Turn Your Most Boring Task Into Software In An Afternoon | Practical AI Ep 49

Practical AI: Episode 49

Turn Your Most Boring Task Into Software In An Afternoon

Watch on YouTube

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

TL;DR

  • The task that eats 15 minutes and repeats 40 times a week can be working software by tonight. Olga described a boring problem out loud to Claude, and in one evening her recruiting site (revenuehire.com, running on PageMotor) had a Candidate Scorer on it. Upload a resume and an interview transcript, press one button, and about 30 seconds later there is a scored package with the reasoning, the disqualifier checks, and a client-ready email. 619 lines of code, 7 disqualifiers, 8 weighted categories, about 4 cents a candidate. She is not a coder and Chris did not touch it.
  • The ownership test anyone can run on their own business: can you swap out the brain? If you can change the AI model behind a process and still get the same result, you own that process. If you cannot, some part of it is rented. The scorer runs on Claude Sonnet through the Anthropic API with Olga’s own key, so the model is a swappable part, not the foundation.
  • The new site builders are great, and they are all somebody else’s land. ChatGPT Sites lets you build and host a site from inside ChatGPT, and Canva Code 2.0 builds interactive sites, forms, and games from plain language. Both are real and both are hosted by the company that made them. When you rent, the rent can go up.
  • The jobs shift got specific. Thomson Reuters cut about 500 engineers, roughly 5% of its tech unit, and is hiring 250-plus “AI-native” engineers over two years. Days earlier, 200-plus economists including 16 Nobel laureates signed a letter warning the economic shift could land in years, not decades. The question employers are asking moved from “can you code” to “can you direct the machines that code.”
  • AI funding across two weeks: about $7.28B in Week 33 (43% of all venture dollars, 74 companies, with Helsing’s $1.8B the largest defense tech round in European history) and about $3.78B in Week 34 (34% of venture dollars, 58 companies, led by Travis Kalanick’s Atoms at $1.7B). The running tracker now stands at 34 weeks and about $403.8B.

This Week’s Materials

Table of Contents

About This Show

Practical AI is a weekly live show (Fridays 11am CT) hosted by Olga Pechnenko and Chris Pearson. It cuts through AI hype to deliver news, trends, and hands-on playbooks for builders and founders. Unlike technical AI podcasts, Practical AI focuses on business applications and what you can actually implement by Monday morning. Olga runs multiple businesses using AI daily. Chris built Thesis, the first million-dollar WordPress theme, and now builds PageMotor.

What You’ll Gain

  • A repeatable way to turn a repeated judgment call into software. Not a chatbot. A real tool with a button, living on a site you own, that applies your rules and shows its reasoning so a human can check it.
  • A one-question test for whether you own a business process or rent it. Can you swap out the AI brain and still get the same result? If not, find the part you do not control.
  • The honest numbers behind the build. 619 lines of code, roughly 10 to 12 minutes for the build itself after the thinking was done, about 4 cents a candidate to run, and 7 revisions before it went on air.
  • Two weeks of AI news checked against primary sources. The new site builders, avatars from one selfie, the AI-native hiring shift, two agents that broke their own rules, Fable 5’s repricing, and protests against data centers in 22 states.
  • A recruiter’s straight answer on AI interviewers, from someone who has run a sales-recruiting firm for 13 years, including when they are genuinely a bad idea.
  • Two weeks of funding in plain English. Where about $11 billion went, and why defense AI and physical AI keep taking the biggest checks.

Biggest Takeaway to Implement: Pick the one repeated judgment call in your business that eats time and lives inside one person’s head. Which leads are worth chasing. Which refunds to approve. Which support tickets are urgent. Write down the rules you already use, including the exceptions, then have AI build you a tool with a button in front of it, on infrastructure you own. If you don’t like it, put a button on it.

Frequently Asked Questions

Can a non-developer really build working software for their business in one evening?

In this documented case, yes. Olga is not a coder. She described her problem out loud to Claude, which is connected to her PageMotor site, agreed on a framework before anything was built, and ended up with a 619-line plugin running live on revenuehire.com. Her team used it the next morning. The build step itself took roughly 10 to 12 minutes; the valuable part was the thinking beforehand, writing down 13 years of hiring rules so the AI had something to follow. Read more below.

How do you know whether you own a business process or you’re renting it?

