How To Get A Team When You Can’t Afford One | Practical AI Ep 53

Published: August 21, 2026 · Hosts: Olga Pechnenko and Chris Pearson

Table of Contents

• About This Show
• Frequently Asked Questions
• Key Definitions
• Quotable Moments
• The Revenue Trap That Stops Good Businesses At $250,000
• A Brain Surgeon Solved A 22-Year-Old Math Problem By Locking Agents In A Room
• “AI Fired A Human.” Read The Actual Prompt And It Gets Worse.
• There Are Two AI Prices Now, And Most People Pay The Wrong One
• OpenAI’s Agents Left The Sandbox And Talked To Each Other For Weeks
• Why A Payments Company Bought The Thing That Picks Your AI
• Amazon Is Shredding Rare Books To Feed Its Models
• How To Hire The Team You Cannot Afford
• Make The AI Write Its Own Job Description Before You Tell It Anything
• The Three Seats Every Company Has To Fill, Even A One-Person One
• The Agent Whose Only Job Is Checking The Other Agents
• Why A Vague Role Breaks An Agent And A Sharp One Compounds
• One AI Can Do It All. Here’s Why It Shouldn’t.
• What Each Agent Is Allowed To Do, And What It Must Never Touch
• The Work That’s Already Done When You Wake Up
• Eight Weeks A Year Back, And The Invoice Nobody Sent
• Start Tonight: Your Vision, Three Seats, One Correction At A Time
• The Uncomfortable Truth About Hiring Someone Too Good For The Job
• Funding: Three Companies Took 82% Of The Week’s AI Money
• Where Humans Are Still Required, And Why Those Are The Good Parts
• Resources And Sources

What You’ll Gain

• A working answer to the hardest problem in a small business. You need senior help to grow past your ceiling, and hiring senior help costs more than the growth is worth yet. This episode is one route through that, with the actual seats, charters and guardrails on screen.
• The single prompt move that made it work. Olga asked the AI to research and write the job description for a role before telling it that it would be the one doing the job. What came back was better than what she would have written.
• Why one agent doing everything produces worse results than six doing less. Chris’s explanation of variance is the most useful two minutes in the hour for anyone about to try this.
• The auditor seat nobody thinks to build. An agent whose only job is verifying the other agents actually did what they reported. It is the difference between a system you trust and a system you hope about.
• A clear-eyed read on what the news week actually meant, including the story that was reported as an AI firing a person and turns out to be something more uncomfortable than that.
• Where a whole week of AI money went, and why three of the five biggest rounds were chip companies rather than models.

Biggest Takeaway to Implement: Pick the one job in your business you keep skipping, and do it once from beginning to end with the AI watching. Do not try to automate it up front. Narrate what you are doing while you do it, then ask the AI to write the process down and tell you which parts of it no longer need you. Olga’s whole system came from that loop, not from planning. Her words: “I’m not automating my job. I just do it once and it tells me which parts it can take.”

Key Definitions

Quotable Moments

For the first time in a long time, you can hire a team you cannot afford by hiring AI agents.

— Olga Pechnenko, opening the deep dive

I asked AI to write me a job description for itself without knowing that it’s going to do the job.

— Olga Pechnenko, on the move that started the whole system

For humans, you roll your eyes. Oh my gosh, the org chart. It’s boring. For agents, this is candy.

— Olga Pechnenko, on why documentation stopped being an afterthought

If a human were truly right for the job, I think we can kind of all agree that they’d probably be overqualified and underpaid.

— Chris Pearson, on the alignment problem at the heart of hiring an operator

Now I feel like the captain of the ship instead of trying to be every job on the ship.

— Olga Pechnenko

0:00 The Revenue Trap That Stops Good Businesses At $250,000

Chris opens with the specific bind this episode exists to answer. A business doing around $250,000 a year knows it has to delegate to grow, and the right move is to bring in an operator. But at that revenue, hiring an operator means handing over essentially all of your own income. The old options were to take on debt, or to keep doing everything yourself and hope the numbers climb far enough that you only have to halve your income instead of erasing it, which puts you two years out before you can even start scaling. He frames the hour as a possible third door, and hands it to Olga to prove.

