Generative AI in the enterprise: what's strategy, not tooling

I use Claude Code daily, with domain skills and scheduled agents. But when a company asks me to help drive generative AI adoption, the question that comes up first, "which tool should we pick," is almost never the one that decides whether adoption actually sticks.

The easy question: which copilot to pick#

Comparing benchmarks, trialing two or three assistants over a sprint, picking the one developers like best: that's the visible part, the one where most company discussions stop. It's necessary, but it's only a fraction of the real decision, and it's rarely what makes a rollout fail.

The real question: where does the data go, and who answers when it breaks#

Before talking about tooling, a serious company has to answer questions that aren't technical in the strict sense: what data can flow through the model, under what retention regime, with what contractual guarantees. In a regulated environment, this is often what blocks a rollout for months, not the quality of the code suggestions. It isn't a tooling question, it's a governance question, and it comes before the first generated line of code.

Measuring something other than "it feels faster"#

The productivity a team perceives when discovering an AI assistant is almost always higher than the productivity actually delivered, because the visible effort (typing code) drops faster than the invisible effort (understanding, reviewing, fixing) rises. A company that steers its adoption on collective gut feeling rather than real delivery metrics (end-to-end lead time, review rejection rate, production incidents) discovers its true gain well after it's already made its investment decisions.

The link that doesn't move: accountability#

An AI assistant changes who types a line of code, never who answers for it. The strategic decision that matters isn't "which tool generates the best code," it's "where does human review sit in the chain, and who signs off." A company that hasn't answered that question before deploying a tool finds out, at the first production incident caused by generated code, that it had delegated a decision it had no right to delegate.

What this changes in practice#

Picking the tool takes one meeting. Data governance, measuring real impact, and placing accountability get built over several months, before and after the rollout. It's that invisible part, not the benchmark, that decides whether generative AI adoption holds up over time or fades once the novelty wears off.