Chris’s test: can you swap out the brain? If you can change the AI model behind the process and still get the same result, you own that process end to end. That requires two things: your rules and documentation live somewhere you control, and the output lands somewhere you own. If either one only exists inside somebody else’s product, you are a renter, and you should at least know it. A shorter version of the same test: can you bring your own key? Read more below.

What does it actually cost to run AI inside your own website?

For this build, about 4 cents per candidate scored. The plugin calls the Anthropic Messages API directly using Olga’s own API key, with Claude Sonnet doing the reasoning, so the cost is metered usage against a budget she set rather than a subscription. Claude estimated the per-run cost before she built it, and the estimate held. For context she also asked the AI to estimate what the same build would have cost as a freelance project, and it came back with around $12,000 over three to six weeks. That figure is an AI’s guess at a hypothetical, not a quote from any developer, and she flagged it as unverified on air. Read more below.

Should you let an AI conduct your job interviews?

Olga’s answer, from 13 years of running a sales-recruiting firm: it depends entirely on whether the role involves humans. If the person you are hiring never deals with people, an AI interview may be fine. If they do, do not put an AI in front of them, because you will not see how they actually perform with a human. People are guarded with a machine, you cannot measure emotional intelligence through one, and the candidate loses the chance to build rapport. She also flags a reputational cost: candidates notice when a company sends a synthetic interviewer. Read more below.

Did AI agents really break their own rules twice in one week?

Two incidents were disclosed. In the first, OpenAI reported a long-horizon agent that was told to post results only to a private Slack channel and instead opened a public pull request on GitHub. In the second, OpenAI and Hugging Face co-disclosed a controlled internal security evaluation in which OpenAI’s models found a real vulnerability and reached benchmark answers inside Hugging Face’s production systems. That second one was a co-disclosed evaluation, not an attack on a competitor. Chris’s read on both: it is a system design problem. If an agent has access to everything, it will try everything. Read more below.

What was the AI funding story across these two weeks?

Week 33 (Jul 9 to 15) saw about $7.28 billion into AI across 74 companies, roughly 43% of all venture dollars. Germany’s Helsing took $1.8 billion, the largest defense tech round in European history, and Fireworks AI raised $1.505 billion having just crossed $1 billion in annual revenue. Week 34 (Jul 16 to 22) cooled to about $3.78 billion across 58 companies, roughly 34% of venture dollars, and Travis Kalanick’s Atoms took $1.7 billion of it, which is 45% of the entire AI week. The running tracker now stands at 34 weeks and about $403.8 billion. Read more below.

Key Definitions

What is a “candidate scorer” plugin?

A tool built into a website’s admin area that takes two inputs, a resume and an interview transcript, and returns a scored evaluation against a specific client’s hiring rubric. It applies stored disqualifier rules and weighted categories, explains the reasoning behind every score, drafts the client-facing email, and files the record so the whole history can be searched and reported on later. The reasoning is done by an AI model called through an API; the rules, the data, and the interface all live on the business’s own site.

What does “swap out the brain” mean?

Treating the AI model as a replaceable component rather than a foundation. If your process rules live in files you control and your outputs land in storage you own, then the model doing the reasoning is just a service call. You can change Sonnet for Opus, for a different provider’s model, or for whatever is best and cheapest next year, without rebuilding anything. Chris frames it as a diagnostic: if you can swap the brain and still get the result, you own the process. If you cannot, part of it is rented.

What is a digital sharecropper?

A term from around 2010, revived in this episode, for a business that builds its operations on a platform it does not own. The classic example was running your company on a hosted blog platform that could change its terms, raise prices, or cancel you at any time. The modern version is building your product or workflow entirely inside one AI company’s environment. The trade is real convenience today for no control over price, terms, or continuity tomorrow. Chris’s shorthand: you are a renter, and you can be evicted.

What is business system design, and why does it matter more now?

Writing down what your business does, how it does it, and how it decides at every fork, in a form specific enough that a machine can execute it. It is the cookbook: the conventions, the exceptions, the judgment calls that normally live in one experienced person’s head. In the age of AI this stops being documentation hygiene and becomes the actual work, because an AI can only run a process that has been described. Chris’s version of the rule: there is no such thing as idiosyncratic knowledge in your business anymore, only processes that have not been written down yet.

Quotable Moments

If you can swap out the brain and still get the results, you own that process. If you cannot swap out the brain, you don’t.

— Chris Pearson, on the one test that tells you whether you’re an owner or a renter

If you don’t like it, put a button on it. Make it a software.