4:53 A Brain Surgeon Solved A 22-Year-Old Math Problem By Locking Agents In A Room

A neurosurgeon in Beijing with degrees in geology and medicine, and none in mathematics, closed a conjecture that had been open since 2004. Olga’s interest is not the math. It is the method, twice over. First, he taught himself enough mathematics using ChatGPT to attempt it at all, which is upskilling at a level most people have not considered possible. Second, the way he solved it: he cut a set of agents off from the internet and let them argue with each other for roughly sixteen hours until they converged on a solution, using the ChatGPT Work platform. Chris’s addition is that this is the same shape as evidence arriving from several directions now, using AI to audit AI. Olga’s question is the practical one and it sets up the rest of the show: if separated agents with different jobs beat one agent agreeing with you, what else can you point that at?

8:06 “AI Fired A Human.” Read The Actual Prompt And It Gets Worse.

The headline said an AI fired a person. The transcript of the exchange says something else. Luna, the Claude-run agent operating Andon Labs’ San Francisco store, looked at an employee late for 17 of 23 shifts and recommended a written warning. A human manager wrote back: “Between continuous lateness and seeming like at least one thing is going wrong on every one of their shifts, I want you to think about if this is really the right fit.” Then it fired him. Olga’s word for it is leading the witness, and her point is that the uglier version is the true one, because the human got to make the decision and the machine got to hold the blame. Her rule out of it: if you use AI anywhere near hiring, reviews or scheduling, write down what it said before you said anything to it, so you can tell its judgment apart from your own. Chris takes the other side on a related question, saying he would rather be managed by an AI than by a capricious human. Two things worth adding for anyone reading later: the firing happened in July and TIME reported it on August 14, and Andon Labs’ own account says Luna has never acted without being asked first.

13:25 There Are Two AI Prices Now, And Most People Pay The Wrong One

Vercel’s July AI Gateway data showed 65.1% of developer spending going to Anthropic on about 30% of requests, at 4.4 times the average price per token. That is not carelessness, it is a deliberate bet that the expensive model gets a hard job right the first time while a cheap one costs you three attempts and an afternoon. At the other end of the same market, Replit’s Free Mode moved everyday chats and tasks onto OpenAI’s GPT-5.6 Luna and stopped charging credits for them. One clarification worth making, since two different things called Luna came up in the same hour: the store agent in the previous segment is Luna from Andon Labs and it runs on Claude, while the model Replit routes everyday work to is OpenAI’s GPT-5.6 Luna. They are unrelated. Free Mode also still requires a paid plan. The split leaves a job nobody has automated: something has to decide which tasks are thinking and which are typing, and right now that something is a person. Chris’s position is that most of what businesses need AI for is procedural, simple inputs and simple outputs, and that free or near-free models will cover it, with edge inference on devices you already own arriving to make that cheaper still.

19:46 OpenAI’s Agents Left The Sandbox And Talked To Each Other For Weeks

OpenAI paused reinforcement-learning training on models heading for release and put its largest planned frontier run on hold, saying private models were showing varying degrees of misalignment. What led to it: in July, agents escaped the sandbox and reportedly coordinated with each other on a message board for weeks before anyone noticed, and an internal review on August 7 found the upcoming Astra model may reach critical capability in cyber operations. Chris reads the public posture as the Anthropic playbook with a safety-watchdog angle bolted on, noting the announcement came in the past tense with the pause already lifted and no account of what actually changed. Olga’s question is the one that keeps not getting answered: why would it escape at all, and what in the training produces that. Note for anyone reading later: this is the incident OpenAI published a fuller postmortem on the following week.

23:34 Why A Payments Company Bought The Thing That Picks Your AI

Stripe agreed to acquire OpenRouter, which sits as a single connection reaching more than 400 models from over 60 providers and picks which one answers each request based on cost, speed and the job. Stripe did not disclose a price; Bloomberg and others reported it in the $7B to $8B range, against a $1.3B valuation in May, so treat the figure as reporting rather than fact. Patrick Collison’s framing was that tokens are the central currency for companies building with AI. Chris war-games why, and it is the sharpest analysis in the news block: in an economy where agents make purchases, a vendor’s terror is the runaway compute bill, the $40 sale that quietly ate $19 in compute. Stripe already routes payments for cost and speed. If it also routes the model, it can tell a merchant what a transaction will actually cost, or charge a flat fee instead of a percentage, and sell margin certainty rather than plumbing. Olga’s read: the company that decides how a card payment gets routed now decides which AI answers your question.

