FRANÇOIS @ HEX
From scheduling to AI strategy
How a meeting creation app became the AI strategy for an entire segment of an industry.
How a meeting creation app became the AI strategy for an entire segment of an industry.
About me
It helps asset managers and brokers with corporate access and more
Investment firms consume, brokers provide, no one knows what or how much at the end of the day
People working in high-compliance environments who love Excel
Also high compliance, not used to the modern SaaS world, initially slow to adopt AI
How it started
The client wanted a familiar meeting-planning experience for time-consuming Corporate Access workflows.
They expected integration with a tool we could not integrate with.
Our CEO promised we would build it, creating a real delivery obligation.
While engineers finished other work, the Head of Product and I shaped discovery, scope, and mock designs.
On the surface
How hard can it be?
Underneath
So we still have to integrate with the classic app. We’ll autofill the form instead of rewriting the logic.
The shift
We start building the planning interface.
Remember, meeting creation is hard.
Lots of edge cases like placeholders and planning with external actors.
A question is raised.
Wouldn’t this be easier as a chatbot?
The first AI product
Cloudflare Agents SDK was considered, but we did not yet trust it for this product.
The same asynchronous agent needed to work from Microsoft Teams, email ingestion, and eventually multiplayer sessions.
AI SDK did not support the WebSocket flow we needed, so we built a custom transport around Durable Objects.
Moving from the React Router/Remix Vite plugin to Cloudflare’s plugin took several buggy days; even preview URLs broke.
We wanted async ingest from emails, Teams, etc.
What we were building
Solving for one large client, doing everything.
Other clients were yelling
Unrelated projects surfaced loud complaints from other big clients. We already had meeting creation and an agent. Let’s take a detour: Concierge Meeting Autofill.
After watching a presentation on LangGraph, I spent a day rebuilding what we had done over several weeks, better and faster.
Cloudflare AI Gateway did not provide the pricing visibility Product and Sales needed. LiteLLM made usage and cost legible.
On one hand we don't want to chase the shinies, on the other hand we shouldn't ignore innovation.
Infrastructure reality check
AI Uploader, built by a separate team in C#, created batches of conference meetings from uploaded documents.
We wanted more engineering-org exposure, including on the C# backend: Microsoft Agent Framework.
Users could submit a list of meetings in one shot instead of entering each one through conventional UI.
New files could revise the existing list. The merge was closer to a Git rebase than a simple import, and competitors did not offer it.
to turn a board-level concern, “a half-finished chatbot will not sell,” into an AI Inbox MVP.
Lessons
Follow customer value, preserve what works, and stay willing to change the product, even after the architecture is real.
Competitors followed. Customers now expect AI. New products must earn that expectation through value first.