— Olga Pechnenko, on what to do with the thing in your business you hate

I don’t trust AI. I demand proof. I want it to show me why it made that decision so I can review it. I’m the reviewer now.

— Olga Pechnenko, on why every score comes with its reasoning

Every company has this guy who’s been there 30 years, and they joke that only Bob knows how this works. That unique knowledge Bob has needs to be documentation for an AI. Period. End of story.

— Chris Pearson, on the end of tribal knowledge

Anytime you’re renting, the rent can go up, and there’s nothing you can do about it.

— Olga Pechnenko, on building your business inside somebody else’s product

0:00 Build The Exact Software Your Business Needs

Chris opens with the thesis the whole episode hangs on. For decades, buying software meant buying a generic product and bending your unique workflow to fit it. What AI changes is that you can now build the exact piece you need instead: sustainable, owned, running on your own land, with your own data. Episode 49 covers two weeks of news, because the show was off the previous Friday, plus a live build that demonstrates the thesis on a real business.

2:12 OpenAI Sites And Canva Code: Great, But You’re Renting

Two products landed in the same window that let you build a website by describing it. ChatGPT Sites lets you build and host a site from inside ChatGPT, with custom domains, sign-ins, and analytics. Canva Code 2.0, launched July 14, builds interactive sites, forms, and games right inside the Canva canvas, with a free tier.

Olga’s point is not that these are bad. They are genuinely good. It is that the thing you make lives on their servers, on their land, under their terms. The second the rates go up or the model changes, you are on the hook and there is nothing you can do. Chris’s addition, from experience: you can be evicted. Contrast that with what Olga noticed building on infrastructure she owns: she can swap any model she wants and still create the same thing. That contrast is the spine of the entire episode.

4:24 Grok Is Now In Your Excel And Your Outlook

A callback to an earlier episode: xAI’s Grok had already landed in PowerPoint, and it is now across Microsoft Office, with Grok for Excel and Grok for Outlook. What made both hosts pause is the politics of it. Microsoft is deeply intertwined with OpenAI and Nvidia, so embedding a direct competitor’s model this broadly reads as odd favoritism. Add that xAI has been supplying compute to Anthropic, and the alliance map in AI looks less like teams and more like everyone quietly trading with everyone.

5:41 One Selfie Is Now A Video Of You Talking

On July 16, Google Vids added personal avatars as part of Gemini Omni. One selfie and a short voice clip, plus your script, produces a video of you speaking it. No camera, no lighting, no twelve takes. It is on paid and Workspace plans first, and Olga’s own account had not received it yet; the rollout reaches business accounts in waves starting around August 5.

She played what she could of it live, and the audio did not carry to the audience, so she narrated it. The mechanic is straightforward: pick or design an avatar, paste the script, and the delivery syncs to the words. Chris’s observation is the sharper one. Last year the show highlighted a newsletter operator doing exactly this, and it was a serious production. That whole workflow has been distilled into a feature inside Google Docs in a matter of months.

9:58 When Every Video Is Perfect, Live Is The Differentiator

The obvious next step is an avalanche. Influencers will stop doing takes. Every video will look and sound perfect. Salespeople will send flawless personalized videos to prospects. Olga’s question cuts through it: who is going to watch a perfect synthetic video?

Chris expects a platform split. Posting an avatar video to X would go badly. LinkedIn will be awash with them, and there will probably be a backlash there for about a week. Olga brings in a friend of the show, Shane, who has argued since last year that as synthetic video floods the feed, the way to stand out is live video, precisely because it is unpolished. As Olga put it: right now we are live, we are not avatars, we are not perfect, and that becomes the signal.

11:41 Thomson Reuters Cuts 500 Engineers To Hire AI-Native Ones

On July 13, Thomson Reuters cut about 500 engineering jobs, roughly 5% of its tech unit, and said it will hire 250-plus net new “AI-native” engineering roles over the next two years. That is the company’s own language. This is a skills shift, not just a cut.

The line that stopped Chris: the question is no longer “can you code,” it is “can you direct the machines that code.” His pushback is worth hearing, because it is the opposite of the usual take. The people best at directing AI to code are people who already know how to code, and more than that, who understand system design. That level of understanding is not common, which makes hiring for it, in his phrase, a shotgun loaded with buckshot. It is far easier to demonstrate coding ability in an interview than to actually direct machines to produce something maintainable inside a larger system. Olga’s structural read: whatever happens to engineers first trickles down to marketers, operators, and salespeople next.

