An AI agent at Sophos now resolves a security case in about 89 seconds. The same work used to take a human analyst around 38 minutes. That number will get the attention. The interesting part is the rule that sits next to it.
Three stories from this week, from cybersecurity, warehouses and public relations, look unrelated. Read together, they describe the same design problem. Capability is no longer the hard part. Deciding who presses send is.
Signal one: Sophos wrote the permission tiers down
In a case study published by OpenAI, Sophos described how agents built on OpenAI models now handle its managed detection and response cases. Its sensors produce trillions of events a day, distilled into roughly 1,000 to 2,000 cases for nine security operations centres. An investigation agent gathers context, a planning model builds and executes a plan, and analysts review a summary with recommended actions.
The results: average response time for agent-handled cases fell from about 38 minutes to 89 seconds, and 52 percent of cases are now resolved end-to-end by AI. That is a remarkable operational shift.
But the part I would copy is the permission model. Sophos runs three modes with each customer.
Notify, where Sophos investigates and recommends but the customer acts. Collaborate, where both sides agree before action. Authorise, where Sophos responds directly. The same boundaries apply whether the work is done by a person or an agent.
That last sentence is the whole lesson. Sophos did not invent new rules for AI. It wrote down the trust levels it already had with humans and applied them to machines. As its CTO, John Peterson, put it, anything the team is not comfortable with an agent handling goes to human judgement.
Signal two: at Amazon, the robot sets the tempo
Ars Technica published a deep look at Amazon's robotics program. Amazon runs more than 1 million robots in its warehouses and still employs more than 1.5 million people, which makes it the world's second-largest corporate employer. Its Sequoia system, which coordinates swarms of mobile robots, can identify and store items up to 75 percent faster.
The quote that stayed with me came from Brian Greenwood, a business professor at George Mason University. For a century, he said, most automation happened away from workers. Collaborative robots change that because they are physically close and because the robot can dictate the pace of work.
That is a different kind of handover. Nobody formally gave the robots authority over people. But when a human workstation is fed by machines, the machine's speed becomes the human's target.
The trigger moved without anyone deciding it should. I explored a version of this in why AI agents won't replace whole jobs yet: the tasks split, and the human ends up working at the seams of the automated ones.
The same thing happens in offices. When an AI system drafts fifty responses an hour, the reviewer's job quietly becomes keeping up with the system. If nobody sets that pace deliberately, the tool sets it by default, and review turns into rubber-stamping.
Signal three: Honeyjar will not hit send for you
The third story is smaller and more explicit. TechCrunch profiled Honeyjar, an AI tool built by two PR veterans to give founders and non-specialists access to senior communications judgment. It drafts announcements, finds journalists to pitch, builds media lists, tracks coverage and surfaces speaking opportunities. Its waitlist reportedly drew 10,000 signups in its first 24 hours.
The design choice I find interesting: Honeyjar can draft pitches and posts in a customer's voice, but the customer must send or publish them. CEO Michelle Masek said they never want a world where the human is removed from the loop. She also said the product might tell a founder they are not ready for coverage yet.
That is the trigger held deliberately by the human, in a field where a bad send damages reputation instantly and permanently. A pitch to the wrong journalist, at the wrong moment, with the wrong claim, cannot be recalled.
What this means for your own automation
Put the three side by side and a simple framework appears. Sophos decided who holds the trigger, explicitly, per customer, per action. Honeyjar decided the human always holds it, for anything public. Amazon shows what happens when the question is never asked: the machine ends up holding the pace even if the human holds the button.
Most marketing and operations teams deploying AI right now are in the Amazon position without realizing it. They have tools that draft, schedule, bid and respond. They rarely have a written rule for which actions an agent may take alone, which need approval, and which stay human. Without those rules, the default is set by whatever the vendor shipped.
Here is the exercise I would run this month. List every action your AI tools can take. For each one, assign Sophos's three levels: notify, collaborate or authorise.
Then check the review workload against the output volume, so humans approving work have the time to actually read it. If you want to build automation with those boundaries designed in from day one, difrnt.ai (difrnt.ai) builds custom AI workflows end-to-end with approval tiers as part of the architecture.
Speed is easy to buy now. The decision about who presses send is the one you have to make yourself.