👋 Hey, this is Artem and this is The Blueprint. At WIQ, we spend each day forward deployed at Fortune 500 companies working through the hairy journey of AI transformation.
A process can take weeks to finish even when each step takes minutes. That creates a problem when we try to learn it by watching people work: we can collect a lot of activity without seeing a single case reach the end.
SAP's September 22 announcement brings this into the agent discussion. Its Process Consulting Agent is adding process mining across SAP and non-SAP systems to identify bottlenecks and recommend where agents could help.
I'm glad the conversation is moving toward understanding the process before automating it. But the first question I have about the resulting analysis is fairly basic: did we see the job finish?
Whoever does the estimate needs the full assignment.
What an unfinished case leaves out
At WIQ, we plan the observation period around how long each run of the process takes. A team that completes requests throughout the day gives us repeated examples quickly. When a process has long waits between steps, it can take weeks before we see one case from start to finish.
Consider an invoice with an amount that differs from the purchase order. The operator opens the invoice, checks the purchase order, and asks another department to explain the difference. The observation period ends before the answer arrives.
We have enough information to reproduce those steps. But what happens when the answer comes back? Someone still has to decide whether the explanation is acceptable, whether the invoice needs a correction, and whether payment can proceed.
I need to understand that part before I hand the invoice to an agent. More recordings of people opening invoices won't tell us how they resolve the disagreement.
The decisions behind an exception
We can't keep recording indefinitely. A monthly reconciliation might fall outside the observation period. A disputed invoice might arrive too rarely to appear at all.
For those cases, I want an experienced operator to walk me through an example they already resolved. Which records did they check? Why was one answer acceptable and another wasn't? The recording may show the action but miss the unwritten rules behind the decision.
Sometimes the operator describes the usual practice and the case in front of us doesn't match. Then we ask: did the procedure change, or did this invoice need an exception? We need to resolve that difference before turning either version into the agent's instructions.
Some workarounds only make sense to their owner.
The work we can assign now
We may be ready to let the agent retrieve documents, compare amounts, and prepare a question about the difference. The operator can keep resolving disputes and approving payment while we work through the decisions we haven't captured yet.
For that assignment, I want to test whether the agent finds the right documents and identifies real differences. Completed cases give us records and outcomes to check its work against. We keep those test cases separate from the examples we used to write the instructions.
I don't need every exception solved before we start. I do need to know which decisions still belong to the operator. In this example, the agent can prepare the question. We don't know yet how it should act on the answer.
We work with many Fortune 500 companies on this. Let's talk through the process you're assessing, what you already know, and where WIQ could help.
|
The Blueprint. Stories and lessons on AI transformations from founders and operators making it happen.