<- all tokdocs
2026.10.26@joe.koenig

Joe Koenig says he exited "right at the very beginning of the AI stage" in late 2024, but the agentic coding tools that make vibe-coding a CRM possible shipped two months after he left, and his zero-to-$10M ARR credential is unverifiable

Business
Watch on TikTok

View on TikTok ->

The artifact is a 76-second vertical clip, 1080x1920, in a QuickTime/MOV-family MP4 container (ffprobe reports format_name=mov,mp4,m4a,3gp,3g2,mj2), carrying an HEVC Main-profile video stream at level 150, yuv420p, 30/1 fps, 2,282 frames, 603,952 bps, alongside an AAC HE-AACv2 audio stream at 44.1 kHz stereo, 1,641 frames, 48,105 bps; overall container bitrate is 658,804 bps and the file is 6,264,134 bytes, matching yt-dlp's format_id of bytevc1_1080p_658801-1 and its SDR dynamic-range tag. Container and video durations both read 76.066667 s, the audio stream reads 76.040998 s, and the extracted audio.mp3 reads 76.094694 s. The post timestamp is epoch 1793038382, which is 2026-10-26T18:13:02Z. The uploader handle is joe.koenig (uploader_id 6963694783513052166) and the display nickname is "Joe | Profit+Purpose"; the audio credit is "original sound" attributed to the same nickname, so there is no licensed music track. The description is 50 characters and identical to the title: "Here's what I would NOT do with AI in my business." Engagement counts were captured at epoch 1793056537, which is 2026-10-26T23:15:37Z, five hours and two minutes after posting, and read 892 views, 25 likes, 6 comments, 1 repost and 5 saves, so these are early-life numbers rather than a settled distribution. I read 17 of the 38 extracted frames: 001, 002, 004, 007, 010, 012, 014, 016, 017, 018, 020, 024, 029, 032, 035, 036 and 038, which at the 2-second sampling interval covers the clip end to end. The transcript is 278 words across 16 segments, language tagged en. Every frame I read is the same unbroken medium close-up: a bald man with light stubble in an olive green crew-neck t-shirt, a DJI lavalier mic clipped at his sternum, a metal-band smartwatch on his left wrist and a gold wedding ring, seated against a charcoal wall hung with guitars (a cherry-red semi-hollow, a sunburst Stratocaster-style solid body, an acoustic) under track lighting and a white coffered ceiling. A black sticker caption reading "I would NOT do this with AI" sits at the top of frames 001 and 002 and is gone by frame 004. Burned-in auto-captions run the whole clip in bold white outlined text: frame 001 "I sold my business at the end of 2024", frame 002 "as a software company and this was kind of like", frame 004 "and I've thought back a couple times", frame 007 "you know if AI capabilities were what they are now", frame 010 "the more I actually would not do that", frame 012 "which is building a lot of tools internally", frame 014 "for example um", frame 016 "and yes", frame 017 "and I think at the end", frame 018 "we were paying Hubspot about 30 grand a year", frame 020 "by building our own but", frame 024 "in order to not just build the initial product", frame 029 "as you're looking at what AI can do for your company uh", frame 032 "to move faster at what you're doing", frame 035 "I'm Joe I grew my company from 0 to 10 million", frame 036 "a R R uh", frame 038 "so follow along for more". No screen recording, no CRM interface, no HubSpot invoice or contract, no code, no terminal, no chart, no company name, no logo, no lower-third and no end card appears in any frame I read, so not one of the figures in the clip has on-screen support. The spoken argument is a single counterfactual: he sold a software company at the end of 2024, he has since wondered what current AI could have done for it, and the one thing he would not do is build a lot of internal tools, with replacing a $30k-a-year HubSpot subscription with a self-built CRM as his worked example, because pulling developers off the product to build and then maintain that CRM costs more than the license. The transcript-versus-audio check is clean: the last segment ends at 75.90 s against an audio duration of 76.09 s, so nothing is timestamped past the end of the recording and no segment is hallucinated.

The credential the whole video rests on is the one claim I could not source

The argument is built on the speaker's standing, stated twice. At 0.00 s he says "I sold my business at the end of 2024 as a software company," and at 69.06 s he closes with "I'm Joe, I grew my company from zero to 10 million ARR, exit at the end of 24." Neither statement names the company, and no frame I read shows a company name, logo or product.

