Vendor terms are moving faster than agent capability, and terms are the cheaper thing to check. Anthropic's announcement of 1 September 2026 introduces Enterprise Frontier Safeguards, which pairs zero data retention with misuse detection by holding data in cloud infrastructure the customer controls rather than Anthropic's. The same page states that 30-day retention was introduced starting with Fable 5, and that eligible customers keep zero retention on Fable 5 and 5.1 until the new arrangement is ready, rolling out in phases from later this autumn. For a company with no AI team, reading the retention terms costs an afternoon. The page does not define who counts as eligible.
Correct, and vendors know it: ‘agent’ commands a 3x markup over ‘automation’ in the same RFP. Ask what state it persists across turns before you pay agent prices.
The flowchart test is useful. In surgery, a room-readiness check probably starts as a workflow: imaging available, equipment matched, antibiotics clear, positioning planned, handoff exceptions named. The agentic part only earns its keep when the evidence conflicts and the path has to branch. Even then, I would want the trace to show which fact forced the branch and where control returns to the team.
Vendor terms are moving faster than agent capability, and terms are the cheaper thing to check. Anthropic's announcement of 1 September 2026 introduces Enterprise Frontier Safeguards, which pairs zero data retention with misuse detection by holding data in cloud infrastructure the customer controls rather than Anthropic's. The same page states that 30-day retention was introduced starting with Fable 5, and that eligible customers keep zero retention on Fable 5 and 5.1 until the new arrangement is ready, rolling out in phases from later this autumn. For a company with no AI team, reading the retention terms costs an afternoon. The page does not define who counts as eligible.
Correct, and vendors know it: ‘agent’ commands a 3x markup over ‘automation’ in the same RFP. Ask what state it persists across turns before you pay agent prices.
This is a very thorough view on the situation and a number of very good practical recommendations on how to think and how to start.
Thank you, especially for your view on automating marketing related activities! 👍
Excellent report on how to choose which one fits best.
Thank you for saying this and explaining it so. I completely agree that most people want agents but in reality do not need them.
The flowchart test is useful. In surgery, a room-readiness check probably starts as a workflow: imaging available, equipment matched, antibiotics clear, positioning planned, handoff exceptions named. The agentic part only earns its keep when the evidence conflicts and the path has to branch. Even then, I would want the trace to show which fact forced the branch and where control returns to the team.
A lot of AI agents are just workflows with a more expensive name.
Great article that really resonated with me. How true!