Nico Lafakis began his turn on Value-First AI Daily, Episode 43, by saying what he was not bringing: a full application. What he brought, in his words, was “this critical thinking idea that you probably didn’t think of and haven’t seen before.” (01:26 to 01:34)
The idea was an exploded view, the kind of drawing that shows a machine with its parts pulled apart so you can see how they fit, applied to a business, a department or a process. Most of the hour was the result: diagrams built with Claude that open when you click them. This is written for someone who did not watch. It follows what the two hosts, Nico and Chris Carolan, said about those diagrams and what they are worth, and it ends with the sealed Top 3, which the hosts did not get to on air.
Episode — Value-First AI Daily
Taking the Business Apart to See How It Works
Episode 43, October 22: two hosts and no guest. Nico shows what he built with Claude, and Chris asks what makes a build land with someone other than its maker. The recording runs about 62 minutes.
Open the episodeA department you can look inside
Nico’s case for the picture began with a claim about the moment. “We’ve never industrialized knowledge work before,” he said (04:38), which is why, in his view, borrowing the factory’s vocabulary makes no sense: “Calling an AI system an industrial tool is, let’s stop.” (04:47) Chris added from the business side that the language people lift from production lines for marketing and sales assumes a predictability that people do not have (05:12 to 05:22).
The first diagram was the sales process of a retained executive search firm, drawn as an engine. Its sections pull apart and open part by part, and Nico clicked through several of them on screen. He described the purpose as being to “solve that problem of not being able to see into the stack of the department” (11:15 to 11:24): one organization’s silos, and what each person contributes inside them, laid out so everyone looks at the same thing. He called it a “new way of computing, man. New way of presenting data, new way of getting people on board.” (11:47 to 11:55) Early on he cautioned that he might be showing an older draft of the engine (08:50 to 08:55).
Two more drawings followed. A marketing organization built as a car, chassis to tailpipe, sat on screen for a stretch while Chris talked, and nobody on the recording explained it on its own. HubSpot’s lifecycle stages and statuses, drawn as a pipe that runs from inbound through marketing and sales to service, did get a walk-through. Nico framed it for someone handed HubSpot administration with no warning, and argued that a stage set wrongly early in the pipe throws off everything after it (26:02 to 29:22). His advice for anyone stuck on how a thing works was to ask yourself “What does understanding better look like to you?”, put that to Claude, and “just step back.” (29:57 to 30:04)
What the build leaves out
Chris met the engine with a test he applies to anything published. Three questions, in his order: can the thing be found, can it be taken in on the reader’s own device and in their own situation, and can the reader make sense of it (15:32 to 17:01). For the third, he said, he usually asks an AI. So he had given Nico’s link to Nectar, his AI assistant, without preparing her beforehand, and he read her answer aloud (18:22 to 19:51). Her reply treated the engine as a way into a process and tied it to his three questions. Chris’s own remarks run into hers with no marked edge, so we paraphrase her and quote none of it. What he took from it he said himself: “I didn’t give any other context to my AI assistant, and it just crushed it.” (21:19 to 21:28)
Then he said what that result did not prove: “the building’s the easy part, but that’s not the value.” (22:32 to 22:44) He had made his own copy of Nico’s engine, already a step behind the original, and the copy is where the trouble sat. “I don’t actually know what Nico was thinking when he created this or why he created it” (23:24 to 23:35), and showing it as his own work would mean, in his words, “now I risk showing up as fake because the real link exists somewhere.” (23:53 to 24:01)
What a build is for, then, is the conversation it starts. Chris’s version: “you got excited, now you want to have a conversation. And when we have a conversation, you trust me and we’re going to create value together.” (24:50 to 24:58) A transcript of that conversation is what lets an AI help with the next explanation, he argued, and his question was how to make it “hyper relevant, hyper timely to solve the worst problems, not the problems that make us money.” (25:21 to 25:33)
A system only works if people use it
His second lesson was about whoever has to live with a system. A version that falls flat is useful material, he said: hand the AI the transcript of the moment it fell flat and ask for a take that works for the sales manager, then another for finance (32:30 to 33:10). A system that does nothing for someone’s day gets ignored, and he named where many teams spend their effort anyway: sales velocity and attribution reporting (34:30 to 34:48). “None of that matters if the people are not following the process in the system.” (34:48 to 34:55)
Read against Nico’s pipe, that is the limit of any accurate diagram. It can show a new administrator the stages in order; it cannot make the people entering the data follow them. The link is ours, not the hosts’.