I searched for a software founder named Joe Koenig with a late-2024 exit at roughly $10 million in annual recurring revenue and found nothing that matches. The closest candidate is a different story. Embrace Education, a special-education software company, was sold to Everway, and Everway's own announcement is dated 30 January 2025. That release names Craig Powell as CEO of Everway and Gus Brecht as CEO of Embrace Education, and it describes Embrace as serving over 1,100 districts with a 98% client retention rate and twenty-five years of history. The sell-side advisor's transaction page dates the deal to January 2025 and gives Embrace a 1999 founding date. A Joe Koenig does appear in search results associated with Embrace, but the company's chief executive in the acquisition release is Gus Brecht, and a business founded in 1999 does not fit "I grew my company from zero to 10 million ARR" told by someone who also frames the sale as happening at the dawn of AI. Neither release discloses revenue or ARR. I am not willing to connect the TikTok account to that transaction on this evidence.

That leaves the central credential unverified. The $10 million ARR figure, the zero-to-ten growth and the end-of-2024 exit date are all first-person assertions with no filing, press release, acquirer announcement or trade coverage behind them that I could reach. This matters more than it would for most creator claims, because the entire persuasive force of the video is that a person who actually ran a software company at scale has thought this through. Treat the advice on its merits and the resume as unsupported.

"Right at the very beginning of the AI stage" is wrong about AI and right about coding agents

At 0.00 s through 8.00 s he says his exit "was kind of like right at the very beginning of the AI stage." For general-purpose AI capability that is off by roughly two years. For the specific capability his argument depends on, which is an agent that can build a working CRM from a conversation, the framing is closer to correct, and the correction runs in his favour.

Take the first half. Six months before his stated exit, Anthropic published its Claude 3.5 Sonnet announcement on 21 June 2024, which states that the model, given appropriate tools, "can independently write, edit, and execute code with sophisticated reasoning and troubleshooting capabilities," and reports that in an internal agentic coding evaluation it "solved 64% of problems, outperforming Claude 3 Opus which solved 38%." A model vendor making that claim publicly in mid-2024 is not the very beginning of anything. (I tried to fetch OpenAI's own ChatGPT introduction post at openai.com/index/chatgpt for the November 2022 launch date and received HTTP 403, so I am not citing it.)

Now the second half. The tooling that turns "a model can write code" into "a non-engineer can build a CRM over a weekend" arrived after he left. Anthropic's Claude 3.7 Sonnet announcement, dated 24 February 2025, introduced Claude Code as a "limited research preview" and described it as "an active collaborator that can search and read code, edit files, write and run tests, commit and push code to GitHub, and use command line tools." That is roughly two months after the exit he describes. So the honest version of his opening is that he exited after two years of broadly capable models but two months before the agentic coding products that make his hypothetical realistic. The video compresses both into "the very beginning," which overstates how early 2024 was and understates how recent the relevant shift is.

The behaviour he is arguing against post-dates his exit, and he is candid that he is speculating

At 24.18 s he says "I see people replacing their CRM, for example, you know, just kind of like vibe coding their own CRM." The phrase itself is newer than his exit. Per Wikipedia's article on vibe coding, Andrej Karpathy introduced the term in a post on X in February 2025, defining it as coding where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." The same article records that Merriam-Webster listed it as a slang and trending expression in March 2025 and that Collins English Dictionary named it Word of the Year for 2025. I did not fetch Karpathy's original post, so I cite the encyclopedia entry rather than the tweet.

This is a chronology point, not a gotcha. Everything he says about replacing a CRM with a vibe-coded one refers to a practice that did not have a name, and barely had a toolchain, while he still owned the company. He is honest about this. At 4.04 s he frames the whole piece as a counterfactual: "I've thought back a couple times and really wondered, you know, what would have been possible for our company, you know, if AI capabilities were what they are now." The video is a thought experiment about a decision he never had to make, presented with the confidence of experience. The transcript supports that reading and the frames add nothing to it, because there is no demo, no screen share and no artefact of any kind.

The thing he says he sees companies doing is real and measured, and CRM is squarely in it

His premise that companies are replacing bought software with internally built tools holds up against survey data. Retool's 2026 build-versus-buy report, published 17 February 2026 from 817 respondents (36% software engineers or developers, with the rest spread across operations, product, data, marketing and sales ops, IT, business analysis and finance), reports that 35% of teams have already replaced at least one SaaS tool with a custom build and 78% expect to build more of their own tools in 2026. The category breakdown puts workflow automations at 35%, internal admin tools at 33%, BI tools at 29%, CRM and form builders at 25%, project management at 23% and customer support at 21%.

