Maybe you could ask ChatGPT for a catchier term. I’m certainly not claiming to be the ultimate authority on this subject. I’m just a human who has worked in the software industry for most of my life now, and I’m still willing to learn and change my mind.
Maybe I should start up a show around this topic. The world could always use another podcast. Who wants to come on my show?
Copyright ©2026 Intellyx B.V. As of the time of writing, none of the organizations mentioned in this article are Intellyx customers. Not a word of this article was written by AI. Image source: Oleg Yunakov, wikimedia commons, CC4.0 license.
