Match the tool to the team.

The question I keep getting asked is “which one should we standardise on?” A new class of desktop AI assistant has arrived — tools that sit on your machine, watch what you are doing, open your apps, read your email, and actually do the work rather than just chat about it. Three of them come up again and again: Claude Cowork, OpenClaw, and Microsoft Scout. And the instinct, every time, is to pick one and roll it out to everybody.

I think that instinct is wrong. Not because standardising is bad, but because these three are not really competing for the same seat. They sit at different points on one line — from fully managed at one end to fully self-hosted at the other — and which end suits you depends entirely on who is sitting in the chair.

One spectrum, three points on it

Strip away the feature lists and the difference is simple. It is about how much you want to set up yourself, and how much control you want in return.

Claude Cowork sits at the managed end. Almost nothing to set up, the strongest model out of the box, and it works well whether or not you live inside Microsoft. You open it and it works. The price of that ease is that you cannot host it yourself, and the controls a big company wants — who can do what, what gets logged — are still maturing.

OpenClaw sits at the other end. You can host it yourself, point it at whatever model you like, and bend it into almost any shape. That freedom is real, but it is not free: someone has to do the setting up, and someone has to build the guardrails by hand. It rewards the person who enjoys that work and punishes the person who does not.

Microsoft Scout sits closest to the Microsoft world. If your company runs on Outlook and Teams, it plugs straight in, and the enterprise controls are the strongest of the three. Step outside the Microsoft world and it is only fine, not special — and, like Cowork, you cannot host it yourself.

The tool is not the decision. The person is.

Once you lay it out like that, the choice stops being about the tools and starts being about your people. A finance manager who wants help clearing her inbox and drafting a board paper does not want to configure anything — she wants Cowork. An engineer who wants to wire the assistant into a build pipeline and swap models on a whim does not want a managed box — he wants OpenClaw. A team that already runs its whole day inside Outlook and Teams will get more from Scout than from anything that sits outside that world.

There is no best tool. There is only the best fit between a tool and the person who has to live with it.

This is the same lesson we have learnt with every other piece of technology, arriving again in a new coat. We do not give everyone the same laptop or the same software and call it fair. We match the tool to the work and to the person doing it. AI assistants are no different — except the gap between a good fit and a bad one is wider, because these tools sit much closer to how a person actually thinks and works.

Why standardising too early hurts

The pull towards one tool for everyone is strong, and I understand why. One vendor is easier to buy, easier to bill, easier to support. But pick too early and you pay for it twice.

If you hand the managed tool to your tinkerers, they will feel boxed in and quietly go around you. If you hand the self-hosted tool to people who just want to get their work done, they will never set it up, and your shiny rollout will show up as a flat line on the usage chart. Either way you have spent money and taught people that the company’s AI is not for them. That lesson is expensive to unteach.

The better move early on is to treat the spectrum as a menu, not a verdict. Let a managed tool be the easy default for most people, give the self-hosted option to the few who will genuinely use it, and lean on the Microsoft-native one where the work already lives in Microsoft. You are not buying one tool. You are matching a small set of tools to a few clear types of person.

What to actually do

So before you ask “which tool is best”, ask a different set of questions. How much setup will this person tolerate before they give up? How much control do they actually need — not want, need? How much of their day already happens inside Microsoft? What does our security and governance posture allow today, not in a year? The answers point at the tool far more reliably than any feature comparison.

The market will keep adding tools to this spectrum, and the names will change. The decision underneath will not. Standardising on a single AI assistant feels like leadership, but most of the time it is just buying convenience for yourself and paying for it with everyone else’s adoption. Match the tool to the person, and the adoption takes care of itself.