Don't add AI to a broken process: map the workflow first
Automating a workflow you haven't mapped usually automates its problems. A practical, step-by-step way to map a business process before writing any code or adding AI.
Every team has two versions of its process: the one in the document and the one that actually happens. Automation built on the first version fails on the second. In our experience, automation projects rarely fail on the happy path. They fail at the exceptions nobody wrote down.
Why the documented process misleads
Process documents describe intent. Real work includes the side spreadsheet one person keeps, the WhatsApp message that unblocks an approval, the customer who always sends the wrong form, and the manager who checks every order above a certain value even though no policy says so. These are not failures of discipline. They are how the business actually copes, and they are exactly what an automated system will collide with.
What a useful workflow map contains
| Element | What to capture | Why it matters |
|---|---|---|
| Actors | Every person, team and system that touches the work | Shows where handoffs happen and who must be involved |
| Inputs and outputs | What arrives, in which format, and what must leave | Decides whether you need extraction, integration or both |
| Volume and timing | How often it happens and when it spikes | Tells you whether automation is worth it and how it must scale |
| Decisions | Where someone chooses, approves or rejects, and on what basis | These become rules, AI judgments or human checkpoints |
| Exceptions | What goes wrong, how often, and who fixes it | The part automation most often breaks |
| Waits | Where work sits idle between people or systems | Usually the biggest source of delay, and the cheapest to fix |
How to map in practice
- Pick real cases, not the procedure
Choose five to ten recent cases, including at least two that went wrong. Follow each one from the moment it arrived to the moment it closed.
- Watch the work being done
Sit with the people doing the job. Note every system they open, everything they copy, every message they send to unblock something.
- Record decisions with their reasons
When someone makes a call, ask what they looked at and what would have changed their mind. Those reasons become rules or evaluation criteria later.
- Draw the flow, including the waits
Lay out actors as lanes, steps as boxes and waits as gaps. Time the waits where you can. They are often longer than the work.
- Play it back to the team
Walk the people who do the work through the map and ask what is wrong with it. The corrections are the most valuable part of the exercise.
The output is a map the team recognises as true. That is the only kind worth automating. If people look at it and say that is not how it works, keep going until they say that is exactly how it works.
Reading the map: four kinds of opportunity
Once the map exists, opportunities usually sort themselves into four groups. Most teams are surprised how few of them need AI.
- RemoveSteps that exist for historical reasons
Approvals nobody acts on, reports nobody reads, duplicate data entry. Removing them costs nothing and frees time immediately.
- IntegrateData that moves by hand between systems
Copying from email to CRM, from CRM to invoicing. An integration fixes these more reliably and cheaply than AI.
- RulesDecisions that follow a clear policy
If the rule can be written as a few conditions, write it as code. It is predictable, testable and cheap to run.
- AIReading, judging or writing unstructured information
Classifying messages, extracting data from documents, drafting responses, summarising cases. This is where AI earns its place.
The highest-value AI opportunities are often invisible until someone studies how decisions, data and approvals actually move.
A worked example
Consider an illustrative accounts-payable team that receives supplier invoices by email. The documented process says an invoice is checked against the purchase order and approved. The map shows something more detailed: invoices arrive in three formats, a third need a phone call because the PO number is missing, approvals above a threshold wait for one manager who travels, and the finance system is updated by hand at the end of each day.
The redesign follows the four groups. The daily batch upload is replaced by an integration. The approval threshold rule is written as code, with a deputy approver when the manager is away. AI extracts invoice data from all three formats and flags missing PO numbers with a drafted query to the supplier. A person reviews every mismatch. The AI is only one part of the improvement, and it works because the rest of the flow was fixed around it.
Mistakes to avoid
- Interviewing managers instead of watching the people who do the work
- Mapping only the happy path and leaving exceptions for later
- Treating the map as a deliverable rather than a basis for decisions
- Choosing the technology before the map is finished
If a step cannot be described clearly on the map, it is not ready to be automated. Fix the description first. That single rule prevents most of the expensive surprises in automation projects.
Frequently asked questions
Why map a workflow before automating it?
Because the documented process is rarely the real one. Mapping exposes exceptions, handoffs and hidden decisions that automation would otherwise break or copy.
How long does workflow mapping take?
For one workflow, usually days rather than weeks: a few sessions observing real cases, interviews with the people who do the work, then a map they review and correct.
What should a workflow map include?
Actors, inputs and outputs, frequency and volume, decisions and their rules, exceptions and who resolves them, and the handoffs and waits between steps.
Which steps should be automated with AI?
Steps that involve reading, sorting, judging or writing unstructured information at volume. Steps that are pure data movement usually need integration rather than AI, and some steps should simply be removed.
Get the next note by email
One email a month at most. Practical, from real work. Unsubscribe in one click.