I Built a Negotiation Workflow with Microsoft Copilot Studio — The Honest Version

Dear Negotiation Explorer,

Over the past few days I built a negotiation workflow inside Microsoft Copilot Studio — two AI agents that work in turn to prepare you for a real negotiation. It works. It was much harder than it should have been. And you almost certainly have the same tool I used, sitting inside your company's Microsoft account.

Here's exactly what I built, what it took, and whether it's worth your time.

You open Microsoft Copilot in Teams — the assistant you already use. You upload two files: the negotiation case you're walking into, and the LinkedIn profile of the person across the table. You type one line about what you want out of the deal.

A few minutes later, the briefing comes back in two parts:

I checked every number in it against the case file. Not one was invented. With AI that isn't a given — and in negotiation preparation, that's what matters most.

Three parts run in turn. For each agent you write its instructions — what it does — and attach a knowledge base — what it knows. Here, the knowledge is a full negotiation method.

You build all of this in Copilot Studio. Once it's built, it lives inside Microsoft Copilot — the assistant you open in Teams. You build it in one place and use it in the other.

Copilot Studio isn't hidden, but it isn't simply switched on either. In a company, IT controls it. So the first real step isn't building anything — it's asking whoever runs your Microsoft setup to give you access.

That's also where I had it harder than you will. I run a one-person operation, so I had to stand up the whole Microsoft foundation underneath myself, from scratch. You won't — your company already has it. You just need the door opened.

Then the walls. Here's the honest sequence.

Chaining the agents. Copilot Studio has a visual tool for connecting one agent to the next. I tried it. The first agent ran fine; the second ran for zero seconds and produced nothing, and the knowledge each agent was supposed to draw on quietly dropped out. So I stopped trusting it and built the chain by hand — wiring each handoff myself, so I could see and control every step.

The files came in as code. This was the make-or-break one. When you upload a document, it doesn't reach the agent as words — it arrives as a long block of machine characters. The agent reads nonsense and replies, politely, that no file was provided. The brief comes back full of blanks. The fix: add a step that turns each file back into readable text before any agent sees it. Nothing worked until that was in place.

Getting it into Teams. To use it where I actually work, the workflow had to be reachable from Microsoft Copilot. But the moment I connected it, the uploaded files vanished — the system carries your typed sentence and silently drops your documents. So I built a short guided conversation that asks for each file in turn — "upload the case", "now the profile", "what do you want to achieve?" — and carries each one through properly.

The hundred-second wall. A tool that answers inside a chat has to reply within a hundred seconds. Two agents thinking in sequence take longer than that. So the whole thing kept dying on a timeout — every time, after everything else had finally come together. The fix was to stop answering all at once: now it replies instantly — "preparing your briefing, give me a couple of minutes" — keeps working in the background with no clock on it, and delivers the finished brief when it's ready. That wait isn't a flaw; it's how these tools work.

The 8,000-character ceiling. This is the one I couldn't build around. Each agent in Copilot Studio holds only 8,000 characters of instructions. That is not much. A serious negotiation method is far bigger — so I was constantly cutting Deepak's instructions down to fit, and you feel what gets left out. It's the hardest limit in the whole tool, and it doesn't go away.

There were wrong turns too, and I'll leave them in. One was Brama's — my AI partner, Claude Code, which I call Brama. It suggested a setting to make things faster that instead made the workflow forget its place between steps, so the agents started and never finished; it said so plainly the moment we saw it. And at one point I got angry: Brama wanted to step back and build less than the full workflow, and I pushed to keep both agents in. I'd set out to build the real thing, and I wasn't going to shrink it to make it easier.

What it costs you, plainly: more steps than the easy tools — a two-agent workflow is closer to eight or nine pieces wired by hand than the two or three you'd expect. The 8,000-character ceiling on every agent, which you fight the whole way. And for now the finished brief lands in a separate chat rather than the window you asked in — a rough edge I'll smooth. While it runs, you don't see it working.

What you get for it: it's inside your company's walls — your data never leaves, and there's no outside tool for anyone to approve. It lives in Teams and Microsoft Copilot, where your work already happens.

My honest verdict: if you already live inside Microsoft, your team works there, and you trust the environment your files and data sit in, it's worth the fight.

Not a toy, and not for developers only. I built a real two-agent workflow on real uploaded documents — and the same structure takes a third agent without any new trick. No code. What it asks of you is the access, and the patience to push through a build that fights you more than the easy tools do.

And if you'd rather not do it alone — if you want something like this built for your team or your organization — I'd be glad to show you how.

Would you want to build this yourself inside Copilot Studio — or have it built for your organization?

Reply and tell me. I read every response.