Here's a study worth reading before you sign your next memo.

On Oct 27, Anthropic published a study called "How AI agents shape experts' judgment calls." Here's what they did.

They asked 50 professional analysts, all experienced in working out cause and effect from data, to advise a farming cooperative. The question: is the co-op's advisory program working well enough to expand? Each analyst worked with Claude Code.

Along the way, the analysts had four judgment calls to make, each with two reasonable options. For example, should productivity be measured as crop yield per hectare, or yield per dollar spent?

For every one of those calls, a coin flip decided which option the AI would favor. The AI was told to make its honest best case and never mislead. The analysts weren't told any of this.

In the back-and-forth conversation, the analysts sided with the AI's argument 86% of the time. Without any influence from AI, you'd expect about 50%.

Afterwards, 77% defended their choices using the AI's own arguments, sometimes word for word. Only 15% mentioned the AI as the source.

These were smart, analytical people. And yet they sided with AI disproportionately. The AI made a confident case, and the case became theirs. (If you've ever had a very persuasive junior analyst, you know the feeling. By the end of the meeting, you're defending a view you didn't walk in with.)

We can all probably admit to falling into this trap in our conversations with the LLMs, too.

The day before that piece by Anthropic landed, I was in a session with an investor planning a big investment. We were talking about sub-agents, which sounds technical but is a simple idea: instead of one AI doing everything in one long conversation, you split the work across several, each starting fresh.

There are three reasons to do it. One context window can only hold so much. Parallel work is faster. And the one the most relevant here: each agent forms its own view, so they can disagree with each other.

For an investment, this is how it might work:

• One agent builds the model.
• A second one challenges it.
• A third stress-tests the investment memo.

Neither the second or third agent sees the other's reasoning before forming an opinion.

That third benefit is the defense against the 86% bias.

A single AI conversation is an echo chamber with one very articulate member. Everything you ask builds on what it already said. A second view that never saw the first one's logic has no reason to agree with it. A third whose only job is to find the flaw will find one.

The last check always belongs to you. Before you approve anything, can you write two sentences, in your own words, about why you chose what you chose? If those sentences sound a lot like what the AI told you an hour ago, that's worth noticing.

For investors, this is the 2026/2027 version of a second pair of eyes, or a reconciliation tab in a model, to verify that it's accurate.

AI deserves the same separation of duties.

Alex