We Asked AI to Document a Business Process. Here's What Happened.
What happened when we used AI to turn a conversation into a structured business process
Most businesses rely on processes that aren’t properly documented.
People know what needs to be done because they’ve been doing it for years. They know where the information comes from, who needs to be involved and what happens next.
But much of that knowledge sits in people’s heads rather than in a document someone else could pick up and follow.
And that becomes a problem when someone leaves or changes roles, or when the business wants to improve or automate the process.
So I wanted to try something.
What if we could simply have a conversation about a business process and let AI do the first round of documentation?
In Episode 16 of the FP&Ai Podcast, Anthony Wilson, Donovan Moses and I gave it a go.
Anthony brought his knowledge of our consulting delivery and capacity planning process. Donovan brought his experience of AI and process documentation, and I guided the discussion and challenged some of the assumptions along the way.
We used a custom GPT, essentially a ChatGPT assistant we’d configured with instructions for documenting business processes.
And we deliberately didn’t prepare a process map beforehand.
We wanted to see what would happen in a real situation, including when things didn’t go according to plan.
Watch Episode 16: Can AI Document Your Business Processes?
We started with capacity planning, not a process document
The process we chose was capacity planning in our consulting business.
We need to know what work has been committed, what new projects might be coming in and whether we have enough people available to deliver that work.
Anthony knew the process well. He had a few notes, but nothing formally documented.
That was enough to get started.
We began by identifying five main steps:
- Collect current work. Understand the backlog and work in progress.
- Collect future demand. Look at the sales pipeline and the likelihood of new work coming in.
- Account for internal commitments. Include leave, training and internal projects that affect availability.
- Analyse capacity. Compare the work planned against the people and hours available.
- Make decisions. Adjust schedules, resources or priorities to balance the workload.
That gave the AI a structure to work with.
But the five steps were only the beginning.
A process is more than a list of activities
When Anthony started explaining how we collect current work, we quickly got into the detail.
We needed information about our backlog, which is work that’s been agreed but not yet completed.
We also needed work in progress, including work already completed but not yet invoiced.
Those two things affect planning differently.
Backlog represents future work that still requires capacity. Work in progress can represent revenue that hasn’t yet been invoiced, but the people have already done the work.
If we don’t understand that distinction, our capacity and revenue forecasts can be misleading.
We then looked at the sales pipeline.
Which opportunities are likely to become projects? When might they start? What probability should we assign to them?
And we had to account for internal commitments.
Consultants take leave. They attend training. They work on internal initiatives.
Those hours aren’t available for client work.
As Anthony explained the process, Donovan and I asked questions about the information needed to make it work.
- Who is responsible for each activity?
- Where does the information come from?
- Which systems need to share data?
- What requirements must be met?
- What could go wrong?
- And how would we know whether the process was working properly?
That’s where process documentation becomes more than simply writing down a sequence of steps.
A good process document should explain what happens, who does it, what information is needed and what the outcome should be.
If the underlying information is wrong, the process won’t work
One thing that came through quite clearly was how dependent capacity planning is on accurate, up-to-date information.
Take timesheets.
If a consultant hasn’t submitted their timesheet, we don’t have a complete picture of the work they’ve performed.
Or consider project plans.
If a project manager hasn’t updated a schedule, we might think someone is available when they’ve already been committed to another project.
A few weeks later, we discover that we’re over capacity.
And the problem isn’t necessarily with the capacity planning model.
It’s with the information we’re feeding into it.
We also discussed how different sources of information need to come together.
Backlog, work in progress, sales pipeline and internal commitments all need to be consolidated before we can compare planned work against available capacity.
That creates requirements around data completeness, accuracy, timing and integration.
It also introduces risks.
An integration might fail. A sales forecast might be too optimistic. Someone might forget to update a project plan.
Documenting those requirements and risks is part of understanding how the process actually operates.
Turning a conversation into a Work Breakdown Structure
Once we’d talked through the process, the GPT began organising the information into a Work Breakdown Structure, or WBS.
A WBS breaks a process into smaller activities and provides a structure for recording the information associated with each one.
In our case, the AI captured details such as:
- The activities and sub-activities
- The people responsible
- The information required and outputs produced
- The systems and integrations involved
- The KPIs used to measure performance
- The requirements and potential risks
We could then transfer that information into Excel or Smartsheet.
Instead of having a transcript of a meeting that somebody still had to interpret and document, we had a structured first draft of the process.
One that the business could review, correct and expand.
We could add more detail to individual tasks, include links to the relevant Smartsheets and identify what was required at each stage.
That could support training, handovers, process improvement and eventually automation.
And it happened during a relatively short working session.
The recording itself was around 40 minutes, and the whole session took roughly an hour.
Normally, documenting a process involves workshops, taking notes, drawing process maps and circulating documents for review.
We hadn’t finished documenting the process in an hour. There was still work to do.
But we’d reached a point where we had a structured document that people could review, challenge and build on.
And that would normally take considerably more effort.
Then the AI started getting things wrong
This was probably the most revealing part of the demonstration.
When we asked the GPT to produce a more detailed version, things didn’t go entirely according to plan.
One export was incomplete.
In another version, information we’d already captured was missing.
At one point, the KPIs disappeared.
We had to go back, ask for corrections and check that all the steps were included.
It’s easy to be impressed when AI produces a neatly formatted table with headings, responsibilities and detailed descriptions.
But a professional-looking document isn’t necessarily an accurate or complete document.
And if you don’t know the process yourself, you might not even notice what’s missing.
- Have all the steps been included?
- Are the responsibilities correct?
- Have we captured the requirements and KPIs?
- Does the document reflect how the business actually operates?
We also learnt something about how we should capture the information.
During the demonstration, we recorded parts of the discussion in separate sections, which made it harder to manage larger outputs.
Next time, I’d be inclined to record the entire meeting, generate a transcript and upload that transcript to the GPT.
That would give the AI one complete source to work from, rather than several recordings captured at different stages.
It wouldn’t guarantee a perfect result, but it’s a more practical approach I’d want to try.
The real opportunity is what happens next
What struck me wasn’t simply how quickly AI could organise the information.
It was what we could do with the result.
Think about the processes your finance team relies on every month.
Month-end close. Budgeting. Forecasting. Management reporting.
How many are properly documented?
And if the person who runs one left tomorrow, how easily could someone else pick it up?
Once the process is documented, we can start asking better questions.
Why do we need this activity?
Could information move automatically between these systems?
Are we duplicating work?
Which approvals and controls really matter?
Where are people spending time on repetitive tasks?
Before you automate something, you need to understand how it works.
And if a process has never been properly documented, that can be surprisingly difficult.
AI doesn’t remove the need for people who understand the business.
Anthony still needed to explain capacity planning. Donovan and I still needed to ask questions. And we still had to review and refine the output.
What AI helped us do was turn that knowledge into something structured much more quickly.
AI doesn’t need to know your business better than you do. It needs to help you capture what your people already know.
And perhaps that's the real opportunity. Not replacing the people who understand the process, but making their knowledge easier to capture, share and build on.
In Episode 16 of the FP&Ai Podcast, you can watch us work through the entire process, including what worked, what went wrong and what we learnt along the way.