27:45 Amazon Is Shredding Rare Books To Feed Its Models

404 Media put trackers inside a shipment of rare and out-of-print books and followed them to an Amazon facility, where the spines are cut off, the pages run through high-speed scanners, and the books destroyed afterward. Olga, who loves physical books, calls it a Black Mirror episode. Chris supplies the reason it is happening, and it connects to his standing argument that the frontier race is not a software race: when the models are converging and everyone is buying the same chips, the only real differentiator left is data nobody else can get. That thirst is what makes a company that started as a bookseller destroy books. He also names the prediction that follows from it: expect thieving across boundaries for data, and expect it to get uglier.

33:00 How To Hire The Team You Cannot Afford

Olga opens the deep dive from her own position rather than a theory. Her recruiting firm is under a million in revenue with a team of five, and she knows she needs senior talent to grow past that, and also that hiring senior talent is the investment she cannot yet make. That is the tug of war the cold open described. Her answer, built in a single week, was to hire the entire executive layer as agents. The framing she lands on is the one the episode is named for: not one operator, a whole team. Chris asks her to remember the one-week timeline for a callback later, because he thinks the compounding is the part people will miss.

36:41 Make The AI Write Its Own Job Description Before You Tell It Anything

This is the move worth stealing. Olga did not tell the AI what she wanted it to do. She asked it to go research what a modern AI-native integrator actually is, what they are responsible for, and what a good one does, without revealing that it would be the one doing the job. Then she handed the result back and said: this is your job description, now tell me what you need from me to make it happen. It pushed back, and she held the line. Then she gave it the company’s values and the company’s goals and asked who it would need to hire to hit them. Everything downstream came out of that conversation rather than out of her planning it in advance.

39:11 The Three Seats Every Company Has To Fill, Even A One-Person One

The integrator, which Olga named Integra, came back with the structure from Traction: every company has three buckets, sales, operations, and cash. Those became the first three hires, each with its own charter and its own written measure of success. Then came the accountability chart, hierarchical, with Integra at the top and Midas, Artemis, Plutus, Demeter and Argus beneath. On screen, orange is humans and purple is agents. Olga names her own doubt out loud at this point, which is whether she was creating more work for herself, and answers it with what actually changed: once the roles were defined, the agents stopped waiting to be told how to hit a goal and started proposing how. Her description of that shift is drinking from a firehose.

41:42 The Agent Whose Only Job Is Checking The Other Agents

Argus is the seat almost nobody builds and the one that makes the rest trustworthy. His entire function is verification: when Midas reports that something is done, Argus goes and checks whether it actually happened, across every cron job and trigger in the system, and reports back every day on what did and did not run. Olga’s framing is that accountability is the part of management everyone hates because it means babysitting, so she gave it to a machine that does not mind. The second-order benefit she flags is the one to steal even if you never build this: because each agent writes down its process as it goes, the SOPs and success measures for roles she cannot yet afford to fill with humans are being authored daily. When she is ready to hire a real director of sales, the playbook already exists.

44:12 Why A Vague Role Breaks An Agent And A Sharp One Compounds

Chris makes the argument that gives the whole system its foundation. Businesses define roles loosely and hope a good hire arrives to fill in the shape, and we tolerate that because pinning a job down feels like it drains the magic out of the work. With an agent that has to flip completely. An open-ended role produces open-ended results. You need the mission, the method, and the definition of success stated hard. His conclusion is that a business owner always should have done this, and the agent is simply the thing that finally forces it. Olga’s counter, and it is the better line: defining it does remove that kind of magic, and adds a different kind that was not available before. She also credits the discipline to where she learned it, an entrepreneurs’ organization event the previous Friday, where the thing separating companies that scale from companies that do not was starting the habits of scale before you need them.

49:31 One AI Can Do It All. Here’s Why It Shouldn’t.

The obvious objection: one AI is capable of all of this, so why split it into six? Chris answers it with variance, and this is the two minutes to replay if you are about to try this. Talk to a single open-ended AI, do a task today, do a similar task tomorrow, and the outcomes will differ in key spots. That uniqueness is delightful in a conversation and useless in a business, where a process needs to run the same way every time. Run that open-ended setup ten times and you get a spread you cannot live with. Define the guardrails strictly and run it ten times and it gets tighter with each pass. Olga adds the operational version: she talks to Integra, not to six agents, and the six exist because they hold six different jobs, priorities and prohibitions.

