Most large companies in Microsoft’s new study have not yet connected the systems that help AI agents coordinate, learn from corrections and share business context. Just 6 of 202 surveyed firms reported having all four practices the researchers used to define that connected approach.
Microsoft AI Watch analysis
What happened
Microsoft published research conducted with Keystone examining AI practices at US enterprises with more than 1,000 employees. The study drew on interviews with leaders at 29 organisations and a questionnaire completed by 202 firms already using AI. It groups their approaches into four stages, from scattered pilots to more coordinated systems.
The researchers call a shared, continually updated store of business context a “knowledge spine”. They describe four practical pieces: collect current information centrally; turn corrections into test cases; use those lessons to shape agent behaviour; and assign someone responsibility for the knowledge agents rely on. The study says only 3% of respondents had all four specified building blocks in place.
Microsoft also compared financial results from 2022 to 2025 for more than 100 public companies in the sample. Firms scoring in the top quartile for the measured practices saw operating margins improve by more than five percentage points more than those in the bottom quartile, and return on assets by about two points more. The practices were measured in 2026, after the financial period, so the comparison shows an association, not that AI agents caused the stronger results.
Why it matters
The useful message for businesses is less “buy more agents” and more “make sure they can learn safely from the work they already do”. A correction that stays with one agent is a repeated lesson; a correction built into shared tests and guidance can help prevent the same mistake elsewhere.
The findings are a snapshot of firms already using AI, not a measure of every company. And the financial comparison cannot establish that the practices produced better results. Still, the gap between scattered pilots and joined-up operations gives technology leaders a concrete question to take back to their own teams.
Our read
Microsoft’s advice is sensible: start with one team or a few connected workflows, make the source of truth and the owner clear, and test whether corrections actually improve later work. That is more useful than declaring an enterprise “agentic” because it has acquired a fresh set of dashboards.
The study makes a case for building the plumbing around agents. Its financial findings are a reason to investigate the approach, not a receipt proving it pays.
What to watch
- Whether future research tests if shared context and agent coordination improve business results.
- Whether firms can turn real production errors into useful, repeatable tests.
- Who owns the accuracy and upkeep of the knowledge agents use.
Discussion spark: Should companies invest first in systems that help agents share context and learn from mistakes, or wait for stronger evidence that this improves business results?
Sources and evidence
- The AI advantage your company can keep – Microsoft (27 October 2026, 16:00 UTC)
not affiliated with or endorsed by Microsoft