How AI Native Are You, Really?

What You’ll Gain

• A scale you can put yourself on today. Seven levels, from “I don’t really use it for work” to “work happens while I’m not there,” each with what it sounds like out loud, what keeps you stuck there, and the single next move.
• Six questions that end the guessing. They work on a job candidate, on a team, and on yourself. The first one alone separates beginners from everybody else, and it takes about a minute.
• The difference between a prompt, a workflow and a skill, which is the thing that decides your level. A single task is barely leverage. A documented sequence is a system.
• Real numbers on AI and jobs, from a central bank rather than a headline, including the number that explains why the job market feels worse than the layoff data says.
• What actually works with an AI voice agent, and what the law will not let you do with one. Plus a free template you can clone and point at your own offer.
• A security beat that applies to everybody learning to build with AI: the risk does not need you to click anything. It needs you to open a folder somebody sent you.

Biggest Takeaway to Implement: Find your level honestly, because the honest read is usually one level lower than the flattering one, and only the honest one comes with a next move. Chris’s version of the assignment: take the transcript of this episode, hand it to your AI, and ask it where you land and what gets you to the next rung. Olga’s: a beginner describes something they have, and somebody at the top describes something that happened while they were not in the room.

0:00 What Happens When You Ask People To Rate Their Own AI Skill

Chris opens on the thing that set up the whole hour: they started asking people to self-assess their AI fluency, and the answers came back generous. Olga expected honesty, on the grounds that an honest answer is the one that helps you. What she got instead was the dating-market answer, everybody an eight out of ten. Chris frames why it matters beyond the joke: every company is trying to gain an edge with AI, that edge runs through existing people and existing processes, and the human half is the part nobody can measure. You cannot move somebody up if you do not know where they are standing. He also teases his own segment for later, which he describes as cloning, and insists it is not the sheep kind.

2:23 OpenAI Shipped Its Most Dangerous Model Into The Plan You Already Pay For

OpenAI launched GPT-6 Astra the day before the show. It is built for multi-step work across applications: using your computer, browsing, writing code, filling spreadsheets, building documents, all in one continuous run. It reaches Plus, Pro, Business and Enterprise over the following days, with no subscription price increase. The thing that makes it a story rather than a release note is the timing. Two days earlier OpenAI published that this same model is the first it has ever rated Critical for cybersecurity capability under its own preparedness framework, which in OpenAI’s own definition means a model that can find and exploit unpatched flaws in hardened real systems. They shipped it anyway, with the advanced cyber access gated to testers. Chris’s reaction is dry: he is sure people will use it responsibly. Then he spends a minute on what he actually saw people build with it, which is 3D — somebody rendered all of Manhattan at street level over a week and a half, and there are camera sweeps through fully built environments that did not exist before.

5:56 The AI Layoff Number Is 4%. The One You Actually Feel Is Somewhere Else.

The New York Fed published survey data on September 1. Among firms in its region, 61% of service firms and 51% of manufacturers now use AI. Among those users, 4% of service firms reported AI-related layoffs in the past six months and zero manufacturers did. 13% said AI actually increased their hiring, often to run the tools. More hired than fired. Then Olga puts up the third number, which she points out everybody leaves out: 15% said they hired fewer workers than they would have without AI. That is larger than the layoff figure. Chris names it the dog that doesn’t bark, and calls it a ghost firing — the role that never gets posted, which no survey catches unless it asks. It is why the street feels worse than the layoff numbers read. One honest limit, said on air: this is a regional survey covering New York and northern New Jersey, not the country.

7:29 Eighty Agents, One Cancelled Subscription, Nobody Hired

Olga makes the abstract number personal with her own count: 80 agents now running across her businesses, and she has not hired anyone. She is rethinking what the people who do work for her should be doing instead, because her assistant is no longer doing most of what she was hired for — AI handles email management and a long list of other jobs. That same morning she cancelled Canva, because Claude does her design now. This is the 15% from the survey, from the inside: not a firing, a job that quietly stopped needing to exist in its old shape. She also flags a hiring pattern she is now seeing at client companies, which is a request to prove AI cannot do the job before a req gets approved.