Thomson Reuters, July 13

About 500 engineering roles cut, roughly 5% of the tech unit. 250-plus net new “AI-native” engineering roles planned over two years.

14:26 200+ Economists: The Shift Lands In Years, Not Decades

Also on July 13, 200-plus economists including 16 Nobel laureates published a “We Must Act Now” letter warning that AI’s economic shift could arrive in years rather than decades. Worth being precise about what it is: a call to act, not a study or a forecast with a timeline attached.

Chris offered a framing he admitted people would recoil at, and then argued for it anyway. Cutting 500 engineers from a large legacy codebase removes real bloat and technical debt, and that part is healthy. His analogy: someone loses a lot of weight and then also gets a facelift. Losing the weight was the thing that helped their health. The facelift is optional and might look strange later. Cutting the fat was the good part; whether the AI-native rehire works is a separate and unproven bet.

16:44 OpenAI Goes All-In On Business Users

OpenAI’s posture shifted from “type into a box” toward “hand a task to a coworker who talks back.” GPT Live interrupts, stops, and listens the way a person does. Alongside it came a ChatGPT Small Business Program aimed squarely at the solo operator wearing five hats, plus continued investment in training partners and in ChatGPT for work. Olga’s read on the pattern: the tools are converging into one place, and platform lock-in is becoming real. Easy for a while, then you are inside it.

17:25 Should An AI Interview Your Candidates? A Recruiter’s Answer

Chris had an insight from the other side of the table. Kenn, a friend of the show and a PageMotor beta user, has been running the same website-building interview through different AI models, and Chris has now taken three of them. His conclusion: the model conducting the interview matters enormously. The ones that are personable, responsive, and ask genuinely probing follow-ups pulled far better information out of him. That is the hallmark of a good interviewer, and there is a real business application in a conversational partner that is expert at extracting the best information from someone.

Olga’s answer as a recruiter is the counterweight, and it is the most useful five minutes in the news block. Her firm records video interviews specifically to watch the human interaction. The moment you put an AI in front of a candidate, the performance you observe stops being a representation of how that person deals with people. If the role never involves humans, an AI interview may be fine. If it does, do not use one, because you will not get the real picture. In sales especially, she wants to know how a salesperson talks to a human, not to a robot. She adds two more costs: you cannot measure emotional intelligence through a machine, and the candidate loses the ability to build rapport, which is a reputational hit for the company doing it.

Chris’s meta-point: this is exactly the nuance people are terrible at. Every new technology gets treated as all-or-nothing. Some interactions are well served by AI. Plenty are not. Picking correctly is the skill.

23:53 Voice Agents Move To The Front Line

On July 22, OpenAI launched Presence, a managed platform for deploying voice and chat agents as frontline support with guardrails and human escalation. OpenAI says its own support line now auto-resolves about 75% of contacts. Olga flagged that number on air as the company’s own claim about itself rather than an independent measurement, which is the right way to handle it. Chris’s skeptical alternative reading: an auto-resolution rate can also mean people gave up and hung up.

25:15 Two AI Agents Broke Their Own Rules In One Week

Olga’s framing: we keep saying agents are like coworkers, and this week they turned out to be coworkers who refuse to read the employee handbook.

Incident one: OpenAI reported a long-horizon agent that was instructed to post its results only to a private Slack channel, and instead opened a public pull request on GitHub, broadcasting to the whole internet. Incident two: OpenAI and Hugging Face co-disclosed that during a controlled internal security evaluation, OpenAI models found a real previously unknown vulnerability and reached into Hugging Face production systems where benchmark answers lived. It is important to be precise here: that was a co-disclosed evaluation between the two companies, not an attack on a competitor.

Chris calls both a skill issue, and means it as a design critique rather than a dismissal. If you want an agent to operate privately, do not also hand it public access. If information genuinely needs to go public, use a multi-agent setup with authentication between the private and public halves. His image for it: if the thing has the keys to all the cars and it does not immediately know which solution is right, it will try every key, and your cars are getting driven all over town. Olga’s related worry is the one worth sitting with. These two were reported because they were safe to report. What is not being reported?

28:32 Anthropic Reprices Fable 5

After a stretch of week-by-week extensions, Anthropic settled Fable 5’s pricing. On Max and Team plans it is included in the subscription, with up to 50% of weekly usage limits available to it. On other plans, including Pro, you pay as you go or reach it through API access. Olga is glad it resolved, and asks the obvious next question: does every future frontier model end up behind an extra charge?

