Elliot Katz

The through-line

The single idea running through Elliot Katz's work on AI is that the value sits in the seam between machine output and human judgment, not in removing people from the process. He co-founded Mixus around the argument that businesses chasing fully autonomous agents are building on an unreliable base: "businesses right now are trying to deploy fully autonomous AI agents, but the problem is that AI is highly mistaken, right? And so fully autonomous AI agents are highly mistaken" [1]. His answer is what he calls colleague-in-the-loop agents, where a person or several people "verify or correct AI outputs before they cause harm" [1]. The company name itself carries the thesis: for businesses to deploy AI agents safely, "they need to mix us," a play on mixing artificial and human intelligence, and the logo, which he calls the lumen, is "a literal human in the loop" [1]. The same argument holds in his account of legal work, where AI agents "can do so much of the work. But, they can't do all of the work" [3].

The second, older strand of his output is about men and leadership in relationships, built on a book that grew out of his own divorce and his search for what he had failed to understand [2]. Different subject, recognisable temperament: someone should step forward, make the call, and own the outcome rather than leave it to run unchecked.

On why fully autonomous agents are the wrong bet

His case against autonomy starts with error rates from the model builders themselves. He points to an OpenAI benchmark built to evaluate the ability to answer straightforward fact-seeking questions, on which the o3 model "hallucinated over 50% of the time, meaning every other answer was incorrect," and the o4 mini model performed worse still, fabricating answers "nearly 80% of the time" [1]. The rate matters less than where the models are being pointed. Being wrong is harmless if a consumer "just wanted to write a poem about my dog," but "these models are being deployed by businesses today in AI agents that power customer service and decision-making systems, right? Places where being wrong is of consequence" [1]. Left fully autonomous at current hallucination rates, agents produce what he describes as massive undetected mistakes [1]. The bargain he proposes is to keep "the full power of AI agents, the efficiency and the time savings" while removing "that downside risk of AI mistakes going undetected" [1].

On colleague-in-the-loop as an operating model

The mechanism is a verification gate placed inside a multi-step agent rather than a review bolted on afterwards. His worked example is an agent that reads call transcripts, drafts a customer follow-up email, and sends it: the drafting is step two, the sending is step three, and between them the agent stops until a named human approves or edits [1]. Verifiers do not have to be the person who built the agent. They can be a boss, a teammate who was on the same call, or several colleagues at once, and if a tagged verifier is not in the product at that moment they "instantly get an email notification" so a reporter on deadline can get an editor's blessing and move on with the story [1].

He is emphatic that the drafter is often not the person best placed to check the work. He runs Mixus as the business co-founder alongside his technical co-founder Shai, and when he drafts follow-ups after sales calls that touched engineering questions, he tags Shai as verifier "to make sure that the AI is actually getting these technical questions correct and that anything that I'm adding in quite frankly is correct as well" [1]. The verification is also a record. Once a human signs off, the step is marked in the chat, which means "you have traceability. You have auditability," and if something goes very right or very wrong you know who was supposed to check it [1]. Verifiers and colleagues can also talk to each other and to the AI inside the same chat to fix a problem before anything ships [1].

On operationalizing oversight

What his early customers actually worry about, he says, is putting AI into the hands of people at their company who have little experience with it "and then not really having a mechanism for oversight" [1]. His pitch is that Mixus lets a company "operationalize that oversight" as a rule rather than a hope: before anything is published, a designated editor or colleague has to press verify, confirming that what the AI produced is real "and not total slop" [1]. He uses the Chicago Sun Times summer reading list as the canonical failure, where a syndicated writer asked AI for fifteen good summer reads, checked nothing, and ten of the books did not exist, including one attributed to Isabel Allende that anyone reasonably informed would have flagged instantly [1]. He does not treat the task itself as unsuited to AI. A human does not know every book worth reading, and a model can genuinely help with that; the failure was the missing gate, not the request [1]. Asked whether it would have been simpler to Google the fifteen titles, he concedes it might, and reframes the point: the value is not the individual check but a repeatable organisational mechanism, so that a company knows it is "just not letting these autonomous AI agents run wild" [1].

On making agents buildable by people who have never used AI

He treats accessibility as part of the safety argument, since oversight only works if non-specialists can both build and inspect the agents. A common first reaction he gets is "what is an AI agent? I don't even know," and his answer is that "someone who's never used AI, someone who doesn't even know what an AI agent is" can build and run them, because agents are described in plain English and, "as long as you can read and write," the platform assembles the multi-step version in seconds [1]. He demonstrates this with a research agent that pulls the latest news on AI in journalism and PR, generates three podcast topics from it, and emails the summary, with the option to schedule it to run weekly, which he describes as being like an intern agent [1].