8:47 Only 22% Have Scaled AI. 85% Are Spending More Anyway.

A private research firm rather than a central bank: Gartner surveyed 1,303 leaders at companies above $50M in revenue. Only 22% have scaled AI across multiple business units or adopted an AI-first approach (Gartner does not separate those two groups). 11% do not know what they spent on AI last year. 85% plan to spend more anyway. Chris explains the mechanism from the buying side: budgets already existed by department, some got reallocated to AI without a plan, and the point was to be able to show the executive team a line item rather than to get a result. He describes analysing one Austin company’s AI transformation package the day before and finding it thin — a good salesperson arriving at the right moment into budgets that were already there. Olga adds the other half of the leak, which is the licence sprawl underneath: $200 here, $100 there, all on a card, nobody counting.

11:15 Uber Cut 3,300 And Insisted It Wasn’t AI. California Says Your Boss Can’t Be One.

Two policy-adjacent beats. First, Uber cut roughly 3,300 roles on September 2 and said the reason was organisational complexity, explicitly not AI. Worth noticing because Uber attributed an earlier round this year to AI, which makes the denial the interesting part. Second, California’s legislature approved the No Robo Bosses Act over the weekend of August 30 and 31 and sent it to the governor. It is not law, an earlier version was vetoed last October, and if signed it does not take effect until July 2027. Chris’s objection is practical rather than ideological: he would run every process through AI and simply not call it the boss, routing the report through a human on paper. Then he asks what a boss actually is, and answers that hiring, firing, accountability and coordination are procedural — the part that is not procedural is managing psychology. Olga’s observation just before this one is also worth keeping: the larger the company she talks to, the more likely the answer is Microsoft Copilot or Gemini rather than Claude or ChatGPT, which she reads as a sign of how early this still is.

13:41 AI Sent 117% More Traffic, And ChatGPT Ads Just Passed $1B

Similarweb published referral data showing traffic sent from AI assistants to websites up 117% year over year, at 770 million referral visits a month. That runs directly against the thing every creator has been told for a year, which is that AI answers swallow the click. Chris does not celebrate the number, he interrogates it: a site with 2,700 visits where most are machines is not the same business as a site with 2,000 human ones, and the question nobody asked in 2012 either is how much of the traffic was ever qualified. His conclusion is a reframe of what a website is for. You are no longer optimising for the landing. You are making sure the site sends the right signal to the AI, so the AI conveys the right thing to the person, who then books, buys or calls. Olga names the conflict out loud: Similarweb sells traffic analytics, so it has a commercial interest in AI referral being a growing category worth measuring. The other half, from August 31: ChatGPT Ads reached a $1B annualised revenue run rate and opened its self-serve ads manager across India, Europe, the Middle East and North Africa, seven months after a US-only pilot. Ads run on the free and Go tiers only, so paying subscribers never see one, which is why neither host had. Chris’s warning on the way out: nobody should buy ads on a channel before they understand its audience, and being early is the only structural advantage a small advertiser ever gets.

21:11 ElevenLabs Made $1M In Pipeline And Gave The Template Away

A callback to the founder who went on paternity leave and let a voice agent work his pipeline, and now the company behind it has published its own numbers. ElevenLabs says the voice agent qualifying its own inbound leads generated more than $1M in pipeline last month. The part that matters more than the number: the template behind that agent is public and free to copy — inbound lead qualifier, sales representative, phone taker, restaurant host, home services qualifier and more. You clone one, point it at your own offer, and it takes the first call. Olga walks the library live and reads what is actually inside: a very well-designed prompt setting personality and environment, the model it runs on, and that is largely it. The free plan gives you 15 minutes of agent call time a month, no card. Her honest caveat is that the $1M is ElevenLabs’ own unaudited number about its own product, and she went looking for a case study of this being used on real cold prospects and could not find one.

23:51 Why Isn’t An AI Answering Every Small Business Phone Yet?

