Your AI pilot worked and it still died
Most AI pilots in portfolio companies die for one reason: they were built by somebody who left. The fix isn't better software. The people who will run it have to build it with you.
The autopsy
That's the whole autopsy on most of them. The pilot was fine. The model was fine. The demo was good enough that everybody clapped.
Then the consultant rolled off, or the vendor moved to the next account, or the internal champion changed jobs, and six months later there's a tool nobody owns, nobody trusts, and nobody can fix when it breaks at 4pm on a Friday.
Software that nobody owns isn't an asset. At exit it's a diligence finding.
Who should build it?
The people who are going to run it have to build it with you. Not be trained on it afterwards. Build it.
It's slower for about three weeks and then it's faster forever.
If you can't name the person who owns the thing after you leave, you haven't finished.
Short answers
Why do AI pilots fail after the pilot works?
Because the person who built it leaves. The consultant rolls off, the vendor moves on or the champion changes jobs, and the tool is left with no owner. Nobody trusts it and nobody can fix it.
Who should build AI tools in a portfolio company?
The people who will run them, working alongside whoever brings the skills. Training them afterwards is not enough. If you cannot name the owner after you leave, the project is not finished.