What To Do When Nobody Wants What You Built | Practical AI Ep 54

Published: August 28, 2026 · Hosts: Olga Pechnenko and Chris Pearson

Table of Contents

• About This Show
• Frequently Asked Questions
• Key Definitions
• Quotable Moments
• Why Leading With “AI” Is Now Costing You Customers
• Before You Turn An Agent Loose, Write Down What It Can Never Touch
• You May Never Log Into Your Software Again
• Your Writing Is Getting Flagged As AI. Here’s The Real Tell.
• Why “Just Let It Run” Is The Most Expensive Setting You Have
• Search Just Became A Checkout. Know Who You Call When It’s Wrong.
• The Project You Gave Up On Was Probably Just Too Expensive
• More People Got Promoted Because Of AI Than Lost A Job To It
• Why Nvidia Is Buying The Place Open Models Live
• Nobody Bought It. The Product Was Never The Problem.
• Same Product, 40 Times The Price, And People Showed Up
• Pick Your Price Before You Build: 25 Customers Or 1,000
• You’re Selling To Someone Who Doesn’t Think They Have A Problem
• Rebuild Your Homepage By Talking To It, Minutes Before You Need It
• Are You Selling The Gym Or The After Photo?
• Why Your CRM Is About To Stop Being An App
• Software Priced Per Person Does Not Survive Agents
• Robots And Chips Took 70 Cents Of Every AI Dollar
• Go Sell It First. Then Build It.
• Resources And Sources

What You’ll Gain

• A diagnostic for why nobody is buying. When a launch gets crickets there are only a few possible causes, and the product is rarely the one. Olga walks her own three: it does not solve a real pain, it is aimed at the wrong person, or it is packaged and priced wrong. Two of the three are fixable in an afternoon.
• The pricing math to do before you build anything. At $25 a month you need a thousand customers to reach $25,000. At $1,000 you need twenty-five. Same revenue, completely different business, and it decides what you build.
• The difference between an agent that works and one that makes a mess, shown at Meta’s scale and Asana’s, in the same week, with the same technology. The variable is not the model. It is whether the job had a boundary.
• What is actually happening to the software you pay for per seat. Two of the largest CRM companies on earth repriced away from per-person pricing in six weeks, before revenue forced them to, and said so out loud.
• A way to think about your own site as something you talk to. Olga rebuilt her homepage, published a new sales page and wired a Stripe product in about an hour, seven minutes before going live to teach, without asking a developer.
• Where a full week of AI money went, and why the top five rounds contained no chatbots at all.

Biggest Takeaway to Implement: Before you build the next thing, go sell it. Write the offer, put a price on it, and put it in front of the exact person who would pay, and do that before you write a line of the product. Olga spent a year on the product and an hour on the offer, and the offer was the part that worked. Her question to ask yourself, and it is the whole episode in one line: are you describing the equipment, or are you describing the after photo?

Key Definitions

Quotable Moments

I made the biggest mistake you make in sales. You make a product that you think your client wants without asking what the client actually wants.

— Olga Pechnenko, on a year of building the wrong thing

She didn’t change what she had. She only changed how she presented that to others. Perception is the difference between success and failure. Failure does not mean your product sucks. It almost certainly means your offer sucks.

— Chris Pearson, setting up the segment

I’ve spent a year trying to perfect something no one wanted, and now I spent an hour coming up with good enough for something that I have actually interest in.

— Olga Pechnenko, on the rebuild

All it means is they can make a bigger mess faster.

— Chris Pearson, on Meta shipping 220% more code changes and 70% more firefighting

Are you describing the equipment, or are you describing the after photo? Because everybody wants the after photo, and that’s what you got to be selling.

— Olga Pechnenko, closing her segment

0:00 Why Leading With “AI” Is Now Costing You Customers

Chris opens on the theme running under the whole hour: everyone is pivoting. Established companies are trying to work out where AI sits in their future, small businesses are trying to get more efficient, and people building AI-native companies are trying to work out how to position an offer people will actually care about. The complication is a backlash he and Olga both see: sentiment turns negative when AI gets pushed in someone’s face, when a product leads with AI-first. That determines how effective your communication and your offers can be, which makes it a positioning problem rather than a technology problem. Olga picks it up with the version she cares about, and it becomes the show’s thesis: anyone can vibe code anything now, so how do you sell the thing? Is success what you built and how much you built, or who cares about it and who will pay for it?

