Every process in your organisation has two versions. There’s the one written in the SOP, and there’s the one your team actually runs. The evidence for the second version is already sitting in your emails, Teams chats and meeting transcripts. Copilot Cowork can now read both, lay them side by side, point to the source of every difference, and help you turn what the team agrees into an updated procedure.
That’s a bigger shift than it sounds. For the last few years we’ve mostly asked AI to help one person with one task: summarise this document, draft that email, analyse this spreadsheet. Useful, but it stops at the edge of the task. This is AI working across the systems and conversations that make up an actual business process, with people deciding what happens next.
Why do documented processes drift from reality?
Nobody sets out to ignore the SOP. Processes drift for ordinary reasons. A role gets restructured and the approver named in the document no longer exists. Someone finds a quicker way through a busy week and it quietly becomes the norm. A threshold gets changed in a meeting and nobody updates the document afterwards.
None of that shows up in the SOP. It shows up in the conversations around the work: the email that says “just approve it, I’ll sort the paperwork”, the Teams thread where someone asks which form to use, the meeting where the team agrees to do it differently from now on.
That gap matters more than it used to. When people rely on the written process for compliance, training or audit, a document that doesn’t match reality is a risk. And when Copilot and agents use your documents as their source of truth, an out-of-date SOP gets repeated back to people with confidence.
Where does the evidence of the real process live?
The useful part is that you don’t need to interview everyone to find the real process. Most of it is already written down, just not in the SOP.
- The SOP itself, usually a Word document in SharePoint.
- Email, where approvals, exceptions and workarounds get agreed.
- Teams channels and chats, where people ask how things are done and get told.
- Meeting transcripts, where the team talks through what’s working and what isn’t.
This is process discovery, the job that process mining and process intelligence tools do from system logs. The difference is that it works from the conversations around the work, which for most Microsoft 365 organisations is where the real process actually gets recorded.
Copilot Cowork can read across all four within the permissions of the person running it, which is the important bit for anyone security conscious. It only uses what that person can already see. Point it at a process, ask it to work out how it really runs from those sources, and compare that with the documented version.
How Copilot Cowork compares an SOP with how work really runs
In the video I ran this on a supplier onboarding process in a demo tenant. Cowork read the SOP in SharePoint, worked out the real process from Teams, email and a meeting transcript, and built both versions side by side on a Miro board using the Miro plugin. Every difference got a sticky note with the source it came from. The parts that matched were called out too, so verifying ABN and insurance came back green.
It also produced a variance register, a table of every gap rated high, medium or low, so the team could work through them in order. Then the team commented on the board, agreeing some changes and pushing back on others.
The second half is where it becomes useful rather than just interesting. I asked Cowork to read those comments and draft the next version of the SOP in Word, keeping the original structure, applying only the changes the team agreed, and listing anything still undecided as an open item. It came back as a version 3 draft. Finally, I compared the old and new versions in SharePoint and generated an executive change report showing what changed and why.
The first pass, the comparison and the board, used 371 Copilot Credits. At pay-as-you-go rates of one cent a credit, that’s about US$3.71 to map one process against reality.
Four rules for trusting AI with a business process
Once AI is pulling from emails, chats and meetings to work on an actual process, the question changes from “can it do this?” to “can I trust what it did, and can I show the working?” These are the four things I’d hold any of this work to.
1. Every finding has a source. If AI says the team runs something differently from what’s documented, it should point to the email, chat or meeting it came from. A finding you can’t check is an opinion.
2. It tells you what’s working, as well as what isn’t. A report that only lists problems gets treated as a witch hunt. Calling out the steps that match the SOP gives people confidence the rest is fair.
3. People make the call, and anything undecided stays open. AI finds the gaps. The team decides which ones are fine, which ones need fixing and which ones need more discussion. The draft should reflect those decisions, not guess at them.
4. Every change leaves a trail. A new version of the document, the original kept, and a plain report of what changed and why. That’s what makes the result defensible when someone asks later.
Which processes should you check first?
Supplier onboarding is just the example. The same pattern works anywhere a process has drifted from what’s on paper:
- Approvals, where the real approver or threshold has changed.
- Client intake, where the team has added steps or skipped them.
- Month-end, where workarounds have built up over time.
- Any process heading into an audit, where you’d rather find the gaps yourself.
At a few dollars a run, it’s cheap enough to do regularly. Running your most important processes through it once a quarter would catch drift before it turns into an incident.
The specific process isn’t really the point. The pattern is: AI working across application boundaries, coordinating multiple steps, keeping people in the decisions, and producing something useful at the other end. We’re moving from AI that helps us do tasks to AI that helps us coordinate work.
If you’d like help finding which of your processes would benefit most, that’s part of what the AI Workforce Assessment maps. For more on how Cowork’s costs add up, see Copilot Cowork models: what each one actually costs.
Frequently asked questions
What is Copilot Cowork? Copilot Cowork is the part of Microsoft 365 Copilot that carries out multi-step work for you, rather than answering a single question. It can read across your Microsoft 365 content, use plugins to work with other tools, and produce documents, reports and other outputs. Each run uses Copilot Credits.
Can Copilot Cowork read my emails and Teams chats? It works within the permissions of the person running it, so it can only use emails, chats, meeting transcripts and documents that person can already access. It doesn’t give anyone visibility they didn’t already have.
How much does it cost to compare a process with Copilot Cowork? In my demo, the first pass that compared the SOP with how the process really ran and built the board used 371 Copilot Credits, about US$3.71 at the pay-as-you-go rate of one cent a credit. The cost will vary with the size of the process and how many sources it reads.
Does Copilot Cowork update the SOP automatically? No. It drafts a new version based on the decisions the team made, keeps the original, and lists anything undecided as an open item. A person reviews and approves the new version before it replaces anything.
Do I need Miro to do this? No. I used the Miro plugin because a board makes it easy for a team to review and comment together. You could ask Cowork for the comparison and variance register as a Word or Excel document in SharePoint instead and collect comments there, although that isn’t what I show in the video.
