Your finance agent should be the cautious one

On September 29, a16z published a conversation between Anish Acharya, a general partner there, and David Pawlan, who runs Assistant Benchmark, a public scorecard for the AI assistants you can text. They cover a lot of ground in 50 minutes. The part I keep replaying is about temperament.

Acharya put it as a steelman. For some jobs you want an agent that is "highly presumptuous," one you'd let "make mistakes 10% of the time as long as it often gets things done." Then he named the exception: "our finance agent perhaps where you want it to be highly cautious."

Pawlan agreed and drew the line with an example. If an agent decides your insurance costs too much, it "should probably get your permission before it actually makes that switch." A drafted email or a claimed flight credit needs nothing from you. A new insurance policy is a decision you'd want to make yourself. And his warning about getting that wrong was blunt: "if you cross that line once, I think you immediately lose all trust with your user."

I build Keepfolio, a macOS app for people who run their own brokerage account. That exchange is the closest thing I've heard on a mainstream show to the decision the whole product rests on.

The flight is a cheap mistake

The episode's best story is a failure. Acharya described a post from someone whose agent checked him into a flight and hallucinated his middle name as "Bong Chang." He couldn't board. Acharya wasn't sure the post was real. It barely matters, because the failure is easy to believe.

A missed flight is expensive and annoying. You rebook it. A bad trade in a margin account has a different shape. Fortune's Nick Lichtenberg covered a research proposal in April about exactly this, and one sentence stuck with me. When an agent sits on a brokerage account, "even a single failure can produce immediate, realized loss." Asked who covers that loss today, the piece answered: "Right now, nobody has to."

That's the gap Pawlan was pointing at. A wrong check-in costs you a rebooking. A wrong trade is a realized loss by the time the notification arrives.

What the fine print already says

This isn't hypothetical anymore. Robinhood, Public, SoFi, Coinbase, and Webull have all put AI trading agents in front of retail customers, according to a Finder review updated August 14, 2026. Finder read the disclosures. For Robinhood's Agentic Trading, the summary is short: "Customer assumes all risk for agent-executed trades," and Robinhood "doesn't control, supervise or audit connected agents."

SIPC won't help either. It protects up to $500,000 per customer when a brokerage firm fails. It doesn't cover an investment that loses value, and an agent's bad trade is exactly that. Finder calls it "an investment-decision problem, and SIPC was never built to touch it."

An agent that can trade acts on your behalf, and the disclosures put every mistake it makes on you.

Where Keepfolio drew the line

Keepfolio's brokerage connection is read-only. It goes through SnapTrade and can read holdings, balances, margin, and activity. There is no order path and no transfer capability anywhere in the app, and Keepfolio never takes custody of anything. The read-only connection is the boundary the product is built around.

Inside that boundary the analysis can be as aggressive as it wants. Cassandra, the analyst inside Keepfolio, checks what moved overnight, measures it against the strategy contract you set (a margin ceiling, for example), and writes a morning briefing with the source behind every line. It can tell you margin utilization is creeping toward your ceiling. It can't reduce the position for you, and that's on purpose.

Acharya's framing helps here. The presumptuous part of Keepfolio is the reading: it assumes you'd want to know, and it does the work without being asked. The cautious part is the acting, and there the app does nothing at all. The decision, and the order ticket, stay with you.

The case against

The strongest objection comes from the same episode. Every use case that got the two of them excited involved an agent doing something. Pawlan described a friend who connected his agent to his sprinkler system and the weather, which he said dropped the water bill by 50%. Pawlan has his own assistant contact the airline for a travel credit when a fare drops after he books. That's the magic, and a read-only finance tool leaves the last step to you. It'll always feel less impressive than an agent that just does it.

I think that's right, and I'd still make the same call. Those examples share something: a mistake is cheap to undo. A wrong sprinkler call wastes a little water, and the travel credit comes from an airline that already runs a refund process. Selling, borrowing, and moving cash in a brokerage account can't be undone the same way, and today the disclosures put the error on the customer. Until someone can answer Fortune's question with something better than "nobody," the useful place for an agent in a brokerage account is reading, so the human can act on better information.

The trust you don't get back

Pawlan's line about losing all trust after one crossing is the rule I'd hand anyone building in this space. Software with no order path can't cross it. Keepfolio reads the account and writes a sourced briefing, and any trade that follows is one you place in your own brokerage.