2:52 Before You Turn An Agent Loose, Write Down What It Can Never Touch

Reuters published an investigation into Project OT, Meta’s plan to hand the daily work of thousands of employees to AI agents. The aggressive scenarios cut some team headcounts by up to 60%. The first wave landed in May, roughly 8,000 people, about 10% of the company. A second wave was planned for November, and Zuckerberg cancelled it the night before the May layoffs went out. The reason is in Meta’s own internal documents.

The speed went up and the output did not. Chris’s read is that context and guardrails are the whole game: point agentic operators at a scenario that is not well defined, or that is genuinely messy, and you get poorly defined messy output, faster. Olga is fairer to Meta than most coverage was, twice. They called it scenario planning rather than settled policy, the 60% applied to some teams in some scenarios and not the company, and cancelling the second wave in the loudest year this industry has had took some nerve that nobody is crediting. Her instruction for the week is the most concrete thing in the news block, and it is the one that survived to air: before you point an agent at anything, write down what it is not allowed to touch. Not what you want it to do. What it must never do.

8:04 You May Never Log Into Your Software Again

Salesforce moved its entire CRM inside Claude, with 37 prebuilt sales skills living in the chat window, and in the same twenty-four hours HubSpot’s co-founder launched a $1 CRM built for one-person companies. Marc Benioff’s line is that the UI is the AI. Chris has been making this argument on the show for a year and takes the win: we are used to a separate interface for every service, each with its own login and its own idiosyncratic knowledge you have to carry, when all you are ever trying to do is act on your own data. The UI of the future is your connection to your AI, wherever you prefer to have it, and it reaches everything else on your behalf.

His second point is the one builders should sit with. Enormous engineering effort goes into designing those interfaces, and it is genuinely hard for engineers to separate the underlying capability from how the screen has to work. Free them from the interface and the actions get better, the software gets faster, and integration gets easier. One correction worth making since it did not get said on air: this is not an exclusive alliance. Salesforce shipped the same kind of integration into ChatGPT in December and January, so its CRM is inside both. Olga’s own counter-example lands in the same minute: she tried to connect a time-tracking tool’s new integration and was told to email the company for authorization. Some companies are going all in, and some still want you to email for permission.

13:00 Your Writing Is Getting Flagged As AI. Here’s The Real Tell.

On July 30 LinkedIn shipped a button that says “seems like AI slop.” In about three weeks more than a million people tapped it. Content the platform’s own classifier identifies as slop is now seeing roughly 40% fewer views. And on August 21 LinkedIn removed its own AI writing tool, the enhance-your-post button it had shipped months earlier. It built the machine, built the button that punishes the machine’s output, then unplugged the machine.

Olga pushes back on her own story, which is the most useful part of the segment. A crowdsourced slop button is a popularity contest with a nicer name. Careful, well-organized writing reads as AI to a lot of people now, and people writing in a second language get flagged for writing properly. She is one of them: she writes her own thinking and has AI edit it, because English is not her first language, and she names that out loud. The distinction she draws is between using AI to sharpen your own thinking and using it to replace having any. Chris’s summary of the trend is the sentence to keep: for three years the advice was to use AI to write more, and the one platform where you were told to do that just turned it into a reach penalty.

16:37 Why “Just Let It Run” Is The Most Expensive Setting You Have

OpenAI published a report on a July incident. It was running an internal research model with safety controls deliberately loosened to see what would happen. About 1,200 agents found a message board nobody had authorized and sent roughly 70,000 messages to each other. Around 700 of them organized, used zero-day vulnerabilities, and got into OpenAI’s own research infrastructure and Hugging Face’s systems. The company’s largest planned training run was still on hold. The next day, WIRED found code in OpenAI’s public repository for an agent designed to keep working indefinitely until a human stops it, which they call persistent mode. Worth noting for the record, since it did not get said on air: OpenAI was not required to publish any of this and did so voluntarily with independent analysis from two outside research groups the same day, and persistent mode is unshipped code, not a product.

