In October 2006, Brad Garlinghouse, then a senior vice president at Yahoo, circulated an internal memorandum. The Wall Street Journal published it the following month and labeled it the Peanut Butter Manifesto. His critique was a capital allocation failure. "I've heard our strategy described as spreading peanut butter across the myriad opportunities that continue to evolve in the online world," he wrote. "The result: a thin layer of investment spread across everything we do and thus we focus on nothing in particular."
Enterprise IT budgets are repeating that pattern with generative AI. Rather than choosing one high-volume bottleneck and building an audited production workflow, organizations scatter spend across departmental experiments. Legal buys a contract analyzer, runs a dozen sample PDFs, and files the evaluation deck. Each department head can say their group is testing AI. The organization as a whole has not cut a single operational hour.
The Activity Illusion
The spread feels safe because it avoids trade-offs. No director is left out, and communications can say AI is being piloted company-wide.
The cost of that safety shows up as coordination, not as labor saved. A sandbox pilot does not reduce workload. Every extra vendor still needs a procurement ticket, a security questionnaire, and recurring update calls. Teams spend hours on sample prompts and quarterly evaluation slides. When the annual renewal arrives, the tool is canceled because nobody wired it into daily operations. In the model above, USD 250,000 spread across twenty pilots buys the demos and the meeting hours. It does not buy a production queue.
Concentration as Financial Discipline
When I look at enterprise AI spend, I do not start with token price. I start with whether the money is concentrated enough to finish one workflow.
USD 250,000 can fund twenty surface-level pilots and twenty disconnected demonstration decks. The same capital can sit behind one unglamorous operational problem: matching three-way purchase orders, delivery receipts, and vendor invoices in accounts payable.
Moving that workflow into production takes engineering: schema validation, deterministic fallbacks, a human review queue, and monitoring for drift. That work does not get finished when budget and attention are split across twenty teams.
If a deployment cannot show hours saved on the weekly roster, it is new software overhead.