A practical guide to rolling out AI, without breaking what works.
Most automation fails. Not because the technology is wrong, but because it scales broken processes faster. This framework is how we avoid that, across every engagement. Eight sections, built from years of shipping this in real businesses.
The most expensive automation is the one that scales a broken process.
Most automation fails. Not because the technology is wrong, but because it scales broken processes faster. Automating a flawed workflow just enables the flaws at higher velocity.
The real bottleneck in most businesses isn't the repetitive task. It's the decision-making, the exceptions, and the undocumented knowledge surrounding the task. Fix that first. Then automate.
Build your own intuition before you touch team workflows.
Before automating anything that touches your team, spend time using AI on low-stakes personal work. Email. Calendar. Research. This builds real intuition for what good automation actually looks like, and where the failure modes live.
Three places to start:
- Email organization and prioritization
- Scheduling and meeting preparation
- Information synthesis and research
Personal automation experience is what separates leaders who make good decisions about business automation from those who chase demos.
Score every process before you decide to automate it.
Track tasks across your team for five business days. Note how long each one takes and how often it runs. Apply the formula above. The highest scores are your first automation targets.
The hiring test:
Would you hire someone to do this task using nothing but a written checklist? If the answer is no, you need to document the process before you can automate it. The documentation work is the prerequisite.
Every task has steps that should be automated, and steps that shouldn't.
Tasks aren't monolithic. A workflow is a sequence of discrete steps, each requiring a different amount of human judgment. The question isn't “should we automate this?” It's “which steps, and to what level?”
Four levels of automation:
Example: “Update website listings” is eight steps. The first four and the last two are mechanical. Steps five and six require judgment. That tells you exactly where the automation boundary sits.
Not every process is ready to be automated.
Some processes are ripe for automation tomorrow. Some aren't ready for another six months of documentation. Some will never be a fit. The checklist tells you which bucket you're in.
- Documented and repeatable
- Structured, predictable inputs
- Low cost if something goes wrong
- Runs daily or more often
- Doesn't require tribal knowledge
- Lives in someone's head, not a document
- Unstructured or unpredictable inputs
- High stakes if something goes wrong
- Runs monthly or less often
- Requires context you can't easily write down
Autonomy is earned, not granted.
Every automation starts in supervised mode and graduates through three phases. You never skip phases. Autonomy is earned through demonstrated reliability, not hoped for.
AI assists. A human decides. You build trust in the tools and figure out where they actually work.
AI acts. A human approves before anything executes. You capture edge cases and turn them into exception rules.
AI handles end-to-end. A human audits periodically. You only get here after the earlier phases prove out.
Some things should never run on autopilot.
The embarrassment test:
If this automation sent the wrong output, would it damage a client relationship, create legal exposure, or embarrass your business? If yes, a human approves every output. Full stop.
- External client communications
- Financial transactions
- Legal and compliance actions
- Anything reputation-critical
- Internal data processing
- File organization and routing
- Status updates and notifications
- Monitoring and alerting
Every automation starts in supervised mode. Accuracy gets tracked for two to four weeks. Only then does it graduate to autonomous.
Measure the right thing. Time reclaimed, not tools deployed.
Three metrics that matter:
- Time reclaimed per week (by person and by process)
- Quality and consistency gains (errors caught, edge cases handled)
- Error rate before vs after (absolute reduction, not just percentage)
The goal isn't to automate everything. It's to automate the right things, so your team can focus on the work that actually requires a human brain.