Here’s an unpopular opinion for a site about AI agents: the most important part of pitching an AI company in 2026 is still five people sitting in chairs, deciding whether they believe you. No model. No scoring algorithm. No agent parsing your deck for signal. Just humans, listening, in a room in San Francisco.
TechCrunch just revealed the next wave of venture capitalists judging Startup Battlefield 200 contenders at Disrupt 2026. The names: Nell Daly, Michael Palank, Aditi Maliwal, Chrystal Huang, and Grace Ge. They’ll spend October 13-15 in San Francisco evaluating startups pitching live.
If you follow AI the way I do — trying to explain it to people who’d rather not read a research paper — that detail is more interesting than it looks.
Why a Judging Panel Matters to Non-Technical Readers
A lot of AI coverage right now suggests that judgment itself is being automated. Agents review resumes. Agents summarize meetings. Agents draft the memo. The implication is that evaluation is a solved problem, or close to it.
Then you look at how actual money gets allocated at the earliest, riskiest stage of company building, and the process is almost stubbornly analog. Founders get a few minutes. Investors ask questions. Somebody decides. The whole format rests on a thing agents are genuinely bad at: sitting with incomplete information about a company that has no history, no revenue pattern, no comparable case, and forming a view anyway.
That gap is the single most useful thing to understand about AI agents today. They’re strong where there’s precedent and weak where there isn’t. Early-stage investing is almost entirely the second category.
What the Panel Format Actually Tests
Live pitching is an odd ritual, and it survives because it measures things a document can’t:
- Whether the founder understands their own product when the questions get specific
- How they respond to being wrong in public
- Whether their conviction holds up without their slides
- What they choose to say when the clock is running out
An agent can prepare a founder for all of that. It can rehearse questions, tighten the narrative, and stress-test the numbers. What it can’t do is be the person on the other side who has seen two hundred versions of this pitch and has a feeling about the two hundred and first.
The Honest Version of the AI-and-Judgment Story
I get asked variations of the same question constantly: will AI take over decision-making? The most accurate answer I can give is that AI is very good at taking over the work that surrounds decisions, and much slower at the decisions themselves.
Think about what an agent can plausibly do for an investor: pull together background research, track which companies in a space raised what, flag inconsistencies in a deck, draft follow-up questions, organize notes afterward. All real, all useful, all preparation.
Now think about what remains. Someone has to weigh a founder’s stubbornness against the market’s indifference and decide whether to write a check. That’s a taste question wearing a data costume. It’s the part that stays human not because of some rule about human specialness, but because there isn’t enough structured signal for a model to work with.
Startup Battlefield is a clean illustration of that split. Judges named, dates set, room booked. The evaluation happens in person because that’s where the information is.
If You’re Building Something and Watching From Outside
The practical read for founders, especially first-timers who feel locked out of these rooms: the tools you use to prepare have gotten dramatically better and cheaper. Research, competitive analysis, rehearsal, refining how you explain a technical product to a generalist — all of it is more accessible than it was even two years ago.
What hasn’t changed is the last few minutes. That’s still on you.
Disrupt 2026 runs October 13-15 in San Francisco. Registering early saves money on tickets, which is a small detail that matters more than it should when you’re running on a seed round or nothing at all.
My Actual Take
I write about AI agents for people who don’t build them, and the framing I keep returning to is this: agents are extremely good assistants and unconvincing deciders. Every time a high-stakes process gets examined closely, that pattern shows up again. Hiring panels. Medical second opinions. Grant committees. Five VCs judging pitches on a stage.
The automation story is real, but it’s happening in the layer underneath the decision, not at the decision. Understanding that distinction is worth more than any prediction about what models will do next year.
Five investors. Three days. A few minutes each. Still the format we trust with the risky money, and that tells you something worth knowing about where AI actually sits.
🕒 Published: