Éric Sadin, Palantir, and the Cognitive Desert of the SAP Consultant

In my last newsletter I talked about the corporate guillotine, the end of body shopping, and how the only way out is to stop being the code bricklayer and become the system orchestrator. But I left out the fine print of that contract.

I'm a fan of Éric Sadin. I've read several of his books and right now on my nightstand I've started La vie spectrale. At the same time I squeeze Claude and Gemini every day to ship code and solve all kinds of problems. Sadin isn't a technology-hating Luddite — for me he's a reality surgeon. In his most recent analysis he warns about something I see every day in the industry: the vegetalization of humanity.

By delegating all analytical effort to generative models, we're becoming what he calls coquilles vides — empty shells wandering in a psychiatric desert of ourselves. If the machine writes, thinks, and solves for us, the vital instinct and critical capacity atrophy.

We become stupider. And now there's hard data confirming it. Sadin is right.

Judy Hanwen Shen and Alex Tamkin, researchers from Anthropic and Stanford, published a paper in January 2026 that measured exactly this. They took experienced programmers, set them up to learn a new Python library, and divided them into two groups: with and without AI assistance.

The results should concern you.

The AI group produced the code. Management would have been happy. But their conceptual understanding, their ability to read architecture, and — this is what hurts — their debugging skills dropped 17% compared to the group that fought through the errors alone. And the most perverse part: AI didn't even make them faster. Many spent time formulating prompts instead of thinking.

The researchers identified six usage patterns. The three that destroy learning share something in common: total delegation. "Write me the code" and paste. The three that preserve learning share something else: cognitive friction. Asking "why," requesting explanations, solving errors yourself and then using AI to verify. Participants who delegated everything scored 39% on the evaluation. Those who maintained cognitive effort reached 86%.

AI gave them the output, but robbed them of the competence.

SAP and Palantir: The Earthquake

Does this feel distant? Let's bring it into the SAP trenches.

SAP and Palantir formalized a strategic alliance in May 2025, and in the Q4 earnings call Palantir's CTO announced that their AIP product now executes SAP ECC to S/4HANA migrations in as little as two weeks. The market understood immediately and panic set in: nearly $300 billion in market cap evaporated in a single session.

The cases published on their website speak for themselves. A multinational food company with a mandate to migrate more than 20 markets in 24 months. Impossible with the traditional playbook. Palantir sat with the technical team for 24 hours. The agents ingested raw ECC data, generated S/4-ready datasets, and inferred business rules directly from transactions — skipping months of requirements workshops. In another case, millions of lines of custom code were scanned, cross-referenced with live orders, and half was flagged for deprecation before a destination platform was even chosen. The migration volume dropped more than 60%. And a retailer with over 250 interdependent SQL Server queries migrated to PySpark by a swarm of parallel agents. 100% accuracy. Token cost: under $4,000.

Do you understand the seismic consequences of this?

The model of traditional consulting firms and partners — billing thousands of man-hours for manual mappings and endless blueprints — has an expiration date. Companies are going to stop paying for seat time. They'll pay for agent orchestration and high-level exception resolution.

The Everyday SAP Consultant

But this isn't just a Big 4 or partner problem — they'll figure out how to keep generating revenue. Think about the everyday SAP consultant. The functional consultant who today lives off gathering requirements in workshops, mapping data, and configuring modules. On one side, an agent like AIFD solves half the migration. On the other, Joule for Consultants automates much of the daily work of configuration, documentation, and troubleshooting. Their scope is being compressed from both sides. And if on top of that, this person doesn't question what agents return, doesn't audit, doesn't verify, doesn't understand why the agent made that decision — then they're not orchestrating anything. They're blindly signing off on what a machine decided for them.

This is where the bricklayer separates from the real architect.

You can't guide a swarm of autonomous agents if you don't understand the ERP's internals. You can't audit migrated code if your brain atrophied because you let AI think for you over the last three years. The Shen and Tamkin paper proves it with numbers: the hard technical base, the pain of learning, of breaking things and understanding deep architecture, is non-negotiable.

But watch out: this isn't a Luddite argument. The same paper shows there are ways to use AI that preserve learning.

Wheelchair or Exoskeleton

The key is whether you use AI as a cognitive wheelchair or as an exoskeleton. The wheelchair takes you places, but your legs die. The exoskeleton amplifies your strength, but you're still making the effort.

In a world where a script migrates an entire system in two weeks and AI simulates reasoning to perfection, your only unfalsifiable advantage is your raw cognitive capacity. Mental friction. The intellectual effort of not taking the shortcut when everyone else does.

Don't let comfort hollow you out from the inside. Automate the transactional garbage, but don't let them replace your backbone.

Shall we begin?