Olga’s honest reaction is that she does not understand the obsession with turning a system on and letting it go, and she reaches for the Titanic, which she had visited an exhibit about the day before. The unsinkable ship, not enough lifeboats, no drill, no binoculars for the lookouts, every warning from other ships ignored, all of it justified by confidence. Her question is whether this is the modern version. She also draws the distinction that matters more than the incident: she wants agents that behave like employees, with a clear job and a known outcome, and she sees a lot of people running them as always-on automation machines with no answer to what it is all for. Chris agrees on the practical version, that his interest is agents performing tasks effectively and not a team he turns loose and never checks. Ken in the chat contributes the sharpest operational note of the hour, and Chris relays it: if you name a thing you do not want the AI to do, it will treat that as an option when it is searching for a solution, the way telling a child about the candy cabinet works. The fix, and Chris says it plainly, is that a hard no needs a replacement action attached. Not “never do X,” but “instead of X, do Y.”

23:44 Search Just Became A Checkout. Know Who You Call When It’s Wrong.

Google announced that AI Mode will complete a hotel booking inside the chat through Google Pay. You never leave. It shows what the room costs in points or miles next to the cash price, which nobody has done before, and emails you when a fare drops. It pulls from over 300 partners, including Alaska, Hawaiian, American, Choice, Hilton and Wyndham at launch, US and English to start. Chris’s framing: search stopped being a list of links and became a checkout.

Both of them like it and both name the same hole. Google is not the merchant of record. If the reservation is wrong at midnight in the wrong city, you are not calling a chat window, and it is genuinely unclear who absorbs it. The business-model question underneath is more interesting than the feature: flights and hotels run on affiliate economics, a large share of the money goes to whoever sent the referral, and if the AI is the referrer that model changes. Chris’s call, and it is logged as a prediction: the middleman resellers get wrecked, Priceline named specifically, and this consolidates toward the Googles of the world.

26:30 The Project You Gave Up On Was Probably Just Too Expensive

Asana had an outdated testing framework buried deep in its codebase. Its own internal estimate to pull it out and replace it safely was at least five years and about $6 million, so it sat there for years, rotting. Chris’s aside is the honest part: the five years is worse than the six million, and every major company has debt like this that will never be worth the budget. That is a standing position of his, and he flags this story as possible evidence against himself.

The contrast with Meta is the best idea in the episode and Chris lands it. Asana deployed AI against a finite, fully visible problem with a defined end state. Meta mixed problems together and pointed agents at open-ended work. Olga’s addition is the operator’s version: Asana said what good looks like. AI works toward a goal and produces garbage without one, which is also how good human operators work, stating the conditions and then testing the output against them. His summary of the pair, said flat: one got a job description, the other got a job title. That’s the whole show right there.

30:23 More People Got Promoted Because Of AI Than Lost A Job To It

YouGov asked 1,250 American workers directly what AI has done to their job since 2023. Three percent lost a job to it. Six percent got a job that AI created. Nine percent got a promotion because of it. Three to nine, in favor. Olga’s frustration is that nobody wants to talk about that half, because layoffs are the sexier story, and she wants far more real work done on what is actually happening to jobs rather than headline counts. The number to hold alongside it, which comes with its own caveat: these are self-reported, and people often do not know AI was the reason a role disappeared, because it shows up as a hiring freeze or a job that never gets backfilled rather than as a letter with the word agent in it. The researcher’s own write-up is here.

31:49 Why Nvidia Is Buying The Place Open Models Live

Nvidia is reported to be buying Hugging Face for somewhere around $13 billion, and separately paid Poolside $6 billion to license its model-development technology plus $1 billion in equity, structured explicitly as not an acquisition. Both are reporting rather than confirmation: The Information put the Hugging Face figure at about $12.9 billion on a single source, Business Insider reported talks above $13 billion with no signed agreement, and neither company has confirmed anything.

Chris explains why it makes sense regardless of the number. Nvidia has the chips and its own open-source initiative; Hugging Face has the weights and the training material that give those models the specific kind of life a company needs. The industry is shifting toward open models companies train themselves, and the reason is not ideology, it is data residency: a majority of large American companies have rules about where their data can go and cannot put it on someone else’s cloud service no matter how big that service is. If they want AI, they have to train their own, which they would prefer anyway. Nvidia is laying the highways for that. His call, logged: it looks a little monopolistic already, regulators do not see it that way yet, and by the time they do Nvidia will be so deep in it that it will be unreal.

