From proposal to project: how Cowork takes over the handoff

At Blis Digital we write a lot of proposals. A client gets excited, signs, and then the real work starts. But there’s a stretch in between that nobody really enjoys. Someone has to open the proposal again, dig out the deliverables and tasks, figure out who is on the team, plan a kickoff that fits everyone’s calendar, write an agenda, and pass the project details to finance so it can be set up in the system.

All of that is useful, all of it is repetitive, but also not super straightforward to automate, and all of it adds just enough delay between wining a deal and getting started.

That’s exactly the kind of work where I think: a good junior virtual colleague could take this off my plate. So enter the Agent Academy hackathon, and the Cowork Collective track in particular, which was the perfect excuse to stop talking about what AI agents could do and actually build something.

The trigger: a hackathon as a nudge

Microsoft’s Agent Academy hackathon, and the Cowork Collective track in particular, was the perfect excuse to stop talking about what AI agents could do and build something. The brief was simple: make something practical with Microsoft 365 Copilot Cowork, record a demo and share the results.

Ship it!

We translated the whole kickoff prep into a single Cowork skill. You hand it the proposal deck, and it works in two phases.

The first phase is read-only. The skill reads the proposal, summarises the project, pulls out the deliverables and tasks, and looks up the people mentioned in our tenant. It shows you what it found, and anyone it can’t find, the client contact for example, it lists separately so you stay in charge of that call.

The second phase only runs after you say yes. It finds the earliest slot where everyone is free, sets up the Teams kickoff, writes an agenda based on the deliverables, saves it as a Word document, and sends a structured intake email to finance.

That whole process used to take a couple of days, spread across Outlook, Teams, Word and email. Now it’s a matter of minutes, with a human pressing “approve” at every step. And that last part is not a detail.

If you want to watch a 5 minute demo checkout the video:

Trust is good, control is better

The skill is built around “propose first, act second” from beginning to end. Nothing gets sent or created without your explicit approval. And the skill makes nothing up. No people, no dates, no deliverables, no email addresses. Anything that isn’t in the proposal or can’t be found in the tenant gets a clear placeholder instead of a guess.

Does this work everywhere? It depends 🤷. For a structured, recurring handoff like this one, it’s a great fit. For work where the context is completely different every time, you will need other skills and other approaches, so start experimenting and find what works for you!

Where it got tricky

It didn’t run smoothly right away, I’ll admit. Two things worth passing on.

Recognising people is harder than it looks. An external client contact isn’t in your tenant, so you can’t simply invite them, and the invites bounce. The fix became part of the design: show the unknown names separately and let the user decide whether to leave them out or supply an address.

A long, multi-step run needs patience and transparency. Our first instinct was one giant prompt that did everything at once. The lesson was to keep the skill tightly scoped and to use the fact that you can steer mid-conversation. One targeted correction fixes a single output without redoing the rest.

The biggest lesson, oddly enough, wasn’t in the technology but in the instructions. The behaviour lives in the text of the skill, not in code. Good instructions, like “use only what’s in the source, invent nothing,” plus a hard split between the two phases, did more than any clever automation.

What I take away from it

The speed is still the biggest surprise. That we got something usable up and running in a few days is something I wouldn’t have thought possible a couple of years ago. And starting small works. A familiar, boring task makes an excellent first case. No hype, just time saved that you feel tomorrow.

AI is here to stay. Go use it. Not as a magic cure, but as that little helper that takes the groundwork off your hands while you keep your hand on the wheel. Start small, pick a task you curse every week, and build something that waits for your “yes.”

Want to see the skill and the demo? The code and write-up are here at GitHub - AppieSchot - Client-kickoff-skill or reach out and let me know. Happy to share and chat about it!