Helping CTOs & CIOs turn AI ambition into an operating model: feedback loops, governance, and execution across people, process, technology | CAIO @ Zühlke
@bprasad_ai Short-lived grants move the hard part upfront. Someone has to define the task and sign the grant. That is the gap I see most. You cannot scope or expire access for an agent nobody has listed and nobody owns.
1/ 97% say they are ready for an agent incident. 6% could stop the agent within an hour.
46% would need more than a day to revoke its access to all systems. That is from an October survey of 900 tech leaders in the UK and US by Gravitee, which sells agent governance tools.
5/ An agent can act on its own. It cannot answer for itself. If one of your agents went wrong tonight, who answers for it, and could that person stop it within the hour?
Survey: https://t.co/Rhm4FZalQx
4/ Machines cannot carry responsibility. The deployer chose the agent's permissions, systems and data. Autonomy reaches as far as your gates hold, and revocation is the last gate. If it takes a day, the agent has a day of autonomy nobody granted.
@zyrasproof Machine-checkable is the hard part. Most runbooks are prose, so nobody can prove the fix stayed inside them. And someone has to own the envelope. If it grows a bit after every incident, you are back to unreviewed changes, only slower.
1/
If you need a human to triage thousands of logs, your system is already obsolete.
Gartner expects AI agents in 60% of IT operations management tools by 2028. End of 2024: fewer than 5%.
5/ Be predictable about when you decide. Stop promising what will work.
How many dates on your roadmap were set before anyone tested the idea?
https://t.co/MkZD9EV8Ez
4/ My view: leaders are right to want both. They ask for predictability in the wrong place. The rhythm can be predictable: when we plan, review and decide. The content cannot. Nobody knows in January which idea will work in June.
5/ Blast radius decides where the human stays. The factory grows one class of work at a time. Which class would you let ship without a person first?
https://t.co/kCu5DmzOyG
1/ If someone still looks at every agent run, you do not have an AI software factory.
You have what Kaspar von Grünberg calls "a very busy Level 2 platform".
4/ A gate decides per change. A fix that cannot break much ships with no human. A change that could be fatal waits for an operator.
For medical devices I would draw it the same way: the pump software stays with people, the settings dialog can be generated.
The bundle The Cybernetic Enterprise Complete Edition by Romano Roth is on sale on Leanpub! Its suggested price is $81.96; get it for $17.50 with this coupon: https://t.co/83X2dzkElo @RomanoRoth#devops
5/ A guideline is a preference, and your agent walks past preferences. Which of your AI rules could your agent skip today?
Paper: https://t.co/EkY8fHMusB
Survey: https://t.co/aZ7DE4bJ2O
1/ Not one Swiss software company runs a fully autonomous dark factory.
That is this year's Swiss Software Industry Survey. Almost 90% use AI in development. Swico's CEO: AI supports, people steer and check. Many steer with guidelines and checklists.
4/ Build toward a dark factory and you learn this first: autonomy reaches exactly as far as your gates hold. Better prompts will not fix it. Put the rule where the agent cannot pass it: branch protection, required checks, tests beyond mocks.