In 30 seconds

Useful AI work rarely starts with a model. It starts by understanding what people actually do, making the workflow visible and getting the important information into a reliable form.

The Workflow Ladder gives you eight stages. Move upward only when the stage below is clear enough to support it.

The ladder

REAL WORK
   ↓
1. OBSERVE       Understand what actually happens
   ↓
2. MAP           Make steps, decisions and exceptions visible
   ↓
3. DIGITIZE      Create reliable inputs, states and handoffs
   ↓
4. STRUCTURE     Clarify knowledge, rules and ownership
   ↓
5. ASSESS        Decide where AI genuinely helps
   ↓
6. DESIGN        Define behavior, authority and boundaries
   ↓
7. BUILD & TEST  Start small and evaluate real cases
   ↓
8. OPERATE       Monitor, maintain and improve

You can read the ladder from top to bottom when you are starting a new project. If you are already inside a project, use it to find the stage where the uncertainty actually lives.

Why projects skip the first steps

Someone asks, “Can we use AI for this?” The most visible part of the answer is the AI, so the project begins with prompts, tools and demonstrations.

The hidden work is less exciting but more important. Which version of the process is real? Where do decisions come from? Which spreadsheet is current? What happens to exceptions? Who may approve a change?

If those questions are unanswered, the AI does not remove the uncertainty. It inherits it.

A small example

A shared inbox receives requests for replacement parts.

At first, the opportunity sounds simple: let an AI read each email and create an order. Observation reveals that experienced employees check the machine model, interpret several customer-specific names for the same part and call a supervisor when warranty status is unclear.

That discovery changes the project. The first useful improvement may be a structured intake form and a visible exception queue. An AI-assisted classification step can come later, once the workflow has somewhere reliable to put the result.

The eight stages

1. Observe what actually happens

Use recent examples. Watch the task or ask someone to walk through it. Treat official documentation as evidence, not automatically as truth.

2. Map the workflow

Write down triggers, steps, decisions, people, systems, handoffs and exceptions. A simple map everyone recognizes is better than a perfect diagram nobody uses.

3. Make the work digital enough

Create reliable inputs and visible states. This may mean a form, a shared queue or a clearer record in an existing system. It does not automatically mean replacing spreadsheets.

4. Structure the knowledge

Identify sources of truth, resolve contradictions and assign owners. Knowledge that nobody maintains will become stale whether a person or an AI uses it.

5. Decide where AI helps

Keep stable rules deterministic. Use AI when interpretation helps, especially with language, documents or variable inputs. Some workflows do not need AI.

6. Design the AI workflow

Define what the system may read, decide and change. Separate capability from authorization. Make approval, escalation and failure handling explicit.

7. Build and test

Start with the smallest useful slice. Test normal cases, difficult cases and prohibited cases using real examples from the workflow.

8. Operate and improve

Deployment creates ownership. Monitor results, review incidents, maintain knowledge and decide who changes the workflow when reality changes.

What the ladder does not solve

The ladder is a sequence for thinking, not a rigid project plan. Real work loops backward. Testing may reveal a missing rule. A new exception may change the map. A source of truth may need repair before a pilot can continue.

Use the ladder to locate the work, not to pretend the work is linear.

Good enough to move on when

  • You know which stage contains the next important unknown.
  • The people doing the work recognize the current description.
  • You can explain why the next stage is now useful.