# How Does the Process Run When Nobody's Watching? - Author: Thomy Gölles - Published: 2026-11-07 - Language: en - URL: https://thomygoelles.com/four-questions-process/ - Topics: AI Agents, Microsoft 365 Copilot > Part 4 of five questions we ask before anyone builds an agent. A pilot built on the handbook meets a process where exceptions are the rule, and an agent that only waits for questions never touches the real work. When we ran the next masterclass on 3 November, this time under the title "From FAQ bot to autopilot", the four questions from September had become five. The new one went in right after the question about the problem, and it's the one I'd now least want to skip: how does the process run when nobody's watching? It comes from an article Daniel Kornum, Head of Engagements at Varick Agents, published at the end of October, [Fixing Enterprise AI Pilots](https://x.com/dkfromdk/status/2116291363407044676). He lists the seven reasons he keeps running into when AI pilots fail, ordered by how often they come up, and adds that model selection isn't on the list. Three of the seven sit inside this question. Two more belong to the last part of this series. ## The handbook has six steps, the work takes 22 days Kornum's example is vendor onboarding in procurement, and I'd bet your company has a version of it. On paper there are six steps, the request, the procurement checks, tax and banking details, bank verification, sanctions screening and finally the entry in the ERP. If you follow the SOP, it takes five business days. In reality it takes 22. The form arrives half filled in, bank details show up in every format imaginable, including a letterhead scan from a Gmail address, treasury calls back to confirm, compliance wades through false positives, software vendors need a separate security questionnaire, and one in ten vendors is already in the vendor master under a slightly different name. A pilot scoped on the SOP meets a reality in which exceptions are the norm, according to Kornum often more than 40 percent of the cases, and it falls over. That's the first reading of the question. What actually happens, away from the handbook? If you build from the handbook, you build for the cases that already work. ![Diagram titled "The handbook has six steps. The work takes 22 days." Example vendor onboarding, after Daniel Kornum. The handbook is a straight line with six steps, request, checks, tax and bank details, bank verification, sanctions, ERP, five business days. Reality is a looping, tangled line with exceptions marked in red, often more than 40 percent, taking 22 days. A small marker near the start shows the eight minutes a typical pilot saves. Footer: if you only answer the second reading, you automate the handbook and not the work.](https://thomygoelles.com/img/2026/11/four-questions-process-handbook.webp) ## Eight minutes in a 22-day process The second reason made me grin, because I wrote about it [last week](https://thomygoelles.com/ai-roi-unmeasurable-roles/). Without a real process map, the step that gets automated is the one that demos best. In vendor onboarding that's an LLM reading the tax form and posting name, tax ID and address to the vendor master, which saves the person in procurement about eight minutes per vendor. Kornum does the maths. 22 days are roughly 31,000 minutes, so the pilot shaves 0.03 percent off the cycle time, the vendor still waits 22 days, and the business keeps paying with a corporate card to get around procurement altogether. That's the cappuccino use case again, just measured more honestly. Saving eight minutes in a 22-day process leaves you with a 22-day process. Nearly all of that time sits in handoffs and in waiting for an answer, and none of it is on the screen where someone could demo it. The third reason belongs here too. Without a baseline from before the build, nobody can say afterwards whether anything got better. That's the core of the ROI post, and it holds for a single process just as much as for a whole Copilot rollout. ## A bot waits, an agent goes and gets the work The second reading is about the agent. What does it do without anyone asking it? An FAQ bot waits until someone types, and the title of the November masterclass was about exactly that step, from a bot that answers to an agent that acts. Kornum describes the same split. Sidekicks are, in his experience, used by around ten percent of people and bring improvements of ten to twenty percent, while background agents reached 60 to 90 percent in the projects he's seen. Sidekicks also tend to need their own app and their own login next to SAP, ServiceNow and Outlook, which is where the work actually happens. In Microsoft 365 the background part has a concrete shape. An agent in Copilot Studio can start from an [event trigger](https://learn.microsoft.com/en-us/microsoft-copilot-studio/authoring-trigger-event) instead of a chat message, say a new item in a SharePoint list or an incoming mail, and work through the normal case on its own. The exceptions go to a person. And that's where the first reading comes back in, because to decide what the normal case is and which exception goes to whom, you need to know how the process runs today, and the handbook won't tell you. ## Find out before you build Kornum's answer to how you find out is unspectacular. Talk to the people who run the process every day, and mine the event logs, from the ERP or the ticketing system. In the Microsoft world, [process mining in Power Automate](https://learn.microsoft.com/en-us/power-automate/process-mining-overview) is built for exactly that. It reads event logs and draws the process as it really runs, with every loop and every detour. I've seen how far apart the two can be long before anyone talked about agents. Back when the word was digitalisation, we looked at a process in a company where everyone involved worked at the same site, and we counted three media breaks. Documents arrived digitally and still got scanned in, were processed in the system, and at the handover to the next department turned back into paper, handed over in a folder with a sticky note on the outside. That sticky note was what actually steered the work, and the process knew nothing about it. Process mining wouldn't have found it either, because a sticky note doesn't write an event log. You only find it by sitting next to the people who do the work. On the masterclass slide we closed the question with one sentence, and I'll keep it here. If you only answer the second reading, you automate the handbook and not the work. ## What the answer is worth Answering the question costs time with the people who do the work and a look at the logs, all of it before anyone opens Copilot Studio. Kornum's numbers at the end of his article show what it can be worth. With all seven reasons solved, vendor onboarding went from 22 days to four or five, and people touched only about 15 percent of the vendors, the ones that needed a judgement call. The gain shows up in the cycle time, which is where the business was waiting all along.