Only 6 of 202 companies share what their AI agents learn, and the top scorers earned higher margins
A Microsoft-backed study links a shared 'knowledge spine' for agents to better margins, but it shows correlation, not cause.
In brief
Only 6 of 202 firms let AI agents share lessons
A Harvard working paper surveyed 202 US firms. Top scorers gained over 5 points more operating margin. It is correlation, not proof that agents caused it.
A working paper from Harvard Business School, written with Keystone and partly funded by Microsoft, looked at how companies manage AI agents. Agents are AI programs that carry out work tasks on their own. The team interviewed leaders at 29 organizations with advanced AI use, then surveyed 202 US enterprises with more than 1,000 employees each, between June and September 2026.
The idea: a knowledge spine
The paper's key idea is a shared system that stores the context agents need. It also stores the corrections, tests and lessons that come from their work. Imagine one agent gets a refund rule wrong. With a shared system, every other agent learns from that fix, not just the one that erred.
Four stages
The researchers sorted companies into four maturity stages.
- Stage one: scattered pilots, often just chat, with little business impact.
- Stage two: agents work on real workflows, but humans coordinate and check them. That is hard to scale.
- Stage three: systems, not people, coordinate agents. This is the turning point.
- Stage four: orchestration, governance and a knowledge spine come together, and lessons flow back into shared knowledge.
What the numbers say
Only 6 of the 202 companies, about 3%, reported all four practices that connect AI operations. For over 100 public companies, the team compared results from 2022 to 2025. Top-quartile scorers improved operating margins by more than 5 percentage points beyond the bottom quartile. Return on assets rose about 2 points more.
How to build one
The paper lists four practices: gather business context in one place, turn every correction into a test case, use that information to change how agents work, and name an owner for the knowledge. Companies also build a golden dataset, a fixed set of test cases with known right answers, to check an agent before it ships. Agents that perform well can then be given more freedom.
Would your company know who owns what its agents rely on?
Limits
This shows correlation, not causation. The practices were measured in 2026, after the financial period, and that period precedes most of today's agent systems. The sample is only US firms already using AI, so it does not represent the whole economy. The paper treats movement between stages as a hypothesis, and more research is planned to test cause and effect.
Source: Microsoft Source