The Computer Era Explains Why AI Pilots Stall

The Pattern Is Older Than AI

Robert Solow's 1987 observation about the computer age became famous because it named a real executive frustration: computers were visible everywhere, but the productivity statistics did not immediately move.

That does not mean computers were useless. It means the first wave of adoption placed new tools inside old operating models. Firms bought technology before they rebuilt the work around it.

AI is in the same awkward middle. Individual operators can draft, code, analyze, and summarize faster with Claude Code and Codex. The company still may not see ROI if review, data, approval, and ownership stay unchanged.

Technology Arrives Before the Operating Model

Paul David's work on electrification showed that major technologies often need complementary redesign before they show up as productivity. Early electric motors did not transform factories when managers treated them like a cleaner replacement for steam power. The deeper gains came when factory layouts, workflows, and production logic changed around electricity.

Brynjolfsson and Hitt later made the same point about information technology. The economic value of computers depended heavily on organizational investment, intangible capital, process redesign, and measurement that traditional accounting often missed.

That is the useful analogy for enterprise AI. The gap is not between weak models and strong models. The gap is between tool access and operating redesign.

Where DTC Marketing Teams Are Now

A CRM operator can use Claude to draft campaign copy faster. A social lead can summarize exports faster. An analyst can generate a query faster. Those are real gains, but they are local gains.

The business result still depends on the workflow around the model: where the request starts, what context is trusted, which data definitions matter, who reviews the output, what the CTO can inspect, and how the next request starts smarter.

Without that layer, Claude Code and Codex become a better sidebar. With it, they become part of the department's operating memory.

The Adaptation Layer

• One workflow with a named owner.
• Source-controlled context the CTO can inspect.
• Approved data definitions and business rules.
• Human review before consequential actions.
• Reusable memory so the next request starts with more context.
• A small first install before the department expands the pattern.

Why Half a Point Compounds

Tyler Cowen's growth estimate is not that the economy gets 0.5% bigger once. He is talking about roughly half a percentage point added to annual economic growth.

That sounds small in one year. It is not small over decades. FRED real GDP data shows the US economy grew from about $2.46T in 1950 to about $20.28T in 2020, measured in chained 2017 dollars. That is about 8.25x, or roughly 3.06% annualized real growth.

Add 0.5 percentage points to that annual growth path and 2020 real GDP would be about $28.47T instead of $20.28T. That is roughly 40% larger, or about $8.18T more annual output in 2017 dollars.

That is the enterprise lesson. A small annual growth-rate advantage can become enormous, but only when the organization absorbs the tool into the way work actually gets done.

The Data Zen View

The lesson from the computer era is not to wait for better tools. It is to stop treating the tool as the whole transformation.

For DTC marketing teams, the practical move is to choose one repeating workflow and rebuild the path around Claude Code, Codex, approval, and operating memory. That is where the ROI starts to become visible.