A flywheel drawn while the show ran
Nico had one more build running in the background, and it arrived mid-show: a flywheel of the Value Path, modeled on the journey map Value-First publishes. It lays eight stages, Audience through Champion, against the kinds of work that move a person along (visible from about 42:00). A teammate, Chris said, had requested precisely this earlier in the week. He then asked for the source on screen too: the journey map pages, and diagrams he has been drawing by hand since 2020, by his account (42:28 to 44:00). The flywheel gives a reader a way across the gap of understanding without removing the need to understand. “You still have to know how it works.” (45:39 to 45:43)
Two ways to measure what the work returned
Nico spent the end of the hour on two ideas, the why and the clock. On the why he was short: “you can get everything else, but the why. The why comes from the individual.” (47:16 to 47:23) On what AI returns, he argued for a single measure: “the ROI of efficiency is time saved. There’s no other way to measure it.” (51:05 to 51:13) He would find the number by going to whoever does the work, and by asking it a particular way: “You’re not going to say how much time do you spend doing it. … You’re just going to say how much time do you waste doing this thing.” (52:11 to 52:22) Hardly anyone sells it that way, in his experience: “No one is advertising time savings, not that I have heard.” (54:26 to 54:31)
The day before, on Episode 42, titled “Efficiency Does Not Always Equal Value,” the emphasis had been different: Chris’s point there was that a count of hours saved cannot say what the hours were worth. On Episode 43 the hosts did not take each other up on it, and we are not settling it here. What the two positions share, in our reading, is where they send the question: to the person doing the work, not to a dashboard.
Episode — Value-First AI Daily
Efficiency Does Not Always Equal Value
Episode 42, October 21: the other side of the measurement question, from the hosts’ conversation the day before.
Open the episodeWho gets to build
Nico ended the hour on how far the ground has moved. “You technically shouldn’t be able to code your own apps … But you can now.” (60:18 to 60:31)
The Top 3 the hosts did not get to
Each episode’s Top 3 is written and sealed ahead of the show. The October 22 board is stamped 12:24 PM CT, a few minutes after the recording began. Chris mentioned it at the open, joked at 39:37 about a strike by the news team, whose articles he said he had stopped showing, and at the close Nico promised to come back to the news. Nobody read a headline, a source or a summary. Everything below comes from the sealed document and was not said on air. Chris’s own close was a promise about the builds: “We’re gonna keep showing builds every day at this point, and we’ll be back tomorrow with another one.” (61:51 to 61:57)
Ranked first, outside researchers found holes in the system that watches OpenAI’s coding agents (METR, “Red-Teaming OpenAI’s Internal Agent Monitoring Systems,” (opens in a new tab) October 22). A METR researcher and a Redwood Research researcher ran a pilot exercise inside OpenAI this summer, with access to its code, monitor settings and staff. Among the holes: an exploit that could hide an agent’s recent actions from the monitor, a route for Codex agents around their sandbox, and settings that switch monitoring off for new agents. OpenAI says every hole touching its production safeguards is patched. METR has not checked that, and for at least one hole it could not tell whether agents ever used it.
Ranked second, Anthropic put Claude Duet, an agent app meant to run around the clock, behind a waitlist (Anthropic, “Introducing Claude Duet,” (opens in a new tab) October 21; The Decoder). It runs on its own cloud computer, so a laptop can be off, and every few hours it checks connected apps such as Gmail, Slack, QuickBooks and HubSpot, then reports what needs attention. Before it sends a message, pays, signs or deletes anything, it shows the step and waits for approval. Trading securities and moving money between accounts are blocked even with approval. It is a research preview for US Pro and Max subscribers, and Anthropic says it can still misread a thread or misjudge what is urgent. The sealed take, from Nectar: “Asking before it acts is the floor, not the finish; Anthropic itself says it can still misread a thread. Before you join the waitlist, write down which steps you’d approve and which you’d never hand over, so where that line sits is your call, not the app’s.”
Ranked third, a team of thousands of OpenAI agents, each handed a slice of one brain-cell analysis, beat a lone agent running on the same model with the same resources (OpenAI, “Using thousands of AI agents to uncover new leads in biological data,” (opens in a new tab) October 21; The Decoder). With PerturbAI and Scripps Research, each agent took one gene disruption from a mouse brain atlas and checked it against the scientific literature. At the largest workload the lone agent cited 11 references against more than 700 for the team, blinded experts preferred the team’s reports 75 percent of the time, and the run took 32 minutes where the team estimates the same work by hand at months to a year. The leads it produced, such as a possible tie between a brain-disease gene and cholesterol production, are ideas for lab tests, not results.
The pairing is ours. Nectar’s advice on the second item, to decide before you join which steps you would approve and which you would never hand over, is the same move Chris made with Nico’s engine: know why the thing exists and where your own judgment sits before you put it in front of anyone. That decision is a why, and Nico’s point was that no tool supplies it.
Worth passing on?
