Process mining — from ad hoc work to a formal process

Work that starts from an email and passes between people and agents leaves handover records. Once the same flow has piled up a few times, process mining finds the pattern and suggests it to an admin. When the admin registers it, the process starts on its own from the next similar email, and each step goes to the right person or agent.

Drawing from the top vs. finding from the bottom#

In Chapter 17 you drew a process from the top. Team lead Jinyoung Jang described the flow in words, and the assistant turned it into BPMN. The travel expense settlement in Chapter 18 was also a definition someone had drawn in advance. That works well when you already know the flow.

But most work doesn't start as a process. The Moonlight Coffee Store #2 job also began with one email and was handled case by case through Chapter 9–Chapter 16. Process mining means finding the recurring flow from the bottom, in the records of that ad hoc work.

What gets recorded — handover records#

Lite groups work that started from an email into one case, and inside that case it writes one handover record line each time the work moves on. The format is the same as Process GPT's handover records.

Each handover record line holds who → to whom, what work, when, and which tool (CADCraft, Word, Excel) and skill was used.

Try it 1 — Build up past records#

Mining needs at least 4 cases before it looks for a pattern. The screens in this chapter were captured after loading, as practice data, the handover records of three similar requests Jinyoung Jang's team handled earlier in October, plus one real case: the Moonlight Coffee Store #2 practice mail of Chapter 9, whose drafts were shared with Seoyeon Park (Finance) and then answered. In a real company, these records pile up on their own over a few weeks.

All four went in the same order:

Receive request mail (Jinyoung Jang) → Draw floor plan (drawing agent) → Write proposal (proposal agent) → Prepare quote (quote agent)
  → Finance review (Seoyeon Park) → Reply to customer (Jinyoung Jang)

Try it 2 — The miner finds the pattern#

Every minute, the Lite server asks Process GPT's process miner (a calculation service that finds recurring flows), "Is there a recurring flow?" The miner finds flows with fixed calculations and no AI model, so the same records always give the same answer.

1. It builds the set of activities for each case and groups cases whose sets overlap by 60% or more (so that different kinds of work don't merge into one model).
2. Within a group, it treats the activities that appear in at least half of the cases as core activities, and takes the most common order as the main flow.
3. For each step, it counts who handled it (role and owner), the share handled by agents, and the tools and skills used.

If a case is still in progress (it hasn't reached the last step, or it just moved), the miner waits. Suggesting a flow from a case in progress would leave out the later steps.

In our run, this suggestion appeared within a minute of the reply that closed the live case (the name and summary come from an AI model, so they may differ from run to run). Two Store #2 practice mails had been handled by then — one replied to after a Finance share, one replied to without a share — so the suggestion counts 5 cases.

Try it 3 — The suggestion reaches the admin#

Suggestions go only to admins. The admin's Home shows the notification card "I found a recurring flow: Design proposal & quote." The card description says how many cases were handled and in what order.

The last 5 mail jobs were handled in the same order (Receive request mail → Draw floor plan → Write proposal → Prepare quote → Finance review → Reply to customer) …

When someone who isn't an admin opens the mining screen (address ending in /mining), the server refuses the suggestion list and the suggestion details. In the current English version the screen then simply shows no suggestions (the Korean screen says "Only admins can see this.").

Open the suggestion from the notification card or the mining screen, and you see the following. In this chapter the admin is team lead Jinyoung Jang.

Try it 4 — Register it as a process#

When the admin clicks Register as a process, the miner derives the flow once more from all the cases so far and saves it through Process GPT's own process-definition save path.

Click On hold and the suggestion folds away, and the same pattern isn't suggested again. As more cases pile up, the decision stays the same and only the evidence grows. Even if newly grouped cases overlap more than half with an existing suggestion's cases, no new suggestion is made; they're added as evidence to that suggestion, so the same flow is never suggested twice.

Try it 5 — The next email becomes a process#

Once the process is registered, when a new design request email comes in, the assistant starts the registered process instead of making ad hoc drafts. The chat of the person who got the email shows this:

I'll handle the … mail with the registered 'Design proposal & quote' process. The step agents make the floor plan, proposal and quote, and the finance review goes to Finance.

A process card is attached below it, and from there it's the same as Chapter 18. In our run, when a new email ("[Onyu Clinic] Request for an interior design proposal and quote for our Pangyo clinic") arrived in team lead Jinyoung Jang's inbox, the formal process started within seconds of the mail being read.

When the three agent steps finished (about 17 minutes in our run), the Finance review went to Seoyeon Park. Her Home shows a Decision · Finance review card with the gist of the quote step and chips for the previous steps' outputs (floor plan, Word proposal, quote).

Check: each step has an owner. In our run, the instance's to-dos were assigned like this: Receive request mail and Reply to customer → Jinyoung Jang; Draw floor plan, Write proposal and Prepare quote → the three agents; Finance review → Seoyeon Park.

Next#

Once formal processes start running, instead of handover records you get run records that keep the exact start and end time of every step. In Part 6, "Improving processes," you'll use these records to find and fix where work piles up (Chapter 20), how much time adds value and how much is spent waiting (Chapter 21), and which path sets the total duration (Chapter 22).