Organizations adopting generative AI are pulled in three directions at once: toward real, worth-capturing productivity gains; away from the risk that fluent, confident AI output can quietly erode human expertise and accountability without anyone noticing the substitution; and toward caution, given how unsettled enterprise AI pricing still is — a workflow that is well justified today could require rollback if costs shift sharply. My approach treats responsible AI adoption as an intentional, reflective, continually-revised practice: built on consistent guiding principles, but adapted through the judgment and lived experience of the people actually using the tools day to day. The specific tools in my current toolkit come out of an academic background in epistemology and philosophy of science, and I expect that toolkit, like the technology itself, to keep growing.

The Tension

Genuine productivity gains, worth capturing deliberately: tasks that once took hours — content generation, estimation, drafting, analysis — can now take minutes. That gain is worth designing for directly, rather than either apologizing for it or banning it outright.

The quiet erosion risk: this is where catastrophic errors are actually born. The danger is rarely a single dramatic, visible failure — it's the gradual substitution of fluent AI output for genuine human expertise and accountability, happening slowly enough that no one notices until the underlying skill and vigilance needed to catch mistakes is already gone. Once that erosion sets in, both large and subtle errors can get approved and enacted without real scrutiny, often surfacing only once the consequences have fully matured and become hard to reverse or contain.

Cost volatility as a structural risk, not a footnote: enterprise AI pricing is not yet stable. A workflow that is economically justified today, at current pricing, may need to be unwound if usage costs rise sharply – which means an adoption strategy has to build in reversibility from the start, not treat cost as fixed. That means defining break-even metrics and rollback triggers up front, at the point of adoption, rather than debating them under pressure later: a clear, pre-agreed decision point turns unwinding a workflow into a scheduled operational step rather than a fresh executive battle, and guards against the sunk-cost fallacy that keeps organizations funding a workflow well past the point it still makes sense.

The Approach

Rather than a fixed rulebook, which ages badly against a fast-moving technology, or an unconstrained free-for-all, which surrenders real accountability, I favor an intentional, reflective, and continually-revised approach: consistent guiding principles, applied and adapted through the actual judgment and experience of frontline users who see a tool's real capabilities and failure modes well before any policy document catches up. This mirrors an instinct that shows up across my other work: build something, watch closely how it actually performs in practice, and revise deliberately before scaling, rather than assuming the first design is the right one.

Why This Matters Here

An enterprise-wide AI learning strategy has to work across a large, genuinely varied organization, where frontline realities differ meaningfully across practice areas – exactly the environment where adaptive, principle-based guidance earns its keep over a fixed rulebook, and where planning for cost volatility protects the organization from being locked into workflows it can't unwind if pricing shifts. I'd bring this same set of instincts (consistent principles, continuous revision, and real weight given to frontline judgment) directly into building and evolving any AI learning strategy. None of this works as a document handed down once: it requires continual collaboration across every tier of the organization, from the C-suite to frontline practitioners, to keep the guidelines connected to real priorities rather than drifting out of date the moment the technology or the business moves.