For about eighteen months I watched good teams fail at the same thing. I was not always the one running the pilots, but I sat close enough to read the retros. The shape was identical every time.
A senior partner gets excited about a demo. A small cross-functional group is assembled. A vendor is picked. Six to ten weeks go by. The deck at the end is polite. The pilot produces two or three useful moments and a dozen small, quiet failures. Nothing ships to the wider team. Another vendor shows up with a different demo eight weeks later. The cycle continues.
In October 2025 I stopped being able to ignore the shape.
Inside one week I sat in on four different retros, across four different teams, in four different firms. A legal team. An account team in a consultancy. A post-sales team in a mid-market SaaS. A finance team in asset management. Different stacks, different vendors, different model generations. The write-ups might as well have been the same document.
Each team had chosen to blame a different thing. The legal team blamed accuracy. The consultancy blamed adoption. The SaaS team blamed integrations. The finance team blamed procurement. All four were kind of right and all four were mostly wrong, because the real failure was upstream of every one of those stories. The model, in every case, did not know enough about the team it was meant to serve. It did not know who the client was. It did not know what the team meant by the words it used. It did not know which reviewer to route a draft to. It could not know, because none of that was written down in a form anything machine-readable could reach.
I spent the weekend after that fourth retro trying to write down what all four of them had been missing. I got to five layers — company, team, client, project, user — before I stopped, because the page had started to feel less like a note and more like a product spec.
That was the week. I closed the laptop, wrote a short letter to the person I was working for, and two weeks later I was in Odense in a quiet office trying to build the thing. I have been building it every day since.
I do not pretend this is the only important problem in enterprise AI. It is the one I cannot stop seeing. When I look at a team and their tools, I see the shape of the context they are missing before I see anything else. That has not changed in six months. If anything it has sharpened.
Sempleo is the company that comes out of that week. Everything else on this site — the thesis, the five layers, the MCP choices, the founding-customer cohort — is downstream of it.
If you recognise the shape, and you are on one of those teams, I would like to hear from you. The founding-customer page is the way in. I read every application myself.