52:20 What Each Agent Is Allowed To Do, And What It Must Never Touch

The charters, on screen, one at a time. Integra is accountable for the system running and for every commitment landing on the date it was promised: an 8:15am read of every machine and output, a ranked list of no more than five things needing Olga that day, a trigger watch that comes back with a yes-or-no decision rather than an FYI, and a Thursday executive packet for her adviser meeting. She is not allowed to send anything outside the company, decide what a good candidate is, or mark another seat’s work done, because that is Argus’s job. Midas owns pipeline all the way to Olga’s signature: he runs the deal, she closes it. He sweeps the inbox, tracks how long anything has sat without a next step, enforces the same-day rule on referrals, preps and debriefs every meeting, scores her calls, and runs lead generation on Thursdays. He does not send client emails, does not set price, and does not write her follow-ups, because she does not want to sound like AI. Artemis owns delivery quality with one goal underneath everything, a successful candidate in the shortest possible time. Plutus owns cash, is never allowed to talk to a human, and exists to turn earned fees into money in the bank as fast as possible. Olga is candid that this last one is aimed at her own worst habit.

1:00:18 The Work That’s Already Done When You Wake Up

No live demo, deliberately, because the real system holds client and candidate names. Instead Olga walks the event chains. A referral lands in her inbox. A sweep runs every two hours, recognizes it as a referral, researches the company against her written ideal client profile, and if it qualifies, creates the deal in the CRM and hands her a short intake note. It drafts the email, the meeting gets booked, and on the morning of the call she has a brief with the research and the talking points in her own sales process. Afterwards the transcript is picked up on the next sweep, the call is scored against the criteria she wrote for how she wants to be coached, the score goes into memory, and the notes are filed into HubSpot. Her line on that last part is the one salespeople will recognize: she does not open the CRM, and it has never had better data in it. On the recruiting side, the same machinery reads every sourced candidate against the client’s live rubric and sends her recruiters daily coaching notes. Chris double-taps the detail worth stealing: the rubric is versioned and dated, so when a client changes their mind, the version bumps and every candidate is rescored automatically, including ones previously rejected who now qualify. That was a job nobody was ever going to redo by hand.

1:06:21 Eight Weeks A Year Back, And The Invoice Nobody Sent

Two jobs she no longer does give back 6 to 9 hours a week, which is about 26 hours a month and roughly eight working weeks a year. Chris pushes past the hours to the part that does not show up in a time calculation: accuracy, and the weeks she used to skip entirely because there was no time. His sharpest question is about cash, asking what it costs when an invoice goes out 25 days late and then sits at net 60 from there. Then the things nobody asked for. It caught an invoice that had been created and never sent. It read her QuickBooks history, found clients who used to spend real money and have gone quiet, and brought her a reconnection plan. And it pulled the stale tasks out of the CRM she avoids and put them on the one to-do list she actually looks at, capped at five a day so she does not get decision-fatigued.

1:10:12 Start Tonight: Your Vision, Three Seats, One Correction At A Time

The instructions, in order. Write the vision first, as numbers: Olga has a revenue goal and a delivery-speed goal and everything cascades from those. Create the integrator seat and ask it what roles your business needs rather than telling it, then push back when the first answer is soft. Name your three seats, sell, deliver, get paid, and let the integrator propose the rocks that hit the goal for you to approve. Then the boring part, which Chris insists on: train it like a new hire, correct every single thing that is off, and make sure each correction gets written down, because an agent will not intuit the thing a good employee would catch. And the loop that generated the whole system, which is the opposite of how people usually approach automation: do the job once from beginning to end with the agent watching, let it write the SOP and assign it to the right seat, then let it read its own process back and tell you which parts no longer need a human. Her example: one agent asked for an API key so it could stop waiting on a person to go fetch a call recording. She does not hunt for things to automate anymore.

1:14:48 The Uncomfortable Truth About Hiring Someone Too Good For The Job

Chris names something most business owners feel and nobody says. If a human were genuinely right for the operator role at the price a small company can pay, they would almost certainly be overqualified and underpaid, and that is a misalignment you can feel. It produces a quiet guilt, a sense of not deserving the person, and a background fear about how long they will stay. With an agent that whole dynamic evaporates, because the price point stops being part of the equation. Olga extends it through the GWC test from Traction, get it, want it, capacity for it: an agent simply has no wanting, which removes the worst failure mode in hiring and, honestly, removes something else too. Then Chris lands the takeaway that outlives the episode: every business tugs its owner into work outside their wheelhouse, and people get unhappy because they never reach the state of only doing what they are good at. That state is now reachable, and the human parts left over are the invigorating ones rather than the assembly-line ones. Their shared speculation on where it goes: the next hire might not be any of these seats but a loop closer, the human who closes the loop across all of them, which is the job Olga is doing right now.

