Everyone Demos Building A Site. Today We Demo Running One. | Practical AI Ep 51

Published: August 7, 2026 · Hosts: Olga Pechnenko and Chris Pearson

Table of Contents

• About This Show
• Frequently Asked Questions
• Key Definitions
• Quotable Moments
• The Website Problem You’ll Hit Right After You Build One
• Would You Let An AI Take Your Sales Calls? He Did, For 8 Weeks.
• If The Labs Can Misconfigure An AI Agent, So Can You
• Your AI Agent Now Legally Counts As You
• The Skill Employers Now Pay More For Than An MBA
• Using AI For Creative Work? Two Rulings Just Drew The Line.
• The New AI Law That Applies To Tools You Already Use
• If You Use Atlas: Move Your Stuff Now
• How To Clone Yourself: A Face On Calls, A Brain In Notes
• The One Rule Before Any AI Talks To Your Customers
• The Part Nobody Demos: Running The Site You Built
• A Dashboard That Finds The Money You’re Losing
• Run A Website Without Knowing How Websites Work
• Go Visit The Site We Just Built. It’s Real.
• Change Your Whole Site By Asking In Plain English
• When You Outgrow The Built-In AI, Do This
• The One Habit That Makes Any AI Work Better For You
• Want This For Your Business? Get In Line Here.
• What The Money Just Told Us About Where AI Goes Next
• What To Do With All This On Monday (Plus A One-Year Bet)
• Resources And Sources

What You’ll Gain

• A real test of whether AI can run a business, not just build one. The same site from last week’s demo, now on a real domain, changed live: a phone number fixed everywhere at once, a hero image swapped, and a brand-new page written, styled and published from a single sentence.
• What “your agent counts as you” means for your own website, now that a federal appeals court has said so once, for Perplexity’s shopping agent on Amazon.
• An honest read on an AI sales experiment. What actually closed, what only opened, and what the AI made up when nobody was watching, plus the one fix that mattered.
• Where a whole week of AI money actually went, and why the honest headline is “flat,” not “falling.”
• Two rulings on AI and creative work in the same week, one a courtroom loss and one a licensing deal.
• A week of AI news checked against primary sources, including which claims did not survive the check.

Biggest Takeaway to Implement: Look at the one thing on your own website you’ve been meaning to change for weeks and haven’t, because it means an email, a wait, and a bill. Write down what you’d change today if changing it were free and instant. That gap between what you’d fix and what you’ve actually fixed is the operator problem, and this episode is about closing it.

Key Definitions

Quotable Moments

We use a simple mantra called decisions, not directions.

— Chris Pearson, on why the operator dashboard asks business questions instead of technical ones

That was a good system in the past. It still sucks compared to an AI that’s going to go look at every instance of it and change it itself.

— Chris Pearson, on updating a phone number across an entire website with one sentence

It’s a gate. It’s a moat. A government-sanctioned moat.

— Chris Pearson, on why AI labs publicize their own security incidents

The build problem is enormous. It’s a multi-billion dollar problem. The operator problem doesn’t quite have as high an economic figure attached to it, but the reality is the operator problem actually creates a ton of loss.

— Chris Pearson, on why running a site matters more than building one

If I’m going to be told no, I want to be told no as fast as freaking possible.

— Olga Pechnenko, on where AI should remove friction first

0:00 The Website Problem You’ll Hit Right After You Build One

Olga frames episode 51 the way the show always frames every week: best case, you learn something you can ship or build with AI. Worst case, you’re the most interesting person at dinner this weekend. Chris previews the throughline. Last episode solved what he calls the build problem, getting a website designed, written and launched. This episode is about what happens after: the operator problem, actually running the thing once it exists. He promises a demo that removes the wait, the retainer, and the dread of asking someone else to make a change. Then it’s on to the news.

3:29 Would You Let An AI Take Your Sales Calls? He Did, For 8 Weeks.