A second demonstration covers a reporter's whole workflow: checking what Waymo has said publicly about human assistance and what others have written, surfacing autonomous vehicle experts as potential sources, drafting personalised LinkedIn outreach that references each expert's institution and book, then consulting a stored PR contacts document to see whether the newsroom already knows anyone on Waymo's PR team, and finally drafting an outreach email that waits for verification before it sends [1]. Documents stored in the platform can be referenced by the AI, and the point of the sequence is that work which would otherwise take an hour or several hours for a complex story collapses, with a human gate at the moment of external consequence [1]. He also notes that the company has moved from B2C to B2B [1], and describes the platform as spanning organisations where colleagues hold seats and can be tagged as verifiers [1].

On why law needs both intelligences

Legal work is where he draws the boundary most sharply. AI agents can do amazing things for attorneys and handle the basics well, but "there's so many things that you need human judgment for," because "typically people don't come to lawyers for just straight-up black letter law. They're looking for their interpretation, their judgment, their assessment of risk" [3]. Mixing human intelligence with artificial intelligence is, in his framing, "the way that you're going to get a complete product" [3].

On leadership, decisiveness and responsibility in relationships

His book argues that what a woman most wants from a man is leadership, and that a man who defers every choice makes her feel "like he is a child and she is his mother," when "she wants a husband not a child" [2]. He is careful to separate leadership from domination: "controlling and being a leader are opposites," and overruling a decision his wife has already made purely to demonstrate authority is "undermining her and causing conflict," not leading [2]. Leadership is instead knowing what is going on in your home and stepping forward where a situation would benefit from it [2]. The everyday failures he cites are small: a man who asks a woman out and then asks her where to go, or answers "whatever you make is fine" when asked about supper [2]. Planning a date, by contrast, is what makes a woman feel special, because he took the time to think and decide [2].

Responsibility is the other half of it. If a man sees something going wrong and lets it run, the blame is his, and he tells the story of a divorced man complaining that his wife ran up $50,000 in credit card debt, asking how he let it go on when there were children involved [2]. Nobody will accept the excuse that a spouse pushed you into it [2]. On money he advises living within your means, on the grounds that financial stress can wear a marriage down and destroy it, and that peace at home beats a big house with payments you cannot meet [2]. On shared goals he warns couples not to assume agreement about where they will live, how they will raise children, or how religion and community will feature, because these unspoken assumptions cause the big problems [2]. Communication, in his account, is frequently just complaining under another name, and the discipline he learned is to communicate with someone the way they need to be communicated with rather than the way you feel like saying it, which he treats as an aspect of leadership rather than a separate skill [2].

The book came out of his own divorce, when he stopped blaming the other person and asked what he had to learn [2]. He found nothing useful in contemporary relationship books and turned instead to older teachings that men once passed to younger men, which matched what he heard women say is lacking today: leadership, decision-making, taking responsibility [2]. His explanation for the confusion is generational, that many men grew up without fathers or with fathers who worked long hours [2]. He notes that women are the book's biggest fans and the ones asking how to get their husbands to read it, while men have to be talked to, "because they don't realize what they don't know" [2].

Takeaways

• The case against fully autonomous agents is empirical: on OpenAI's own fact-seeking benchmark, o3 hallucinated over 50% of the time and o4 mini fabricated answers nearly 80% of the time, and those same models power customer service and decision-making systems where errors have consequences [1].
• Put the human gate inside the workflow, not after it: an agent should pause between drafting and sending so a named colleague can verify or correct the output "before they cause harm" [1].
• The best verifier is often not the person who ran the agent; tag the subject-matter expert, as he tags his technical co-founder Shai on engineering answers in sales follow-ups [1].
• Verification doubles as an audit trail, so a signed-off step shows who approved what and gives you traceability when something goes wrong [1].
• The Chicago Sun Times summer reading list, where ten of fifteen books did not exist, failed for want of a mechanism rather than because the task was unsuited to AI; the fix is a rule that an editor must verify before anything publishes [1].
• Adoption depends on accessibility: agents are described in plain English and built by anyone who "can read and write," including people who do not know what an AI agent is [1].
• Law resists full automation because clients want interpretation, judgment and risk assessment, not black letter law, so a complete product requires both intelligences [3].
• In relationships, leadership means stepping forward and deciding rather than pushing every choice back onto a partner, and "controlling and being a leader are opposites" [2].
• Financial stress can destroy a marriage, so live within your means, and talk through where you will live, how you will raise children and what community you will join before assuming agreement [2].