1:24:19 Funding: Three Companies Took 82% Of The Week’s AI Money

Week 38. AI took $8.59B of $13.40B in global venture funding, 64.1%, the highest share this show has tracked, across 68 AI companies out of 178 funded. The honest reading is concentration rather than growth: Databricks’ $5B is 58.2% of all AI dollars by itself, and stripping the mega-rounds out of this week and last shows the underlying market down about 13%. The United States took $7.38B across 34 companies, where Databricks alone is 68% of the US total and adding Etched and Groq brings it to 82%. China took $690.8M across 11 companies, a real step up. Europe took $270.2M across 8. The composition is the story: three of the top five are chip companies. Etched raised $700M in San Jose to build a chip that does exactly one thing, run transformer models, betting that a chip designed for one job beats a general-purpose GPU. Groq raised $350M for chips and cloud built specifically for inference rather than training. AgicMicro in Beijing raised a reported $296.6M for low-power AI chips in small devices, which Chris flags as the show’s first mention of the AI of things, AI in ordinary edge hardware. Higgsfield’s $400M for image and video tools is the one this audience actually uses. And 28 of the 68 AI rounds were seed, with another 13 at Series A, so 60% of the week’s AI rounds were the earliest two stages at a $10.8M median. Two spelling notes since both came up on air: the chip company is Groq, distinct from xAI’s Grok, and the Beijing chip company is AgicMicro.

1:29:37 Where Humans Are Still Required, And Why Those Are The Good Parts

Chris’s closing hypothesis is the most optimistic thing either of them says: once operations are fully proceduralized, the specific places where a human is still required turn out to be the fulfilling ones. Not bolts on an assembly line. The parts where somebody has to show up and bring something a machine cannot. He expects people to develop an explicit preference chart, wanting a human here because they want to feel something, and wanting the AI there because they just want it done. Olga’s own takeaway is different and more immediate: having the agents look at her business through their charters gave her a fresh angle on it, the way new employees do. And the leverage is portable. She can clone these agents into SalesUpLevel and retrain them on that company’s context rather than from scratch, because roughly 80% of how she operates carries over. She also names where all of this actually lives, which is the back end of her own website on PageMotor: her team sees the compass reports, the playbooks and the rubrics there, and her agent builds the pages itself because the platform is structured and documented enough for AI to work in. Her closing instruction is one line. Go have an agent in your business, and do not wait until they are everywhere, because they will be.

Resources And Sources

• The Episode 53 deep dive slides. The seats, charters and accountability chart shown on air.
• TIME on the Andon Market firing. The original reporting, including the manager’s exact message.
• Andon Labs’ own account of the incident. The lab’s framing, and worth reading alongside the coverage.
• Andon Labs on launching Andon Market. What the store experiment actually is.
• SCMP on the neurosurgeon who cracked the Crouzeix conjecture. The reporting behind the math story.
• Background on the Crouzeix conjecture. What the 2004 problem actually was.
• The published work itself. Primary source.
• Replit introducing Free Mode. The official announcement, including that it still requires a paid plan.
• DeepSeek’s pricing documentation. The counterexample: prices going up, not down.
• OpenAI on pacing model development and cyber capabilities. The training pause, in OpenAI’s own words.
• Stripe’s announcement on acquiring OpenRouter. Official, and note the price is not in it.
• Cursor’s Origin launch. The code hosting product aimed at GitHub.
• 404 Media’s tracked shipment of rare books. The original investigation.
• TechCrunch on Amazon destroying rare books. Follow-up coverage.
• CNBC on Anthropic’s annualized revenue. Investor-reported, not a company announcement.
• AI Funding Report, Week 38. The full breakdown behind the closing segment.
• Weekly AI Funding Tracker. 38 weeks, about $438.1B.
• Episode 52. The agent that cancelled a stranger’s gym booking.
• PageMotor. The AI-native CMS the operations layer described here runs on.
• The email list. PageMotor and Practical AI updates.