Confessions of Customer Zero | Agent Operation Centre

By Katy Rumbelow | Published On: 28 August 2026 |

By Louis Arthur Brown (Head of Agentic Operations)

Before we sold the Agent Operations Centre to anyone else, we ran it on ourselves. Same tenant, same tools, same discipline we were about to ask customers to sign up to. If it wasn’t going to work for us, we weren’t going to inflict it on anyone else.

Some time in, we have 553 agents under management running across Copilot Studio and Power Platform. Of course, that number includes the bundled third-party agents which are available via Copilot connectors, but even taking those out, we still had over 300 proper agents. That number surprised me the first time I saw it, and it should probably surprise you too. We are not a huge company. If we have over 300, what do you have? And this is what sparked Agent Operation Centre and the importance of it not just for us, but for everyone.

What it’s taught us

Here is what running as Customer Zero has actually taught me, told honestly, because if I only tell you the wins you will not believe any of it.

We didn’t know what we had. We thought we did. We were wrong. When we ran our first proper discovery pass, we found agents built by people who had left, agents duplicating each other, agents that had been “temporary” for eight months, and a handful that nobody could explain at all. The inventory problem is not a theoretical one. It is the first thing that hits you when you turn the lights on.
Ownership was the real gap, not tooling. The Microsoft governance planes are actually pretty good. Agent 365, Defender for Cloud Apps, Purview DSPM for AI, the Power Platform admin centre. What none of them do is tell you who is accountable when a specific agent starts behaving oddly. That is a human problem, and we had to solve it with a named sponsor, an audit trail, and a monthly rhythm before anything else made sense.
Cost crept up quietly. We caught a Copilot Cowork session that would have consumed £2,500 in a single sitting (but luckily we were in the Frontier early access programme, so this was free at the time). Nobody had done anything wrong. It just did what it was asked to do, at the price the meter was running. If you have not looked at your per-agent spend recently, you probably have a version of that story in your tenant right now. Multiply it out and it stops being funny.
Blocking six agents was the point. In our most recent review cycle we blocked six agents. Not because they were malicious, but because they were unnecessary, unowned, or overlapping with something better. If your governance function is not comfortable saying no, it is not a governance function. It is a rubber stamp.
The EU AI Act is not a future problem. We classified our own estate against Annex III risk categories, GPAI (General Purpose AI), and the exercise was more useful than I expected. It forces you to describe, in plain language, what each agent actually does and who it affects. Half the value is the assessment, not the certificate at the end.
Reporting only matters if someone acts on it. We now produce a monthly report that includes a narrative, key points, and recommendations. The first two we ignored because the operating rhythm around it did not exist yet. Reporting is a symptom of good governance, not the cause. You need the advisory call, the decisions, and the follow-through, or the PDF is just a nicer way of doing nothing.

Being Customer Zero changed the product.

The bundled tier model we sell today looks nothing like the Agent Operation Centre that we drafted in April. Two axes (Coverage and Service Depth), a Critical Agent concept, a Watchlist add-on for the agents that carry real weight. Every one of those decisions came from something we hit ourselves and didn’t want customers to hit blind.

So what would I do differently if I were starting again? Firstly, I would build the operating rhythm before the reporting layer. Secondly, I would name the owners before I bought tooling. Furthermore, I would treat the first ninety days as a discovery exercise and set expectations accordingly, rather than promising a governed estate on day one. And I would take the EU AI Act classification seriously, sooner than felt comfortable.

We built the AOC as a managed service because governance is a habit, not a project. Software will not save you. A dashboard will not save you. What saves you is somebody looking at the estate every day, asking the awkward questions, and making sure the answers are written down.

If you want to see what your own version of this looks like, we run Preview sessions. No slideware. We show you what is actually running in your tenant and where the gaps are. That was the first thing we did to ourselves, and it is still the most useful thing we do for anyone else.