AI Automation

Before any tool: what an automation is made of, what should never be automated, and the dry-run that saves you a fortnight.

3 lessons · 13 min

1. 1.1Trigger, steps, outputYou can write the one-sentence spec for a recurring task, pick the right trigger type, and name who would notice if the output were wrong5 min
2. 1.2What to automate, and what never toYou can score candidate tasks on frequency, rules and blast radius, and defend both your first build and your first refusal4 min
3. 1.3The manual dry-runYou can run a three-input dry-run that either clears an AI step for unattended use or shows exactly which part needs a person4 min

The tools compared honestly, wiring two systems together, and putting an AI step in the middle without letting it run loose.

3 lessons · 20 min

1. 2.1The tools, honestly comparedYou can choose one of the four platforms for a specific job and name the single fact that decided it4 min
2. 2.2Connecting two systemsYou can build a flow one tested step at a time, and read each step's returned data closely enough to catch what would break the next step4 min
3. 2.3Adding an AI stepYou can write a classifier step whose output a branch can match exactly, with a permitted failure word and a default branch that is safe and loud12 min

Automations do not crash. They quietly stop being right, which is worse, and entirely preventable.

3 lessons · 19 min

1. 3.1Silent failure is the real enemyYou can read a run that reports success and tell whether it actually did anything, then wire the two defences that make the answer visible9 min
2. 3.2Designing the human checkpointYou can place a checkpoint where a reviewer can judge an item in under thirty seconds, and test whether it is actually being read5 min
3. 3.3Cost, limits and runaway loopsYou can set the caps and the exclusion filter that keep a flow cheaper than the work it replaces, and prove it cannot trigger itself5 min

Making it someone else's to own, judging when an agent is genuinely worth it, and what to build over the next ninety days.

3 lessons · 15 min

1. 4.1The runbookYou can write a page that lets a colleague find, understand and switch off your automation without asking you anything4 min
2. 4.2Agents: what they are, when they are worth itYou can decide between an agent and a fixed flow for a task, and name the property of the task that decided it8 min
3. 4.3Your ninety-day roadmapYou can commit to a build order for the next ninety days, with the audit that turns things off already in the calendar3 min

What changes once there is more than one: permissions, changing something already live, and chains that fail safely.

3 lessons · 19 min

1. 5.1What your automation is allowed to reachYou can grant the narrowest access a flow needs, move it off a personal login, and keep a register that survives an access review5 min
2. 5.2Changing something that is already runningYou can change a live flow with a captured before-output and a rollback, and catch the break that still looks right to a human4 min
3. 5.3Chains, and the empty inputYou can build a chain where a step that receives nothing refuses to run, and prove yesterday's data cannot be passed off as today's10 min

The three things that separate someone who has built an automation from someone who can be trusted with one: bounding an agent, getting it approved, and owning the output.

3 lessons · 17 min

1. 6.1Scoping an agent so it cannot surprise youYou can bound an agent with tools, caps and a reversible write target, and read a run trace closely enough to name what is wrong in it5 min
2. 6.2The approval caseYou can put a one-page case to whoever approves an unattended process, with the four predictable questions answered as mechanisms4 min
3. 6.3Answering for a system nobody watchedYou can answer the five ownership questions about something you actually run, and fix or switch off whatever fails one8 min