Chris asks the question this segment was really about, using his daughter’s gymnastics place: he can never get anything done by phone, nobody is at the desk most of the time, and the alternative is a phone tree and a voicemail he will not leave because he needs an answer now. There are thousands of businesses shaped like that. Why has AI not swept through those roles like wildfire, when the setup is a prompt and the cost is small? Olga’s half of the answer is the legal boundary. The Telephone Consumer Protection Act, via an FCC ruling in February 2024, treats an AI-generated voice as an artificial voice, so a two-way AI agent cannot open a conversation with somebody who has not consented. Inbound is fine — when you call a company now, the thing answering may not be a person. Cold is not. Chris grants the trade-off is real for legitimate businesses but notes it also raises the cost of spam enough to break the economics. Olga is unambiguously glad: she is already drowning in texts and does not want a robot cold-calling her too.

27:53 $550 Is Where Salesforce Stops Charging Extra For AI

Salesforce collapsed its editions into three tiers: Core at $195 per user per month, Advanced at $395, Max at $550. Agents, Slack and analytics that used to be paid add-ons are bundled in. Olga’s objection is that bundling previously-priced skills into a $550 plan is not the same as making AI free, and that is how it is being described. Her sharper point is about readiness rather than price: most teams still do not know what to do with the CRM they have, let alone skills and agents, so this only works with serious training attached. The second half is only reported, not confirmed: OpenAI is said to be piloting outcome-based pricing with a small number of large enterprise customers, billing for completed work instead of seats. OpenAI has not confirmed it. Salesforce, separately, has charged about $2 per successfully resolved support case on its help agent since June. So one has shipped it and one is rumoured to be testing it, and those are not the same thing. Chris and Olga land on the obvious problem with paying for outcomes: somebody has to define what counts as one, and that definition is where the money is.

29:52 He Faked An NFL Career And AI Vouched For Him

Federal prosecutors charged Daejon Love over a romance and investment scheme that took about $1.3 million from 26 women across four states. He had built social media profiles presenting himself as a San Francisco 49ers player, including convincing photos. Some of the women checked him out with AI, and the AI confirmed he was a player. The FBI affidavit notes that search engines and AI “occasionally stated that Love was a bonafide 49ers player.” Chris is precise about what happened, and it is the part most coverage got wrong: he was not running a campaign to manipulate AI. He faked a profile, and the AI repeated it, giving a fake an authority it never earned. What the AI never did was the obvious detective step of checking the published 49ers roster and noticing his name was not on it. It took the input at face value. Then Chris turns it into the harder question: if an organisation papered the internet with false claims about you, and every AI started repeating them, who exactly do you call? Olga’s answer is that nobody knows yet, and Chris expects a landmark defamation case to settle it. Five days later, the other side of the same problem: Alexa can now tell you whether a message claiming to be from Amazon matches Amazon’s own records. Checking against a record beats judging a vibe — though it only checks Amazon’s own messages, and treating it as a general scam detector would be the same overconfidence again.

33:21 Opening One Folder Can Hand Over Your Keys

The security beat that applies directly to this audience, because it targets the tools this show tells people to learn on. Research published by Manifold shows AI coding assistants automatically run git commands to read a project’s context before asking you to trust anything, without sanitising the repository’s own configuration file first. A repository can carry settings that execute a program on your machine. So opening a shared folder, a client’s zip file or a USB stick can run code as you and take your credentials. Nothing has to be clicked. Named in the research: Claude Code, Goose, Hermes, Qwen Code and Grok Build, with Codex and Cursor affected and already patched. The instruction is simply to update before you open anything from outside. Chris then gives the discipline that matters more than any single patch, on API keys: mint it once, let the tool read it from a file rather than pasting it into a chat where it becomes a record, and delete the evidence when you are done — a key left sitting in a folder your chat has referenced is still reachable. Same forty-eight hours, two more: CISA confirmed active exploitation of an authentication bypass in LiteLLM, which only matters if you run your own AI gateway, and Palo Alto’s Unit 42 published an account of a human attacker using AI agents to break into a company in under ten hours. Olga’s read on the week overall: the news has stopped being about models and started being about agents, marketing with agents, and security — while most people have no idea any of it is happening.

37:52 Everyone Says They’re AI Native. Here’s How You’d Actually Measure It.