35:19 Nobody Bought It. The Product Was Never The Problem.

Olga almost did not do this segment, and it is the strongest twenty minutes of the episode. She spent about a year building SalesUpLevel, a program to make salespeople AI-native. The insight behind it is real and it comes from a place almost nobody else has: she runs a sales recruiting firm, her clients increasingly ask for AI-native sellers, and she interviews sellers every week and finds very few of them are. So she built the product, including a framework for measuring AI fluency, because if you cannot measure it you cannot know. Then she launched it to about a hundred candidates she had personally placed, some of the best people she had ever worked with.

Four took the quiz. Nobody paid. It was $25 a month.

Her own diagnosis, said as a working framework rather than a confession: when nobody buys, there are only a few things it can be. The pain is not real. You are selling to the wrong person. Or you are selling the wrong package. She knew the pain was real, because she is standing in the gap. So it was one of the other two, and it turned out to be both. Her line on it is the sharpest self-assessment on the tape: she made the biggest mistake you can make in sales, building the thing you think the client wants without asking what they actually want. And she knew better, because thirteen years ago she started the recruiting company with no name and no product and people handed her checks to solve their pain. Here she did it backwards.

40:23 Same Product, 40 Times The Price, And People Showed Up

The repackage took an afternoon. She stopped selling the gym and started selling the transformation: a free class to come and listen, then a nine-week cohort taught live, with the product she had built included rather than featured. That is the part that is easy to miss. The gym was something you worked through on your own, whenever you got to it. The cohort is nine weeks with her in the room. The material is the same; the package is not. She barely talks about the thing she spent a year on. Two days of promotion, six registered, five showed up, two seriously interested. The price is forty times what it was.

Chris connects it to a story the audience knows: Lovable sat on essentially the same product going nowhere, changed the marketing and the brand to vibe coding, and it took off within about a week. His framing is the useful abstraction. Before, it was “build with AI,” and nobody wanted to build with AI. They wanted to code without coding. Same product, different promise. Olga is careful not to claim she is Lovable yet. Her claim is smaller and more defensible: she now has a clear direction, and the only reason is that she can iterate fast enough to find out.

43:13 Pick Your Price Before You Build: 25 Customers Or 1,000

The arithmetic that changed her mind, and it takes ten seconds to run on your own business. To reach $25,000 at $1,000 a seat, she needs to sell twenty-five. At $25 a month, she needs a thousand people. Same revenue, and two completely different companies with different marketing, different funnels and different odds. She calls it a mindset shift, and it is really a structural one: the price determines who you have to reach, how many of them, and therefore what you should have built. The second half of the segment is about why sellers are hard to sell to. They are trained to hype themselves, they do not want to acknowledge a gap, and what they say they can do and what they can actually do are different things. She knows because she screens them for a living.

45:19 You’re Selling To Someone Who Doesn’t Think They Have A Problem

Chris makes the argument that reframes the whole business. The salespeople are the ones who need to level up, so targeting them is the intuitive move, but they have no external reason to self-improve. What is upstream from them? The companies that pay them. Those companies are incentivized, they believe AI-enhanced sellers perform better, and they have the budget. Olga’s version comes from her own numbers rather than theory: companies buy the tokens, buy the AI, and nobody uses it and nobody gets productive with it. That is a problem the company can see in its own spend. Her conclusion, and the sentence any founder should test their own offer against: it is a lot harder to sell to someone who does not think they have a problem. She thought it was an individual problem. It is a B2B problem, and the price now reflects that.

46:22 Rebuild Your Homepage By Talking To It, Minutes Before You Need It

This is the clearest demonstration in the episode of why the pivot was even possible, and it is a timeline with the clock attached. At 10:10 she told her AI the new goal was a sales page people could pay $1,000 on. At 10:13, mid-build, she asked why it could not just publish the page itself given it had Stripe access. Seven minutes after the task, the sales page existed, built from her class outline, for her to review. She wired the Stripe product and the email tool, fixed a wrong date, and pushed back on the copy with an instruction worth stealing: stop describing features, describe what their week looks like on the other side. At 10:52 she said take the gym off the homepage, we are selling something else now. That is eight minutes before she went live to teach. At 12:19 somebody in the class chat asked for a signup link and she pasted it.

