A routine reporting workflow can become a decision system one tool at a time.
Consider a hypothetical reporting agent. It reads approved sales data, prepares a weekly account summary, and leaves the draft for a manager. The organization then connects it to a customer-record system so it can keep that system current.
The new connection looks like a small convenience. It also introduces a new decision: when may the agent change the record on which another team relies?
Anthropic’s 2026 State of AI Agents Report, developed with Material, surveyed more than 500 U.S. technical leaders in late 2025. It reports that 57 percent of respondents’ organizations used multi-stage agent workflows, including 16 percent using cross-functional processes. These are respondent-reported practices. They do not establish whether a particular organization’s controls are adequate. [1]
For the hypothetical reporting workflow, start with the first action that changes someone else’s working record. Before enabling it, write down the condition that permits the change. “Keep the system accurate” is too broad to resolve a disputed customer status or a missing source document.
A useful review would make the decision concrete:
Which fields may the agent change, and which must remain untouched?
What source supports each proposed change?
Who reviews an ambiguous case, and what happens while that review is pending?
What record lets another person understand and correct the change later?
Now test a disagreement. The agent’s weekly summary says an account is active; the account owner’s note says discussions have stopped. Which record governs? Does the agent pause, ask, or overwrite? The answer should be settled before a live exception forces the organization to improvise.
This example illustrates the practical consequence of expanding an agent’s authority. The added tool deserves its own approval decision, even when the underlying model and business objective stay the same.
For the next agent review, choose one proposed permission and trace it to the person or team affected by its use. Ask that owner to walk through an exception with the operator. A small, specific rehearsal can expose an assumption that an aggregate adoption figure cannot answer.
Source and further reading
This edition draws on “Anthropic’s 2026 State of AI Agents: An Executive Briefing,” published and reviewed by Techné AI on September 11, 2026. The full briefing sets out the wider framework for workflow ownership, oversight and performance measures:
https://techne.ai/insights/anthropic-state-of-ai-agents-executive-briefing/
[1] Anthropic, The 2026 State of AI Agents Report. Survey location: page 3; methodology: page 5; workflow findings: page 8.
https://resources.anthropic.com/hubfs/The%202026%20State%20of%20AI%20Agents%20Report.pdf