Olga sets up the problem from inside her own business. She runs a recruiting company for sales roles, and about a year ago started building a company that upskills salespeople. The first thing you need to upskill anyone is their current level — and for AI, no such scale exists. Chris confirms he is not aware of a standard one either. Meanwhile clients keep asking her for AI-native salespeople, and when she actually interviews them, very few are. They say they are. They are not, once you start testing.

So she built the scale out of two existing frameworks, which is the part worth stealing whether or not you use her levels. The first is the Common European Framework of Reference for Languages, the international standard used to grade a second language, because AI is used like one: you learn the alphabet, then words, then rules, then you chain things together, and eventually you stop consciously assembling and start thinking in it. Her analogy is learning to drive, where the mirrors stop being a checklist. The second is the Dreyfus model of skill acquisition, published in 1980 by Stuart and Hubert Dreyfus at UC Berkeley, which studied how people get good at anything — flying, chess — and found the same progression. If both hold across domains, they hold for AI. Chris adds the open question neither framework answers yet: there is currently no way to assess fluency in one domain and carry it into another. Somebody fluent in AI for sales who moves to marketing — are they starting from zero? Nobody knows.

44:37 The Seven Levels, And The One Where Most People Quietly Stall

Seven levels in three bands. 0, A1 and A2 are following rules. B1 and B2 are seeing patterns. C1 and C2 are intuition, where you are the musician rather than the sheet music.

0 Outsider — “I don’t really use it for work.” Almost never ignorance. Usually nobody ever showed you a use that was obviously yours, so it stayed a novelty you did not have ten minutes for. Olga says she is surprised how many people are still here. The one move: take the task you dread most and have AI do it badly this week. Done badly still beats not started.

A1 Explorer — “I’ve tried it, I use it when I remember.” A couple of good moments and no habit. Every session starts from scratch and you edit almost everything it gives you. The tell in an interview: asked whether they use AI, they name a tool rather than a task. The one move: write one prompt down in a file. Not in your head, not in the chat window. One you will use again.

A2 Adopter — “I use it for specific things every week.” Research, drafts, summaries, reliably. But every job is a separate errand and the output of one never becomes the input to the next. This is where most people are, and it is a comfortable place to stall, because from the outside daily use looks like fluency. The tell: asked for their AI workflow, they give a list of tools and a list of tasks. The one move: connect two, so it becomes a sequence instead of two errands. Chris adds the thing that keeps people stuck here, and it is not skill. It is a scared operating state where people will not make a move that might create a problem, so they make none, and they wait for somebody to tell them the next step. His answer is that no mentor exists for this. You are Lewis and Clark, the AI is your navigator, and you ask it where to go next.

53:27 The First Level You Cannot Fake, And What Turns A Prompt Into A Skill

B1 Practitioner — “AI is how I work, I have a system.” The big jump, and the first level you cannot fake. You have prompts you wrote and kept and reuse without thinking, and a routine: research before the call, notes after it, follow-up inside the hour. You are visibly faster and more consistent than the people around you doing the same job. What you do not have is proof. You believe it is working and you would struggle to say by how much. The one move: pick one number and write it down before and after. Twenty minutes of research down to five is a real number.

B2 Optimizer — “I test what works and I know my numbers.” You stopped improving by instinct. Version one against version two, keep the winner, point at something real that moved. You are starting to know what will work before you test it, and you still test it. This is the rarest rung on the ladder, and almost every other scale stops one level below it. What keeps you stuck: everything you built lives in your head and dies when you go on holiday. The one move is to write down how it works in enough detail that somebody else could run it without asking you a question. This is where Olga’s prompts become skills — once she repeats something enough, it stops needing her attention and runs itself. The show went from taking hours and hours to a couple of hours that way.

Chris stops here to define the ladder underneath the ladder, which is the most portable thing in the segment. A prompt is for a single task. A workflow is a series of tasks in sequence, taking an input to an output, with multiple prompts along the way. A skill is the thing that defines that whole process. A single task is not that useful. A workflow is useful. Defining the workflow is what building a skill means.