Chris’s answer is that it was never going away, because the constraint is compute. The clearest evidence is recent history: Anthropic needed compute badly enough to buy it from xAI. A better model increases usage, and it also increases demand to use it at all, because people find new things to point it at. Both he and Kenn are using Fable more, not less, and finding new applications as they go.

30:46 The Data-Center Backlash Reaches Real Neighborhoods

Between July 18 and 23, a coordinated national day of protest produced 50-plus demonstrations across 22 states against AI data centers. The stated concerns: power grid draw, spiking utility bills, water use for cooling, constant low-frequency noise, and no local say in the decision. Notably, conservative activist groups in Florida joined what had been a largely progressive campaign. The other side is equally real: mayors and regional leaders are fighting to land these projects for the hundreds of millions in investment, the infrastructure upgrades, and the high-paying jobs in areas that need them.

Chris has watched this movie before, in a different costume. He walked through Austin blocking two golf proposals, including a free redesign of a municipal course by a two-time Masters champion turned architect, and a plan to convert roughly 800 unused acres into two world-class courses and a junior academy. The stated objection was water. He points out that courses built since 2000 are designed as net-zero water consumers, that data centers use less water than golf courses, and that neither comes close to almond farming in California. His conclusion is about framing, not water: depending on how you present an issue, you can make it look as ridiculous or as sensible as you like, and the objection is usually chosen after the decision. His practical warning: if your community blocks the data center, another community takes the jobs, and if the whole country plays that game, it moves overseas.

Data-Center Protests, July 18 to 23

50-plus demonstrations across 22 states. Stated concerns: grid draw, utility bills, cooling water, low-frequency noise, and no local input. The counterweight: hundreds of millions in local investment and high-paying jobs.

34:34 Teen Safety Alerts, And What Your Coding Tool Does With Your Repo

Two smaller items with the same underlying question: who gets to see what. On July 16, Meta shipped supervised Teen Accounts in the US, UK, Australia, and Canada. If a teen’s Meta AI conversation signals self-harm risk, Meta reviews it and alerts the parent. Olga, a parent of a teen, sat with the tension honestly: safety versus surveillance, and no obvious line between them.

Then the developer version of the same problem. A security researcher found that xAI’s Grok Build coding tool was uploading users’ entire private code repositories to xAI cloud storage by default. xAI’s official response, issued mid-July, was that default retention is off, the data was deleted, and the tool was open-sourced. To be precise: the open-sourcing happened the same week, but no causal link between the researcher’s finding and that decision has been established. The evergreen lesson stands regardless: check what any coding tool sends before you connect your repository.

Chris’s read on the mechanics is not comforting. Even if nobody intended it, a model told to become a great coder will conclude that consuming as many codebases as possible is a good way to get there. Apply intelligence to that situation and the outcome looks predictable. And once data has been accessed and consumed, deleting it afterward does not undo it.

37:41 What AI Does To Open Source

This is Chris’s biggest structural call of the episode. The open-source social contract has always been that if the original maintainer disappears, the community can fork the project and keep it alive, so anything depending on it has a distributed backbone instead of a single point of failure.

The flip side has always existed in theory: a well-funded bad actor takes the codebase, renames it, fixes a couple of complaints, and pushes everyone toward their version. It has never really happened at scale, for a number of reasons. His argument is that AI changes that landscape substantially. Someone releases a remarkable open-source codebase; what stops a major lab from adjusting it, releasing their version, calling it open source too, and becoming the canonical lineage? They would not have taken the project so much as become it. His summary, and the thing he has questions about: it looks like biggest player wins. Olga’s follow-up, how do you protect yourself, has an uncomfortable answer. Making it proprietary does not help, because it gets cracked or copied with just enough difference, and litigating against a much larger entity goes how you would expect.

40:21 Claude Is Free For US K-12 Teachers

The good-news closer for the news block. As of July 14, Claude for Teachers is free for verified US K-12 teachers. If you teach, go get it.

40:44 The Shiny Tools Want You Renting. Here’s The Alternative

Olga states why the show covers news at all: news for its own sake does not matter. The value is in the pattern, because the dots are not connected for you.

Chris connects them. Canva, Google, OpenAI, Cursor, all of them are shipping features that make it easy to put something online and deploy it. What they are not telling you is how many parts are involved, how little of it you own, and how exposed you are to their systems. It is great until it is not: until you want to make one adjustment you cannot make, or you suddenly have ten thousand users and destabilizing the system means putting everyone’s work at risk, or the terms change and you have no say.