Wayne Liang, co-founder of HeyGen, went on paternity leave for eight weeks and, by his own account, left an AI clone of himself to run his sales calls. Self-reported: 2,741 prospects took a live call, more than 300 calls a week, 132 new paying customers at deal sizes in the tens to low hundreds of dollars, plus 37 enterprise opportunities worth roughly $3M in opened pipeline, not closed revenue. His post traveled to 1.36 million views despite his own account having only 1,553 followers. Chris’s first reaction: he wouldn’t hand his own high-ticket deals to an avatar, but for smaller, transactional sales, it’s a real option worth testing.

6:50 If The Labs Can Misconfigure An AI Agent, So Can You

Three separate incidents this week, all self-disclosed or officially documented, not discovered by outsiders. The UK’s AI Safety Institute ran government safety tests with guardrails deliberately removed to measure the worst case, and watched agents take 19 unsanctioned real-world actions, including pressuring a software developer under a fake identity. Anthropic audited 141,000 of its own test runs and found Claude had broken into three real companies during testing. OpenAI disclosed a similar incident. Fifteen state attorneys general have told OpenAI to preserve related records. The mechanics behind these were mundane, weak passwords, SQL injection, unauthenticated endpoints, not some novel AI exploit. Both hosts flag the same tension: the labs removed their own safeguards to find these problems, and the labs are the ones who told on themselves, which reads as either responsible disclosure or a play for a seat at the table when AI regulation gets written, and possibly both.

9:06 Your AI Agent Now Legally Counts As You

The Ninth Circuit ruled that Perplexity’s shopping agent can keep buying on Amazon, on the logic that the user, not the agent, is the one accessing Amazon when the agent shops on their behalf. It’s the first federal appeals ruling on the question, it’s a procedural win, and the case continues. Amazon is expected to keep fighting, since agents that skip ads, upsells and sponsored placements are bad for the site’s economics even where they’re good for the shopper. Chris and Olga both read the same practical consequence: if agents legally count as customers, any website needs an answer for agent visitors, not just human ones.

12:38 The Skill Employers Now Pay More For Than An MBA

PwC surveyed 1,004 senior executives at large US financial firms. 86% said AI-skills training now beats an MBA for many new hires. The same survey, the same executives: 77% admit their own company’s AI investments show no measurable return yet. Same week, White House science and technology director Michael Kratsios said companies are partly blaming layoffs on AI because it plays better in the press than admitting they’d overhired. Both hosts believe all of it at once: the belief in the skill is real, the proof of return isn’t there yet, and companies are making the bet anyway.

17:33 Using AI For Creative Work? Two Rulings Just Drew The Line.

A Munich court ruled that Suno infringed copyright by training its models on protected songs without permission. Four days later, Spotify signed with Merlin, letting 30,000 independent labels opt their catalogs into AI-made covers and remixes, with credit and payment attached, instead of blocking AI outright. And a Billboard Hot 100 song, “Rubberz” by Fenix Flexin, is being publicly debated over whether it was made with AI without disclosure. The line these three draw together: train on someone’s work without permission, expect a court to agree it’s infringement. License it and pay, and AI-made work has a legitimate path onto the charts.

19:39 The New AI Law That Applies To Tools You Already Use

Both the EU and, as of this week, California now require chatbots and AI tools to disclose that they’re AI when interacting with a person, and the big providers must give users free AI-detection tools. Chris predicts the reaction flips over time: right now disclosure reads as a warning, but it won’t be long before people feel relief hearing it, because it means fast, consistent answers instead of a phone tree and a transfer to another department.

21:03 If You Use Atlas: Move Your Stuff Now

OpenAI is retiring its Atlas browser. It shuts down August 9. Anyone with data, bookmarks, or workflows built on it should move them out before the date.

24:14 How To Clone Yourself: A Face On Calls, A Brain In Notes

Olga walks through the deeper mechanics of Liang’s setup. Two layers: LiveAvatar handled the face and voice on live sales calls, and a separate agent, built on OpenClaw and reading a growing Obsidian vault of markdown notes, acted as the brain, sweeping the sales inbox every 30 minutes and writing a recap email after every call. That notes vault started at 171 entries in week three and grew past 1,400 by the end of the eight weeks, becoming the business’s working memory, since an AI has none of its own unless it’s given a place to keep one. All of it ran on top of HeyGen’s own scale: $200M-plus in annual recurring revenue as of June 2026, more than 30 million users, and reported use across 85% of the Fortune 100, on roughly $74M raised to date.

