AI Manifesto: Stage 2, Map
The Lean Lesson for AI: Automate the Repetition, Protect the Judgement
AI is unusually good at producing output. That is its superpower and its trap.
The trap is that output can be mistaken for progress. A model can summarise a meeting, draft a response, produce a report, and generate something that looks complete. In many organisations, that “looks complete” quality is enough to trigger deployment momentum.
What Stage 2 Is Actually Asking
Stage 2 of the Human-First framework is designed to stop that momentum from becoming drift. It asks a sharper question: “Where would AI make the biggest difference without removing judgment?” It is not asking where AI can be used. It is asking where AI should be used, and where it should be kept away.
The actions in this phase read like an experienced operations leader talking, not an innovation lab. Audit workflows for repetitive, low-stakes tasks that slow people down. Prioritise augmentation, with AI handling volume so humans can focus on judgement. Flag any tasks where human accountability is non-negotiable and keep AI out. Prefer narrow, specific tools over broad platforms at this stage.
Then the guardrail lands: Never map AI into a role that requires human empathy or ethical discretion.
Why This Is the Toyota Comparison in Its Most Practical Form
This is the Toyota comparison again, in its most practical form. Toyota Production System is often summarised as a socio-technical system. The “social” part is not incidental. It is the recognition that work is done by people inside processes, and automation must serve that reality. In lean environments, machines do not “own” quality. People own quality, supported by standard work and visible control points.
That maps cleanly onto the manifesto’s core line: “Humans Decide. AI Assists.” Mapping is where you decide what assistance looks like in each workflow, and where you design the hand-offs so decisions remain human where they must.
What This Looks Like in Practice
A concrete example is accounts payable. AI can read invoices, extract fields, match purchase orders, and route routine items. That is exactly the kind of volume-handling AI is good at. The moment an invoice mismatches the PO, the supplier details change, or the spend looks unusual, that is not “more AI”. That is a human exception queue. In lean terms, you have a quality gate.
A second example is customer support. AI can summarise long ticket histories, suggest likely fixes, and retrieve the relevant policy paragraph faster than a human can search a wiki. That reduces waste and improves speed. Where it must not go is into the part of the role where empathy, context, or ethical discretion is the job. The mapping stage’s guardrail is explicit, and useful precisely because it stops the conversation drifting into “let’s see what happens.”
Request the full AI Manifesto
Why Narrow Tools Beat Broad Platforms at This Stage
There is a reason the framework prefers narrow tools at this stage. Broad platforms encourage broad behaviour. Broad behaviour creates accidental delegation. This process is about drawing boundaries early, then earning your way into wider deployment later, after you have learned what works and what breaks.
Where Responsibility Has a Name Becomes a Design Requirement
This is also where “Responsibility Has A Name” stops being a slogan and becomes a design requirement. In a mapped workflow, the owner should be able to answer critical questions from auditors and frontline teams without improvising. What data is in scope? What is out of scope? What does the AI do, and what does it not do? What happens when it is uncertain? How do we prove what happened?
For a quality-management perspective, bring in Deming. The phrase “In God we trust, others must have data” is widely traced and discussed as a statistics and quality maxim, and the point is simple: decisions should be testable, not mystical. Mapping is where you decide what data and evidence you will need later to assess whether the AI is amplifying people or creating hidden rework.
How to Run a Mapping Session
The best mapping sessions look like value stream mapping, not ideation workshops. You start with the work as it is, then mark the friction. Where do people copy and paste? Where do they search? Where do they reconcile? Where do they chase approvals? Those are strong AI candidates because they are repetitive, time consuming and relatively low-stakes.
Then you identify the human control points. The contract sign-off. The disciplinary conversation. The credit decision. The patient communication. The safeguarding report. These are places where accountability is non-negotiable, and the framework tells you to keep AI out.
Choosing What Not to Do
If you need a strategic frame to make this feel credible at executive level, go back to Porter: strategy is choosing what not to do. Mapping is that choice expressed in workflow terms. It is where you say yes to augmentation and no to substitution, at least until governance, measurement, and capability mature.
How Map Connects to the Rest of the Framework
This is also where the broader stages start connecting together. Map becomes the bridge between Clarify and Commit. It gives leaders something concrete to fund and defend. It gives Upskill a curriculum tied to real workflows, instead of generic prompt training. It makes Pilot meaningful because the pilot boundary is designed deliberately, not guessed.
Later, when you get to Integrate and Govern, these maps become living documents. They are revised as the organisation learns, as risk appetite evolves, and as the technology changes. That connects to the framework’s emphasis on continuous improvement and the progress loop, where feedback from people and results from technology feed the next cycle.
Removing Waste So People Can Think More
Toyota did not treat automation as a replacement for thinking. It treated it as a way to remove waste so people could think more, not less. If AI adoption is going to be both fast and safe, mapping is where that philosophy becomes operational.
Request the full AI Manifesto