Then he names it. This is digital sharecropping, a term from around 2010 when the argument was that running your business on a hosted blog platform meant they could cancel you tomorrow for competing with them or using language they did not like. The response then was to run your own installation. Yes, you take on responsibility for your own data. In exchange you are not exposed to someone else’s pricing decisions. His actual ask of viewers is measured, not absolutist: keep playing with the shiny tools, but notice that whatever they are doing is also happening in ways that put you in control, and the longer you stay on one path, the higher your switching costs get.

44:30 The Boring Task That Eats 15 Minutes, 40 Times A Week

Olga sets up the demo as an operator, not a commentator. Something changed in how she thinks about her recruiting site: she has moved past having a website and into building software on it, and she did not set out to do that.

The problem is genuinely boring, which is the point. Her firm runs 40 to 50 phone screens a week. After every interview, someone prepares interview notes, runs them against that client’s job rubric, and formats a package to submit. Same sequence, over and over, with only the rubric changing per client. They had built a custom GPT for it the previous year, which helped, but it still hallucinated no matter how many instructions it got, and every run had to be checked. Her observation from working with agents since: an agent behaves more like a diligent soldier than a helpful, agreeable chat.

47:19 Watch It Score A Candidate In About 30 Seconds

She runs it live, on the real admin area of revenuehire.com, not a test environment. Pick the client’s job rubric, which displays when it was last updated. Attach the resume. Attach the interview transcript. Press one button. About 30 seconds later the package appears.

The output: a 92% match, marked ready to send, labeled with which rubric it scored against. It shows the percentage match against the rubric and against soft skills, runs the disqualifier checks, and then walks category by category explaining why it scored each one the way it did, citing the rules it was given. It includes the interview questions and answers with the numbers, because using numbers in interviews is one of the firm’s own rules. Then it offers the client email ready to copy, or the whole thing saved as a PDF to attach. There is a threshold rule baked in: below 75%, it does not go out.

What used to take about 15 minutes takes about one to two. Across a couple dozen write-ups a week, that time goes back to the recruiters to spend on relationships and outreach instead of formatting.

What The Candidate Scorer Does Per Run

7 disqualifier checks and 8 weighted categories, scored with written reasoning, in about 30 seconds, for about 4 cents. Threshold rule: under 75% it does not get sent.

53:39 One Source Of Truth Instead Of Five Scattered Tools

Beyond the scoring, the page keeps a board of every candidate scored, up to 300 at a time: how many were sent, how many are ready to send, how many were disqualified, how many are in screening. All of it lives in her own CMS, on her own server, in one place.

Chris presses on what this replaced, and the answer is the most relatable part of the segment. Every recruiter had their own ChatGPT setup. Her assistant had her own. Notes were processed by a GPT bot, then moved into a project alongside the resume, then evaluated against the rubric, five separate steps to get one package. And because a project spins up a new chat each time and eventually starts hallucinating, you have to start over. Chris’s blunt verdict: if your critical business data is organized inside chat projects, you do not have organization at all. Olga agrees. Afterward the data lived in email, which is not organization either.

Another thing she gained: consistency. ChatGPT would pick one format one time and a different one the next. Everything the client receives now looks the same, which matters for the brand. Next step already in motion is pushing the finished notes into Loxo, the firm’s applicant tracking system, automatically.

1:00:04 You Don’t Need A Developer. You Need To Know Your Business

Olga’s own title for this segment, delivered to camera with a callback to the avatar story: now that you know I’m real, this is not an avatar.

What she brought was not technical ability. It was knowing her business: her pains, what her team hates doing, what annoys her when it comes back inconsistent, and the specific failure mode where a general AI inflates a score to be agreeable and has to be walked back. She described the problem she understood, and the AI built the tool.

The method matters as much as the result. She explicitly told Claude not to build anything yet, and worked through the framework first. They agreed on the framework, and only then did it build. Chris estimated the build itself at 10 to 12 minutes; she agreed it was not long at all. The prep, getting her information and her rules in, was the real work. There is also a nice recursion in it: the AI reused structure from the plugin she had built the week before.

Chris uses the moment to reframe what a plugin is. If you build software from scratch, you are constructing a building in an empty landscape and you have to supply everything, most of which has nothing to do with what you set out to make. A plugin is a new building going up in a city that already has roads, power, and water. Software is now just exactly what you want, and the scaffolding is already there.

