The approach
Augmentation without cognitive atrophy.
The risk in enterprise AI is not adoption failure. It is what happens to analytical judgment when adoption succeeds.
The risk
Cognitive offloading.
When an analyst can produce a polished draft in a single prompt, the formative work quietly stops happening: sitting with the data, building a position from first principles, defending a number under pressure.
The analyst becomes a passive editor of model output. Quality looks the same on the page; the underlying judgment thins out. This is the failure mode a serious firm should be designing against.
The risk in any AI rollout is that talented people become editors of drafts they cannot fully defend.
The principle
AI as adversary, not author.
We treat the model as a stress-tester of the analyst's reasoning, not a generator of it. The analyst forms a position first; the model challenges it. The analyst defends the work; the model probes for weakness.
Used this way, AI sharpens the very skills that unconstrained use erodes. The output is better and the judgment behind it is stronger, not weaker.
Analysts produce the formative work first, then bring Claude in as a critic, a red-teamer, and a verifier.
The practice
Constraints that hold under pressure.
The principle becomes real through how the work is sequenced and reviewed. The shape varies by firm; the discipline does not.
- Analyst-firstThe analyst produces a defensible first pass before engaging the model.
- Show-your-workAppendices make the human contribution legible to senior reviewers.
- VerificationThe devil's advocate pattern is built into the workflow, not bolted on.
- CalibrationSenior-led reviews surface drift before it compounds across the team.
One thesis, both sides of the work
The approach governs how we train and what we build.
It is not a slide at the front of a deck. It is the operating principle behind every session we teach and every system we ship.
How we train
Every session is designed the adversary-first way, so the team builds verification habits as it builds fluency.
What we build
Agents and workflows are designed to challenge and verify human reasoning, never to quietly replace it.
See it in practice
The programmes put the principle to work.
The four-session arc carries the analyst-first principle from first principles through to building with Claude Code.
Explore the programmes →