The Value of Knowing Your Business
Most owners assume the hard part of adopting AI is choosing the technology. It usually is not. The models, and the platforms used to build AI assistants, are widely available, and your competitors can buy the same ones you can. What cannot be bought is knowing your own operations well enough to point AI at the right problem.
The evidence on this is getting clearer. Anthropic, a leading AI company, analyzed roughly 400,000 sessions in which people worked with an AI agent on real software tasks.¹ In a typical session, the person made about 70 percent of the planning decisions, determining what should be done, while the AI worked out how to do it. What separated the strong sessions from the weak ones was how well the person understood the problem, not whether they had technical training. The researchers make the point with an example: an accountant who has never used a programming language, but who tells the AI exactly which rules a reconciliation must enforce and catches the edge case it mishandles at month end, counts as an expert at that task.
The reverse holds as well. In a field experiment in Kenya, entrepreneurs were given an AI assistant. The less experienced ones ended up with lower revenues and profits than those who worked without it, because they acted on the AI’s generic advice that did not fit their situation.² Stronger operators drew out suggestions tailored to their business. The tool amplified whatever judgment was already there.
Better models will not close that gap. A weak instruction is rarely a wording problem. Vague thinking by humans means poor results.
Here is a test worth applying before you approve any AI project that impacts a business process, including one where an AI agent will operate autonomously with limited supervision. Name the person in your company who knows that process well enough to tell the system when it is wrong. If you can name that person, you have a good place to begin.
Sources
¹ Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, Ryan Heller and Peter McCrory, “Agentic coding and persistent returns to expertise,” Anthropic, June 16, 2026, based on a privacy-preserving analysis of roughly 400,000 sessions from about 235,000 users.
² Nicholas Otis, Rowan Clarke, Solene Delecourt, David Holtz and Rembrand Koning, “The Uneven Impact of Generative AI on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya,” Harvard Business School Working Paper No. 24-042, October 2025, reported in “What is really happening to jobs? Separating AI hype from reality,” Stanford Institute for Economic Policy Research, July 2026.