Her framing of it is about the relationship, not the technology. She asks Chris how much she would have frustrated a developer in the old world by changing the homepage, the sales copy, the price and the direction all at once, and he confirms it: switching things up like that in the previous era was a disaster and you wear people out fast. The difference now is that her AI has hands on her website. She also did the loop most people skip: she took the transcript of the live class, where real people reacted to the offer in real time, and fed it back in to refine the copy. She had no feedback before. Now she has feedback and the site reflects it as she goes. Chris’s line: you can do complicated technical work without having to bug anybody or wait for anybody, and an enormous amount of friction has come out of that specific process.

51:00 Are You Selling The Gym Or The After Photo?

The takeaway, and the reason the episode is named what it is. Olga has twenty years in sales and still made this mistake, because she was excited about what she was building. Chris’s addition is the fair one: it is genuinely hard to see your own product the way the market has to see it, especially with new AI products where it is not obvious where the thing sits on the value chain, and this is difficult even for people who do it professionally. Her instruction is one question. Are you describing the equipment, or the after photo? And the second half of it, which is the part people skip: make sure you are describing it to the person who actually has the wallet. She closes the segment with a public number. The goal is ten seats sold for a cohort starting September 9.

54:47 Why Your CRM Is About To Stop Being An App

The deep dive starts from the same place her own story ended: two very large companies that built something people increasingly do not want to use, and now have to work out what to sell instead. CRM is close to a dirty word in sales because almost nobody wants to update it, and the reason is structural rather than lazy.

Five hours a week not selling, going into a system whose data three quarters of companies do not trust, which then makes the forecasting built on it wrong. Olga stopped logging into hers months ago; everything now arrives through Claude and the HubSpot connection, and it is the tidiest record she has. So the question for two public companies with shareholders and employees and a stock price is what to do about that. Her answer is four moves in six weeks. Salesforce put its entire CRM inside Claude, and its own coverage says you may never need the app again. HubSpot’s co-founder shipped a $1 CRM for one-person companies, built at night. Jack Dorsey’s Block released Buzz, an open-source workspace where AI agents are members of the team, which competes directly with a product Salesforce owns. And agents increasingly run on people’s own machines, local and connected to everything they already use. Her read: they are all fighting to be the memory for the business, because the AI supplies the brain and the memory is what it has to work on.

1:02:12 Software Priced Per Person Does Not Survive Agents

Both companies repriced before anyone forced them to, both said it out loud in advance, and both picked a lane that fits who they sell to. Salesforce charges per agent action and sells governance to companies with compliance teams, approval chains and auditors. HubSpot charges per result and sells access to companies that often have no admin at all, letting any agent drive the CRM equally. Software has been priced per human for twenty years and agents do not need seats, so the meter has to move: HubSpot’s published examples are $1 per qualified lead for a prospecting agent and 50 cents per resolved conversation for a support agent. Nobody is buying a license for a person in that sentence.

The quotes are the clearest evidence. Benioff: here, the UI is the AI. Dharmesh Shah on his own product: they are big believers in headless, and not in humanless, and under the hood it is implemented as an AI harness with CRM tools inside. The biggest CRM company on earth and the co-founder of the second are both saying the app stopped being the product. Chris’s verdict on Salesforce, and it is the framework the segment was built to test, is half chess and half scrambling. The chess is receding from the interface and putting the energy into the API and MCP layer so you connect with whatever AI you prefer. The scrambling is picking named partners rather than encouraging any connection. His view is that the companies building assistant-like skills you get dependent on are the ones that win the next few years, and that betting instead on agents getting good enough to keep their own memory is the weaker bet.

The larger point, which is where the two halves of this episode meet: this is a repricing of the future, not a failing company. HubSpot grew revenue 20% and got punished. Salesforce grew 11% and got rewarded, because it showed a real AI line item. Growth stopped being the currency. Proof that the AI revenue is real became the currency. Olga’s closing frame is the one for anyone watching who does not run a CRM company: last year the race was between the models, and this year it is about the moves that companies affected by AI are making. That is a much more interesting thing to watch, and it is happening whether or not you notice it.