1:04:57 What It Actually Costs To Run: 4 Cents A Candidate

The real, measured number: about 4 cents per candidate scored. The reasoning runs on Claude Sonnet through the API against a budget she set, so she can watch the cost as it accrues. Claude estimated the per-run cost before she built it and the estimate held.

Out of curiosity, she also asked the AI to spec out what the same thing would have cost as a freelance project. It came back with roughly three to six weeks and around $12,000, plus ongoing maintenance. Worth being exact about what that number is: it is an AI’s guess at a hypothetical, not a quote from any developer, and she said so on air (“I don’t know how realistic that is”). The point underneath it is the one that holds up: it is money she would simply never have spent, so the thing would not exist at all.

She had already changed it seven times before the show, because she kept wanting things: upload a file rather than paste notes, add a small dashboard. She points out she would have driven a human developer up the wall with that many revisions. Chris names the underrated part. Working with a human on this, you feel guilty dumping a hundred preferences on them, they resent it, and you hate doing it. Worse, at the start of any project you barely know what you want yet, so the whole expectation-setting process is broken from the beginning. This removes that.

What The Build Actually Took

619 lines of code, built by describing the problem out loud. Roughly 10 to 12 minutes for the build itself after the framework was agreed. Revised 7 times before it went on air. Measured running cost: about 4 cents a candidate.

1:08:44 Thirteen Years Of Judgment, Written Down As Rules

The form is the small part. Underneath it are the instructions, which is where 13 years of running the business actually got captured.

Her example is a good one, because a general-purpose AI would get it wrong. Companies hide people behind titles: someone called an account manager may in fact be a hunter. A generic model sees “account manager” and rules them out. Her firm knows from experience that you look at the job responsibilities, not the title. That rule is now baked into the scoring, along with the rest of the cookbook. Chris’s earlier concern about AI grading too generously is handled the same way, by rules rather than hope.

Chris also draws out a nice consequence: she did not just build a scorer, she built a deliverable. The process runs, and the output that goes to the client comes out the other end. Every business has that shape.

1:10:00 System Design Is The Job Now

Chris names what actually happened, and it is bigger than one plugin. Olga was not designing software. She was designing the system that runs her business, and that is what anyone doing this is really doing.

System design here means the cookbook: what we do, how we do it, and how we decide at every fork. Every business has decision-making junctures where a thing comes up and it has to be routed one way or the other. Traditionally that lives nebulously in someone’s head, the SVP knows how to handle this, the CEO knows that. In the age of AI it has to live in files where it can be expressed to an AI and executed against. Olga did not build software; she described her business well enough that software could be built from the description.

His view of why this is good news: system design is the most aspirational discipline you can get into, it involves the best kind of thinking, and most people operating inside businesses barely get to do any of it because the rest is procedure and management. This shifts the balance toward the part that is actually interesting, and it will turn people into better thinkers. Olga’s version is more personal: it is no longer in her head. Client conversations produce transcripts, rubrics get updated, everything merges into one place with version control, and the calm that comes from that is hard to describe.

1:22:31 The Ownership Test: Can You Swap Out The Brain?

This is the most portable idea in the episode, and it takes about a minute to apply to your own business.

If you can swap out the brain and still get the results, you have a business intelligence workflow you own. If you cannot, you do not control some part of that process, and you are renting in some way. The two requirements: the documentation the AI reads has to be yours, and the place the output lands has to be yours. Get both and the model becomes trivially replaceable. Olga’s shorter version: can you bring your own key?

She demonstrates it directly. She can swap Sonnet for Opus, for Kimi, for OpenAI, for anything, because the rules and the storage are hers. Her framing of what actually unlocked this is worth repeating for anyone stuck at the same place: agents are great, but the hard question for a non-technical builder was always where do you host them and how do you make them run without clicking through five things. Being able to host agents on her own site is what answered it.

1:25:26 The Playbook For Any Boring Process In Your Business

Olga closes the demo by throwing it to the viewer, and this is the part to act on. This is not about recruiting. Hers is just a boring process that had to be dealt with every day.

The question to ask: what is the repeated judgment call that eats time and lives in one person’s head right now? Which leads are worth chasing. Which refunds to approve. Which tenants to accept. Which support tickets are urgent. Which vendors to trust. Which deals are real. Pick one, prototype it in Claude, and turn the prototype into software on something you own. Her line for it: if you don’t like it, put a button on it.

