Key Takeaways
- Automating a process you have not mapped tends to speed up its problems along with its useful parts.
- Observe real cases, including awkward ones; process documents drift away from daily practice.
- Measure time, volume and errors per step to create a baseline you can compare against later.
- Classify each step as automate, assist or leave with people, based on risk and judgement.
- Finish with a short ranked list of opportunities, each with a clear next step.
Why mapping comes first
It is tempting to start an AI project with the tool. A demo looks impressive, a vendor says it will handle the whole process, and switching it on feels quicker than studying anything. The difficulty is that business processes are messier than anyone's description of them. They contain exceptions, workarounds, handoffs between teams and informal checks that exist for good reasons. Automate a process you do not understand and you often speed up its problems along with its useful parts.
A workflow audit puts things in the right order. Before anything is built, you find out how work actually moves, where time goes, where errors appear and which steps are safe to change. At Artifex it is the second stage of our process, after agreeing goals with leadership, and it shapes everything that follows. Below is the method behind our workflow audit, written so you can apply it yourself.
The five steps at a glance
- Frame.Agree the process, its start and end points, and the outcome that matters.
- Observe.Follow real cases from start to finish and document every step and handoff.
- Measure.Capture time, volume and errors for each step to form a baseline.
- Classify.Decide for each step whether AI should automate it, assist a person or stay out.
- Prioritise.Rank the opportunities and link each one to a concrete next step.
Steps one and two: frame, then observe
Framing is quick but easy to skip. Write down where the process starts and ends, which teams are involved and what a good outcome looks like. For example, a finance team reviewing supplier invoices might frame the process as starting when an invoice arrives and ending when it is approved for payment, with accurate, on-time payment as the outcome. That boundary stops the audit sprawling into purchasing, budgeting and everything nearby.
Observation is where most of the insight comes from. Sit with the people doing the work and ask them to talk through what they are doing and why. Note every step, each system they open, every time they copy information from one place to another and every point where work waits for someone else. Watch for unofficial steps, such as a spreadsheet kept on the side or a colleague who is always asked to double-check certain cases. They often reveal the real controls in a process.
- Follow a mix of routine and awkward cases, not only the easy path.
- Observe at different times, including busy periods, if the work varies.
- Speak to the people at each handoff, not just the process owner.
- Record the data each step uses and where it comes from.
Step three: measure
Once the steps are clear, put numbers against them. For each step, estimate how long it takes per case, how many cases pass through it in a typical week and how often it produces errors or rework. Perfect precision is not required. Sampling a set of real cases and asking the people involved to check the estimates is usually enough to show where the effort sits. Where systems already record timestamps or volumes, use them, but check them against what you observed, since logged times often miss the work that happens outside the system.
These figures become your baseline. When you later build a solution or train a team, the baseline is how you show whether anything improved. Without it, results become a matter of opinion, which makes the next budget much harder to secure. We explain how finance teams use these figures in building a business case for AI.
Step four: classify each step
With the map and measurements in hand, go through each step and decide what role AI should play. We use three categories. Automate means AI carries out the step with checks and monitoring, but without a person reviewing every case. Assist means AI prepares something, such as a draft, a summary or a suggested category, and a person decides. Leave means the step depends on judgement, relationships or accountability that should stay with people.
The deciding factors are the cost of a mistake, how consistent the inputs are and whether the data needed is available and permitted for use. Extracting fields from a standard delivery note is a reasonable candidate to automate. Deciding whether to extend credit to a struggling customer is not, although AI might usefully assemble the information a person needs to make that call.
If a step exists because something once went wrong, find out what it was before you remove it.
Step five: prioritise and share
The last step turns findings into decisions. List each opportunity with its expected effect on time, errors or turnaround, how feasible it is with your systems and data, and the risks it carries. Rank them, and for each of the top few recommend a next step: build a solution, train the team to handle it or leave it for now. Use-case scoring is a structured way to do this consistently.
Then share the map and findings with the people who were observed. They will spot mistakes, and involving them makes later changes far easier to adopt. A good audit leaves you with an accurate picture of how the work runs today, a baseline to measure against and a short list of changes worth making.