The Nick Standlea Show
← All episodes

Guide

Practical AI for Business: The Model Was Never the Bottleneck

Most advice about AI and business answers one of two questions: will this take my job, or can it write my emails faster. Neither helps much on a Monday morning. The guests collected here were asked something narrower — what did you actually change, and what did it do. Their answers arrive from five directions and land close together. The model is not your constraint. A frontier model handed nothing knows nothing about your pricing, your customers, or how your refunds work, and it will answer anyway, confidently. Give it that context and the same model starts behaving like someone who has read every document in the business. Withhold it and you get a competent stranger's guess, written well enough that nobody catches it. That is the first half of this guide. The second half is what happens once the thing works: which tasks can run without a person watching, which cannot, and what a wrong line between them costs. Then the part most operators find out about last — that being good at AI inside your company does nothing for whether an AI recommends you to a customer. That runs on its own mechanics: one guest here takes them apart, and another has measured where the citations actually come from. Where these guests pull in different directions, the differences are set out at the end rather than smoothed over — including the ones that turn out not to be fights. Two put the bottleneck in different places, but they are measuring different things: one the model's reliability, the other the organization around it. Two others mark the edges of how much work to hand over at all, one arguing that most people ask AI to do far too little, the other that the first draft is work to keep for yourself. Those are live questions, and you will have to answer them inside your own company.

By Nick Standlea · Updated Oct 27, 2026

The model does not know your company

Rick Nucci, Guru's CEO and the founder of Boomi before Dell acquired it, puts the split at 40 percent technology and 60 percent human. His reasoning is specific. Today's models are capable and know nothing about your products, pricing or processes, and they will answer as though they do, which in Guru's experience is one of the largest sources of hallucination. That makes accurate company context binary: a yes or a no. The reason it is hard is structural. Organizational truth is decentralized, living in the heads of subject-matter experts rather than in a document. And the person you most want the AI to inform is usually the one who cannot tell whether the answer is wrong. Geoff Woods, who wrote The AI-Driven Leader, names the step people skip. CRIT stands for context, role, interview and task, and the interview is the one almost nobody runs: you hand the questioning over to the model, and it surfaces detail you would not have known to volunteer. On the context block his rule is that the moment it feels sufficient, it is not. Give three times more. Zapier's co-founder and chief executive, Wade Foster, has the cleanest before and after. A vanilla model returns an intelligent, generic answer. The same model handed the strategy, the customer profile and live account data reads down to ground truth instead of one person's functional view.

Pick the decision before you pick the tool

Igor Pogany, who runs the AI Advantage channel on YouTube, teaches three levels. The first is AI for answers, which he likens to a fancy Google. The second is AI as a thought partner you have loaded with context. The third is the architect's level, where you build the agents and the systems and they run without you. His observation is that almost everyone leaps from the first to the third, which he compares to marrying somebody after one date. Nucci watches the attempt from the vendor's side and has a blunt test for which buyers succeed. The green flag is an operational number the buyer can defend to a finance chief — support handle time that used to run minutes and now runs 30 or 45 seconds. The red flag is a buyer who opens with theoretical architecture and which models are swappable, and has no answer for what they are trying to accomplish. Token volume he dismisses as a vanity metric, descended straight from judging an engineer by the thousands of lines of code they produced. Andy Sack, co-author of AI First, orders the work unglamorously: become an AI-first leader yourself, build a broad base of employees using it daily, write a policy, and only then hire for it. Seth Godin reduces it to one line. Choosing which game to play is the strategy; improving at the one you already entered is not.

Decide what runs without you

Foster puts it as building with AI and running with code. The valuable part is the back-and-forth of designing a workflow; once designed, it should execute deterministically. The mistake he sees is leaving the model in the loop at runtime as well, so a fresh decision gets made at every step — which gives up reliability and spends tokens for nothing. He will let an agent run unreviewed at 99 out of 100, and not below that. He also describes what he calls the lethal trifecta — an agent holding private data, exposed to untrusted outside content, and able to act externally — and builds with only two of those three at a time. Pogany reaches caution from the opposite end. If the underlying model is wrong one time in 20, you cannot stack dependable workers on top of it. His filter for what deserves automating is whether a conversation you already had would improve your life if it ran again next week, and he guesses that is a single-digit share of them. Two guests supply the vocabulary for failure. Peter Weir, a family physician at the University of Utah, prefers confabulation to hallucination, because confabulation describes inventions delivered with so much conviction that nothing in them gives the invention away. Steve Rosenbaum had a model invent his entire career history.

Being the answer, not the tenth link

None of that work decides whether a model puts your name in front of someone who has never heard of you. Different machinery, different job, and Ross Simmonds is the guest who takes it apart. A single customer question does not produce a single search: it produces searches across roughly 10 to 30 channels — Reddit, YouTube and Yelp among them — and the answer gets assembled out of whatever comes back. Assembly here is arithmetic rather than editorial judgment. Variants of the question go out — best, top, highest rated, most reviewed — the results arrive from those channels plus the top Google links, and the business name that comes back from the most of those sources is the one the answer leads with. Simmonds's first instruction follows from that: write down the 20 questions a customer asks before your name has ever reached them, then publish something substantial against each one. Brett Farmiloe owns the pitch platforms that journalists and sources run through, which is what let him measure where citations originate. Out of about 22,000 citations, 64 percent traced back not to the brand's own website but to coverage a writer had published somewhere else, and a third landed on sites whose domain authority sits below 40 — the kind of small blog that never makes a media list. Jay Baer and T.D. Orel supply the discipline underneath. Baer treats organic reach as undependable and puts what actually matters into an email list he owns. Orel is a novelist rather than an operator, and he reaches the same lesson with no business vocabulary in it at all: a thing made for everyone turns out to be for nobody.

What does not get cheaper

Joe Lazer went looking for what AI makes more valuable rather than less, and landed on the small size of a human network. Research puts the circle a person actively maintains at about 150 to 250 people. An AI-written message spends trust drawn from that set, and the set is too small to degrade — Paul Graham, he notes, simply stopped reading email he can tell a model wrote. Lazer's harder claim is that drafting is thinking, so people who hand off their first drafts lose the capability while people who keep it widen the gap. Alan Weiss has written more than 50 books on consulting and expects AI to finish off hourly billing, because hours were always the wrong unit. His alternative is conceptual agreement, settled before a proposal exists: agreement on objectives, which are always outcomes rather than deliverables, on metrics, and on what meeting those objectives is worth. Justin Wolfers, a University of Michigan professor of both economics and public policy, offers the most portable instruction of anyone here, and it is not about becoming an AI expert. Try it on your own work every day. A year of that — 365 days — and you have roughly 300 ways to be more effective than whoever sits beside you.

Where they disagree

Related conversations

Field Notes

The sharpest idea from every conversation.

A short dispatch for people who'd rather understand than scroll — one big idea from each guest, distilled. Free, most weeks. Unsubscribe anytime.