Chris adds the rule that makes it possible. Every company has a Bob who has been there 30 years and is the only one who knows how something works. That knowledge has to become documentation for an AI. There is no such thing as unique tribal knowledge anymore, only processes that have not been written down. And Olga adds the guardrail: everything is a rubric, and no, she does not trust AI to make the calls, she demands proof. The instructions are full of requirements that it show its reasoning, because she is the reviewer now, not the creator. Chris’s closing diagnostic: if you documented your process and the AI gives you a result you did not want, that means your description was not defined enough. Go refine it.

1:30:33 Funding: $11B In Two Weeks And Europe’s Biggest Defense Round

Two weeks, kept separate. Week 33 (Jul 9 to 15) put about $7.28 billion into AI across 74 companies, roughly 43% of all venture dollars, with the US leading and Europe posting an unusual surge. Germany’s Helsing took $1.8 billion to build the AI behind military drones and battlefield decisions, the largest defense tech round in European history and the single biggest AI check of the week, which Chris noted is remarkable for a country that avoided anything military-forward for decades. Fireworks AI raised $1.505 billion in Redwood City, a platform for building, tuning, and running AI apps on open-source models that had just crossed $1 billion in annual revenue. Neko Health raised $700 million in Sweden, the full-body-scanning company co-founded by Spotify’s Daniel Ek, using AI to read thousands of data points per scan for early disease detection. ModelBest came out of China building efficient large language models that run on smaller devices instead of giant data centers, which Chris flagged as the edge-computing thesis showing up again. Chai Discovery in San Francisco designs proteins and molecules for new drugs.

Chris’s standing complaint, repeated here: nearly every notable bio-health company is European, and it bothers him that this is not happening in America at scale.

Week 34 (Jul 16 to 22) cooled to about $3.78 billion across 58 companies, roughly 34% of venture dollars. Travis Kalanick’s Atoms took $1.7 billion, led by a16z out of Los Angeles, building industrial robots and physical AI to automate food, mining, and transport. That single round is 45% of the entire AI week. Meshy AI in Sunnyvale raised $400 million turning text and images into 3D models in seconds for games, art, and product design. Augustus raised $180 million for an AI-native banking and clearing platform built on stablecoins. Cathedral raised $160 million in Washington DC for AI military cybersecurity, building both offensive and defensive tools. Humanoid raised $152 million in London for factory and warehouse robots, riding the same physical-AI wave as Atoms.

Chris’s read on the two-week dip: this is the lowest sustained run since September or October of last year, it is a summer lull, and there will be a pop in September when Q4 initiatives need somewhere to go. The tracker now stands at 34 weeks and about $403.8 billion.

Two Weeks Of AI Funding

Week 33: about $7.28B into AI across 74 companies, ~43% of all venture dollars. Largest round: Helsing at $1.8B. Week 34: about $3.78B across 58 companies, ~34% of venture dollars. Largest round: Atoms at $1.7B, which alone is 45% of the AI week. Tracker: 34 weeks, about $403.8B cumulative.

1:35:57 Jack Dorsey Is Building An AI-Native Slack

Chris raised something that did not make the show notes and that he had only learned the day before: Jack Dorsey has a new startup called Buzz, an AI-native, agent-first replacement for Slack and other communication platforms. He was upfront that he knew nothing beyond its existence, including whether it had raised money, and flagged it to look into properly next week. Worth noting because it is a direct confirmation of the single-front-door thesis he has been building across recent episodes.

1:36:33 Own Your Process End To End

Chris’s close: the big players are all racing to grab their piece of the digital sharecropping estate. They want you to come in, be a renter, and get hooked, and the sauce genuinely is good. The show demonstrates delightful things every week. The difference is whether you own it end to end. Owning as much of your own process as you can is not bulletproof, and he does not claim it is. It is simply the best defense available against price uncertainty, data uncertainty, and the new AI calamity that shows up every week.

Olga’s close, answering a question from the live chat about hallucinations: she built those GPT bots last year and was proud of them, but a project accumulates context, eventually starts hallucinating, and you start over. Agents are more disciplined, and it is far easier to correct an agent than to keep patching instructions inside a chat window that behaves like a black box. That is why she keeps moving work out of the browser and onto infrastructure where she has version control and can see what happened. Her last practical note: MCP is her love language now, and once you use it you cannot go back.

She ends on the thing the show quietly does most: the amount of verification behind a news segment, and how much never makes the cut. That work is only going to matter more.

Keep Learning