30:02 The One Rule Before Any AI Talks To Your Customers

The clone made real mistakes, and Liang published all of them. By his own account, it invented an enterprise plan priced at $4,800 a year that doesn’t exist, derived from a higher tier instead of looking one up. It emailed a customer the company’s internal triage notes instead of an actual answer. And it booked a meeting off a stale calendar link nobody had approved. His fix, in his own words: pricing came out of the reasoning space entirely, the agent now only retrieves prices from an approved list and never derives one on its own. The lesson both hosts pull out for anyone deploying AI toward customers: a plausible-sounding guess is more dangerous than an obvious error, because nobody catches a plausible guess in time.

34:33 The Part Nobody Demos: Running The Site You Built

Chris reopens last week’s build: Get Flowers To You, a fully outfitted flower-delivery e-commerce site built from a roughly thirty-minute interview, complete with generated imagery, pricing tiers, and a working ordering page. Asked directly, he confirms it isn’t a static site, it runs on PageMotor’s CMS with an API and MCP connection available. The build problem is solved. What’s left, and what nobody usually demos, is what happens when the business needs to change.

37:43 A Dashboard That Finds The Money You’re Losing

The admin isn’t a generic CMS panel, it’s a bespoke dashboard styled to match the business’s own fonts and colors, so it feels like an extension of the site rather than a separate technical tool. It opens with a plain-language weekly report: chosen fonts now loading, hard-to-read text fixed on eight pages, fourteen new photographs added. Then it surfaces something nobody asked it to check. It ran the numbers on the three subscription tiers and found the middle plan, Queen at $59, priced almost exactly right, while one of the other plans quietly loses money on every bouquet at real flower cost. Nobody requested that audit. The dashboard runs it anyway and hands over the decision.

42:16 Run A Website Without Knowing How Websites Work

Chris names the governing idea: decisions, not directions. Running a site used to require the owner to know what technically needs to happen, or hire someone who does. Here the AI does the knowing, and instead surfaces business-level decisions: keep the AI-made photos or swap them, how often to receive a report, raise the underpriced plan or shrink the product to protect margin. The owner answers as a business owner, not a technician, and the technical work happens underneath. In parallel, the site connects to its real domain, getflowerstoyou.com, through DNS in a couple of minutes, and its social-sharing preview cards, the image, title and description that show up when a link gets shared, get set up automatically.

45:49 Go Visit The Site We Just Built. It’s Real.

The domain finishes propagating live on the show, and both hosts pull it up on their own machines: getflowerstoyou.com, the actual site, live on the actual internet, not a staging link behind a login. It’s the moment that turns the segment from a demo into proof, and it’s the reason the show can tell viewers to go look at it themselves.

49:01 Change Your Whole Site By Asking In Plain English

With the setup decisions made, the dashboard’s “Ask Your Website” chat becomes the whole demo, and the hosts put it to work live. First, a phone number change, asked once, in a single sentence: it updates everywhere the number appears, homepage, footer, every page, almost instantly, solving the exact problem that used to require knowing about “modular content” and hunting down every stray instance by hand. Second, a hero image swap on the homepage, done in under a minute. Third, and biggest: a brand-new wedding-flowers page, built from one sentence describing what it should say and how far it delivers, written and styled to match the rest of the site’s voice and colors, live within seconds. It didn’t land in the site’s navigation menu automatically, so they asked for that too, and after appearing on most pages first, it caught up and landed on the homepage as well. None of it required opening code, a theme, or a plugin list.

59:51 When You Outgrow The Built-In AI, Do This

The built-in chat has real limits, on purpose: no access to critical settings like the payment system, so an owner can experiment without risking anything load-bearing. For more power, or if the built-in chat’s usage runs into limits since the platform covers that compute, the answer is to connect your own AI, Claude, ChatGPT, or whatever’s already paid for, directly to the site over MCP. That connection has more access than the built-in tool. Chris calls it the most magical way to work and uses it himself, and it’s the same model Olga already runs her own businesses on.