C1 Builder — “other people use what I built.” Your system got good enough that it left your desk. People use your prompts and ask you to set theirs up, not because you are leading anything, but because they can see it working. The difference from the level below is a communication act, not a technical one. The one move: put one workflow on a timer. Take the thing you do every Monday and let it run Sunday night. C2 Architect — “AI is my operating system.” Work happens while you are not there. Things run, report back and wait for your judgement rather than your labour. You have not automated yourself out of a job, you have moved to the 20% that actually needed you. The one move: teach it, because the bottleneck is no longer your capability.

58:53 What People Said About Themselves Against What They Scored

Olga puts up her own data and it is the most uncomfortable chart in the episode: green is where people actually assessed, purple is what they reported about themselves. The mass sits in Outsider, Explorer, Adopter and Practitioner, and almost nobody is beyond that — while the self-reported column refuses to enter the bottom half at all. She has stopped running the self-assessment quiz until she can make it match a live assessment, because asking questions lets you clarify and a form does not. The outside version of the same finding: in The AI Proficiency Report (Section, June 2026), 5,026 US knowledge workers were tested hands-on and 54% rated themselves proficient while 10% were. Her reading is generous and it is the right one. It is not dishonesty. If there is no scale, there is nothing to rate yourself against, and you cannot call yourself proficient before you have met somebody genuinely proficient. Chris adds the rule from psychology: all self-reporting fails in one direction, and it fails up, no exceptions. Then Olga adds the tell she has noticed herself, which is the most useful line for anyone hiring — beginners want to impress you with how many tools they use, and advanced people have one model and get obsessed with connecting everything to it. A long tool list is not a strength signal.

1:02:27 Six Questions That Expose Real AI Fluency

The part you can steal and use today, on a candidate, on a team, or on yourself. You stop at the first no.

1. What is the prompt you are reusing? Not the prompt you are using — the one you reuse. It throws people immediately, because somebody without a system has nothing kept, nowhere to keep it, and starts over every time. That single answer tells you whether you are talking to a beginner, and if there is no reused prompt there is no point walking further down the ladder. 2. How did you come up with it? 3. How do you know it’s working, and how do you measure that? 4. Who else is using your prompt? 5. What do you have AI run without you? This is where C1 and C2 reveal themselves, and where genuinely AI-native people get excited and start describing systems. 6. What’s in your AI stack? The closer, which separates somebody who is AI native from somebody presenting as one.

Chris’s takeaways from the segment: you cannot rely on self-reporting at all, so if you are serious you need an assessment — and if you hire a firm to find AI talent for you, ask them whether they have one. Second, businesses should be looking for cross-team repeatability: can somebody new arrive, start a process, and have it run in the absence of the person who built it? That is the test of whether an AI initiative is real. Until the tasks are connected, you have nothing.

1:07:26 Chris Thinks He Just Cloned Himself

Chris pays off his cold-open tease. In the PageMotor beta there are dozens of people, and about ten who have become genuinely fluent operators — their AIs talk to each other and send him reports. For fifteen years he has been able to look at a codebase and see broken patterns, two structures that should match but do not, the outlier that should have been caught. He always wanted others to see that way so the refinements could happen at scale, and gave up on it as unrealistic. What changed is that PageMotor ships heavy documentation of its own rules and conventions, the AIs read it over MCP, and it turns out that is teaching at scale. Now the error reports coming back read like he wrote them. Two systems with identical input and output structure, a third one where a single item is named differently, flagged as an outlier — exactly the thing he has noticed for fifteen years and nobody else cared about.

His conclusion is the one worth keeping: cloning is not going to arrive as a humanoid in a meat suit. It arrives as a thinking style, propagated through documentation and skills across a lot of users. This is from ten people. He thinks that at roughly ten thousand the software landscape starts to change, on the argument that destabilising about 10% of a large system threatens the whole thing — a hive mind that knows how things are supposed to be, pointed at all the existing software.

