Without the map, any AI solution is a guess. Process mapping turns impressions about “where things are slow” into evidence of where the real bottleneck is—and only the real bottleneck is worth tackling first.
Illustrative diagram — simplified value stream map with an identified bottleneck.
🔗 Value stream mapping
VSM is a snapshot of the entire flow: from input (a customer request, a document received, data created) to output (delivery, response, decision). The goal isn't a pretty diagram — it's reveal where the time goes.
How to build the VSM
- 1.Define the scope: which process marks the beginning and end of the map?
- 2.List all the steps in sequence (use sticky notes or a whiteboard)
- 3.For each step: who does it, which system they use, how long it takes
- 4.Add wait times between stages (handoffs)
- 5.Calculate the total lead time and compare it with the pure processing time
Total process time (including wait time)
Actual work time at each step
cycle / lead = % of time that adds value
⏱️ Measure time, wait time, rework, and cost
Data without measurement is anecdote. Every step needs a number — without this, prioritization is based on whoever shouts loudest in the meeting, not where the problem actually is.
Metrics by stage
| Metric | How to collect | Why it matters |
|---|---|---|
| Processing time | Time it / ask | Where human work goes |
| Wait time (queue) | System log / observe | Where time disappears without work |
| Rework rate | Correction history | Hidden quality cost |
| Unit cost | Salary × time × people | Business case foundation |
💡 Tip: the annual cost formula
Annual cost of the stage = daily frequency × cost per occurrence × business days
Example: manual analysis of 50 documents/day × R$12 each × 250 days = R$150.000/year. This is the number that justifies investing in automation.
🍾 Find the bottleneck (Theory of Constraints)
Eliyahu Goldratt demonstrated that every system has exactly a bottleneck — the slowest step that determines the speed of the entire workflow. Improving any other step doesn’t change throughput: only the bottleneck can do that.
How to identify the bottleneck
- ✓It’s the step with the longest wait queue before it
- ✓The previous steps sit idle while waiting for her to clear them
- ✓The following steps wait for her to produce
- ✓The team at this stage is always busy / overloaded
The 5 TOC steps
- Identify the constraint (the bottleneck)
- Explore the constraint (get the most out of it without investing)
- Make everything subordinate to the bottleneck
- Raise the constraint (invest to expand capacity)
- Return to step 1 — the bottleneck has shifted
The classic mistake: optimizing the wrong step
Automating a step that isn’t the bottleneck just makes it produce faster while waiting in the bottleneck’s queue. The result is a system that’s faster in 80% of the steps, with the same final outcome. It’s the most common waste in AI projects.
📊 Volume and Frequency Data
Volume × frequency is the impact multiplier. A time-consuming process that happens 3 times a year rarely warrants automation. A quick process that happens 500 times a day usually does.
Questions to collect volume data
- •How many times does this process happen per day/week/month?
- •What are the peaks? Is there seasonality?
- •How many people are involved in each occurrence?
- •Is this volume growing, stable, or declining?
Fewer than 20 occurrences/month. Automation is rarely worth the investment. Consider process optimization or a checklist.
More than 100 occurrences/day. AI automation often has ROI within months. High priority on the matrix.
🧮 Repetitive work vs. judgment-based work
Not all work can be automated in the same way. The key distinction: rule-based work (deterministic) vs. based on judgment (contextual). AI lives in the middle ground.
Deterministic
Fixed rules, no variation
- • Send an email when X
- • Calculate tax according to the table
- • Route by document type
→ Classic automation (RPA, if/else)
Standard with variation
Recognizable but variable patterns
- • Classify ticket category
- • Summarize a lengthy document
- • Extract data from an invoice
→ AI/LLM (workflow or agent)
Contextual judgment
Decision depends on broad context
- • Negotiate a complex contract
- • Legal decision involving risk
- • Differential medical diagnosis
→ Human (AI can assist)
📝 Document and validate the map with the team
The map you drew by yourself is always wrong. Validate with the people doing the work is required before turning the map into a diagnosis. The validation process itself is valuable: it generates buy-in.
Map validation workshop
Never send it beforehand — the reaction to seeing it live is more honest.
Keep a marker handy. Mistakes are expected—you don’t work there.
"Do you agree this step takes ~X minutes?" — group estimates are more accurate than individual ones.
Informal stages, approvals over WhatsApp, checks that “everyone does but no one knows about.”
💡 The valuable side effect
When the team corrects and improves the map together, they become “owners” of the diagnosis. This drastically reduces resistance to change when you present the recommendation — because they built it with you; it wasn’t imposed on them.
🎒 Module summary
Next module:
2.3 — Catalog of AI opportunities by function