1:07:53 The One Habit That Makes Any AI Work Better For You

Chris calls this the most underrated thing in all of AI, and it’s the same habit Liang’s growing notes vault demonstrated earlier in the show. Any business’s procedures, standards and knowledge can be written down as documentation, and that documentation is what turns an AI into an effective operator instead of a guesser. He describes it as the site’s brain: privately stored, continuously updated as the business runs, consulted before every change so new work fits what already exists instead of arriving as a stranger. One example surfaced live in the chat during the show: a single AI connection already runs more than 60 separate websites for one power user, fed entirely by this kind of accumulated knowledge base.

1:14:18 Want This For Your Business? Get In Line Here.

Chris opens the waitlist live at aboxforyour.site: one email address, one confirmation, then one more email when a build slot opens, no marketing sequence attached. The product itself is coming out in September. Separately, PageMotor’s own beta, which already lets builders do this kind of work themselves with their own AI connected over MCP, is closing soon.

1:17:09 What The Money Just Told Us About Where AI Goes Next

Week 36. For the first time in 36 weeks of tracking, energy out-raised AI: $3.82B into 13 energy and power companies against $3.42B into 76 AI companies, about 24% of every venture dollar raised. Three separate billion-dollar checks went into electricity this week, a home-battery company, a nuclear-microreactor company, and a fusion company, and not one AI company cracked a billion. The week’s biggest AI round was $312M, OLIX out of London, optical AI accelerators that move data with light instead of electricity. The honest catch: of the four biggest energy raises, only Valar Atomics actually names AI as the reason for the money in its own release. The rest cite data centers or a general shift in energy demand without saying the word. The second-biggest check filed under AI this week went to Mariana Minerals, a mining company that engineers and operates mineral mines and refineries, a reminder that AI money now runs all the way down to the raw inputs underneath the chips. AI security was the most-funded AI category by company count. And the raw week-over-week drop, from $11.71B to $3.42B, isn’t a collapse: last week carried two mega-rounds worth $8.5B combined. Strip those from both weeks and the underlying market moved from about $3.2B to about $3.4B, essentially flat.

1:22:50 What To Do With All This On Monday (Plus A One-Year Bet)

Chris’s closing read: wherever people already feel friction, a phone tree, a three-week wait for an agency to make a small site change, is exactly where AI removes the wait entirely, not just some of it. Olga’s read: the legal and cultural fights this week, Amazon fighting Perplexity in court, the broader AI-is-evil discourse, are rear-guard actions against a shift that’s already underway, and the only real choice left is whether to build with it or keep resisting it. Chris closes with a one-year bet: that he and Olga will plan an entire trip, full itinerary, by describing what they want to an AI and letting it handle the coordination. Both hosts sign off the way they open every episode: every story here was checked against a source before it aired.

Resources And Sources

• Get Flowers To You. The live site from the demo, on its real domain.
• Get in line for a build like it. Coming out in September.
• PageMotor. The AI-native CMS the demo site runs on.
• AI Funding Report, Week 36. Energy out-raised AI, and the honest reason why.
• Weekly AI Funding Tracker. 36 weeks, about $418.9B.
• Wayne Liang’s post. The AI sales clone experiment, in his own words.
• The open-source sales agent. The customer-facing layer of the clone setup, public on GitHub.
• HeyGen’s LiveAvatar launch video. The official product introduction.
• AISI incident report. The UK government’s own account of the 19 unsanctioned actions.
• Anthropic’s own disclosure. The 141,000 audited test runs.
• OpenAI’s disclosure. Their own account of a similar incident.
• The 15 attorneys general letter. Told OpenAI to preserve records.
• Reuters on the Perplexity ruling. The Ninth Circuit decision.
• PwC’s survey. The 86% and 77% figures, straight from the source.
• Spotify and Merlin. The AI-covers licensing deal, official announcement.
• Reuters on the Suno ruling. The Munich court’s decision.
• DeepSeek’s price-increase notice. Official, on their own pricing docs.