All playbooks
Playbook · AI transformation · 7 steps
AI transformation 3: From assistants to agents
Raise AI's autonomy step by step, with evidence, safeguards and people who can still take over.
“…the more advanced a control system is, so the more crucial may be the contribution of the human operator.”
- 1Workflows Before AgentsWould a fixed workflow do this more reliably than an autonomous agent?Use the simplest LLM design that works; add agent autonomy only when a problem truly needs it.Open
- 2Evals as Fitness FunctionsHow will we know, automatically and on every change, that it still works?AI behaviour is probabilistic and shifts with every model, prompt and data change, so check quality with automated evaluations on real cases, run on every change and tracked over time.Open
- 3The Autonomy SliderHow much autonomy can we verify quickly, and what evidence earns the next step up?Give AI as much autonomy as you can verify quickly — and let people move the slider up as trust grows.Open
- 4Ironies of AutomationWho stays skilled enough to take over when the AI fails, and how do they keep in practice?Automating the routine work leaves people with the hardest tasks, monitoring they do poorly, and fading skills, just when a failure needs them most.Open
- 5The Lethal TrifectaDoes any agent combine private data, untrusted content and a way to send data out?An AI agent that combines private data, untrusted content and a way to send data out can be tricked into leaking it.Open
- 6Guardrails, Not GatesCan AI governance be automatic policy instead of a committee every use case queues for?Govern by setting clear, ideally automated, boundaries and paved roads, and review only what crosses them, instead of making every change queue for approval.Open
- 7The Bitter LessonAre we hand-building rules and prompts that the next model will make redundant?General methods that scale with computation beat clever, hand-built knowledge — eventually, and by a large margin.Open