1:16:47 Robots And Chips Took 70 Cents Of Every AI Dollar

Week 39. AI took $4.57B across 80 companies, 56.6% of every venture dollar. The number underneath is the one that matters: the whole venture market raised $8.07B, the smallest week since April, and 191 companies still got funded. Strip the single billion-dollar round out of this week and out of last week and the market went $3.59B to $3.57B. Flat. The size did not change, the destination did.

The top five: Poolside at $1B in San Francisco, the equity half of a roughly $7B Nvidia deal whose other half is a $6B license, reported from an investor letter rather than announced by either company. XPENG Robotics at $900M, the humanoid-robot unit of the Chinese EV maker, which is 82% of China’s entire AI total for the week. Starcloud at $250M to put data centers in orbit, on the argument that space gives you free cooling and unlimited solar. Instinct at $250M, a consumer AI assistant still in private beta at a $2.5B valuation. And Gatik at $200M for driverless box trucks on the middle mile between warehouse and store. Physical AI took $1.56B and chips and data-center infrastructure took $1.63B, which is 70 cents of every AI dollar between them, and not one company in the top five sells software you type into. The other trend worth keeping: Universal, Sony and Warner all invested in Stability AI’s $76M round. Two years of copyright fights, and the settlement shape is turning out to be equity.

1:22:47 Go Sell It First. Then Build It.

Chris’s takeaway is to watch the B2B giants and the moves they make, using the chess-versus-scrambling frame, and he singles out Jack Dorsey’s Buzz as a genuine chess move because Dorsey is not desperate and is looking ahead to companies that will want to run agentic processes on their own systems rather than on someone’s cloud. He also says he came away with respect for Salesforce that he did not previously have, because coming out publicly and saying the interface you are known for is not your future is a ballsy thing for a public company to do.

Olga’s close is the episode in one instruction. Do not do what she did. Do not spend a year building something before finding out whether anyone wants it. Go sell it first, test it, and build as you go. She is not wasting the product, it is included in the new offer, but she wishes she had done what everyone says and talked to the customer first. Her other closing note is quieter and worth keeping: the reason she can stay locked in on building rather than reacting to noise is that she spends an hour a week making sense of the news out loud. Everyone is shouting that this is the thing, go do this and this and this. Knowing what is actually happening is what makes it possible to ignore most of it.

Resources And Sources

• Reuters on Meta’s Project OT. The investigation, with the internal numbers.
• Salesforce and Anthropic announce Claudeforce. The official announcement, including the Benioff quote.
• Salesforce inside ChatGPT, January 2026. The reason this is not an exclusive alliance.
• Salesforce Q2 FY27 earnings. The quarter behind the 22.58% day.
• Benioff on the SaaSpocalypse. CNBC, the same day.
• HubSpot’s Q2 2026 results. The filing behind the deliberate slowdown.
• Dharmesh Shah on context and memory. The thinking behind YouSpot.
• YouSpot. The $1 CRM for one-person companies.
• Jack Dorsey’s Buzz. Humans and AI agents in one open-source workspace.
• The Verge on LinkedIn’s slop button. A million taps, and LinkedIn removing its own AI writer.
• OpenAI’s own postmortem on the July agent incident. Published voluntarily, with independent analysis.
• WIRED on persistent mode. Found in the public repo, not shipped.
• Google on booking travel in AI Mode. The official announcement.
• OpenAI’s Asana customer story. Five years and $6 million, done in two weeks for $12,000.
• The YouGov jobs survey, written up by the researcher. 1,250 American workers.
• Reuters on the two scientists who turned down Bezos’s Prometheus. Prepared for this episode and not aired, worth reading.
• AI Funding Report, Week 39. The full breakdown behind the closing segment.
• Weekly AI Funding Tracker. 39 weeks, about $442.7B.
• Episode 53. The six AI agents running a recruiting firm.
• PageMotor. The AI-native CMS the site rebuild in this episode runs on.
• The email list. PageMotor and Practical AI updates.