1:13:01 $4.87B Raised And The Four Biggest Cheques Left America

AI raised $4.87B across 75 companies in the week of August 27 to September 2, taking 39% of all venture dollars. The four largest AI rounds all went outside the United States — Hong Kong, Singapore, Amsterdam, Hangzhou — and the first American company appears at number five. America still funded 40 AI companies, more than every other region combined, and took barely a third of the dollars. 43% of every AI dollar went into robotics, and the two biggest rounds of the week are both robotics companies. Second week running. The top five: Cornerstone Robotics, $700M, Hong Kong, surgical robots, with a stated goal of making robotic surgery affordable in ordinary Chinese hospitals rather than elite ones. Sharpa, $668.8M, Singapore, which builds the hands, joints and actuators other robot makers need. Wonderful, $550M, Amsterdam, enterprise software for building and running AI agents, valuation more than doubled in under six months, with Salesforce coming in as a new investor. VAST, $446.4M, Hangzhou, generative AI for 3D models. Upwind Security, $300M, San Francisco, cloud security that watches what is actually running rather than scanning configuration after the fact — the only round in the top five that neither the company nor its lead investor has announced. Trends: America funded the most companies and the fewest big ones, robots dominated, Israel’s entire AI week was security, and health AI was tiny in dollars while owning the single largest round. Also on air, Chris on the tracker: this is roughly the twelfth straight quiet week, and both hosts prefer the stabilisation to the outlier weeks. The full Week 40 report is here.

1:18:40 Find Your Level Honestly. Then It Becomes A Game.

Chris’s close is about workflows. Fulfilling a single task with AI is some leverage but it is not systematic leverage. A workflow is a series of tasks in order that takes input data and turns it into output, and it requires everything along the way to be documented. Task one, here are the instructions for task one, here is task two. That is what most businesses have never done — the employee responsible knows what they are supposed to do, and it has never been written down as steps. For machines it has to be.

Olga’s close is the assignment. She has laid out what AI-native looks like from A to Z, and the only thing left is to place yourself honestly, because once you are honest with yourself it becomes a game and you can see the next level. Chris’s practical version: take the transcript of this episode, upload it to your AI, and ask it where you are on the scale and how you get to the next rung. You will see the progression the way you can see the belts in karate. And the part he cares about most — once you have your own path, you can stop listening to people selling you a better one.

Resources And Sources

• The AI Fluency Scale. All seven levels with what each one sounds like and the one move out of it.
• OpenAI on GPT-6 Astra. The launch, September 3.
• OpenAI on rating the model Critical for cyber capability. Published September 1, two days before the launch.
• The New York Fed on AI and employment. The 4%, the 13% and the 15%.
• Gartner, on the 22% who have scaled AI. A survey of 1,303 leaders at companies above $50M in revenue, published the same week.
• Uber’s own announcement of the cuts. The company’s stated reason, in its own words.
• California’s No Robo Bosses Act. Approved and sent to the governor, not law.
• Similarweb on AI referral traffic. Up 117%, 770 million visits a month.
• OpenAI on ChatGPT Ads. The $1B run rate and the self-serve rollout.
• ElevenLabs on its own AI SDR. The $1M pipeline figure, unaudited and self-reported.
• The ElevenLabs agent template library. Free to clone, including the inbound lead qualifier.
• ElevenLabs agent pricing. 15 free minutes a month, no card.
• Salesforce on the new three-tier pricing. Core $195, Advanced $395, Max $550.
• The federal complaint against Daejon Love. Including the line about what AI said.
• Amazon on message verification in Alexa. Checks Amazon’s own records only.
• Manifold’s research on hijacking AI coding agents. The one where opening a folder is the attack.
• CISA on the LiteLLM authentication bypass. Actively exploited; affects self-hosted gateways.
• Unit 42 on an AI-assisted intrusion. A vendor account of its own engagement.
• The Common European Framework of Reference for Languages. One of the two parents of the scale.
• AI Funding Report, Week 40. The full breakdown behind the closing segment.
• Weekly AI Funding Tracker. 40 weeks, about $447.5B.
• Episode 54. Nobody bought the product, and the offer was what was broken.
• PageMotor. The AI-native CMS behind the cloning segment.
• The email list. PageMotor and Practical AI updates.