So the specific example he reaches for, the CRM, sits fourth on a list where a quarter of surveyed teams say they are replacing it. One caveat on the sample: these are Retool's own customers and builders, a population selected for being willing to build internal tools, so the absolute percentages should be read as high. The same report notes maintenance burden at 26% as a barrier to further AI automation, which is his argument stated by the people doing it, but it offers no data on whether teams regret the replacements or abandon them, so the follow-through question he raises stays open.

I attempted to fetch McKinsey's State of AI 2026 survey twice for a second, broader sample and the request timed out both times, so I am not citing it.

The $30,000 HubSpot figure is plausible against published list prices and otherwise unverifiable

At 29.14 s he says "we used HubSpot. And I think at the end, we were paying HubSpot about 30 grand a year." The burned-in caption at frame 018 renders the company name as "Hubspot"; the company styles it HubSpot, and the Whisper transcript gets it right, so no product name in this clip is mangled.

The number itself is a private contract, so it cannot be verified directly. It can be checked for plausibility against HubSpot's published list prices, which I fetched. HubSpot's Marketing Hub pricing page lists Professional at $800 per month billed annually, including three core seats and 2,000 marketing contacts, with additional core seats starting at $45 per month and a one-time $3,000 onboarding fee; Enterprise starts at $3,600 per month including five core seats and 10,000 contacts, with a one-time $7,000 onboarding fee. HubSpot's Sales Hub pricing page lists Professional at $90 per seat per month billed annually ($100 billed monthly) with a one-time $1,500 onboarding fee, and Enterprise at $150 per seat per month with a one-time $3,500 onboarding fee.

Run the arithmetic. Marketing Hub Professional at list is $9,600 a year. Sales Hub Professional is $1,080 per seat per year, so roughly nineteen sales seats on top of Marketing Hub Professional lands near $30,000, as does about twenty-eight Sales Hub Professional seats on their own. For a company at the $10 million ARR he claims, both shapes are ordinary. The figure is consistent with list pricing and is neither remarkable nor suspicious. It remains an unsourced recollection, hedged in the clip itself with "I think."

One thing worth separating out: these are current published prices, fetched today, not the prices he would have paid through 2024. HubSpot's seat-based packaging has changed since, so the arithmetic above establishes that $30,000 a year is a normal HubSpot bill for a company that size, not that it was his bill.

The opportunity-cost argument is the strongest claim in the clip, and the measurement behind it has since been withdrawn

From 37.64 s he makes the case that actually matters: "we could have saved that 30 grand a year by building our own. But I think the opportunity cost would have been absolutely massive, I definitely would have had to pull some of my dev team away from what they were doing, in order to not just build the initial product, but then ultimately maintain it add new features, updates, those kinds of things."

The structural point is sound and does not need a citation: a license is an operating expense that stops when you cancel it, while an internal CRM is a permanent claim on engineering capacity that grows with every integration, migration and compliance requirement. The maintenance half is where the evidence is thinnest in both directions, and I want to be precise about what can be said.

The best-known attempt to measure whether AI assistance actually speeds up experienced engineers is METR's randomised trial, published 10 July 2025. Sixteen experienced open-source developers worked 246 real issues, with tasks randomly assigned to AI-allowed or AI-disallowed, using Cursor Pro with Claude 3.5 and 3.7 Sonnet. Developers took 19% longer with AI tools available, while estimating afterwards that AI had made them about 20% faster, having expected a 24% speedup going in. That result reads as direct support for his position. It should not be used that way any more. The post now carries a prominent banner stating "These results are out of date" and linking to results current as of early 2026, with the authors saying they believe the historical findings no longer reflect the current impact of AI models on open-source developer productivity. So the strongest piece of quantitative evidence against the build-it-yourself case has been retired by the people who produced it.

What survives is the qualitative objection, which is about ownership rather than speed. The vibe coding article quotes Simon Willison: "Vibe coding your way to a production codebase is clearly risky. Most of the work we do as software engineers involves evolving existing systems, where the quality and understandability of the underlying code is crucial." A CRM is exactly that kind of system. It holds customer records, it is load-bearing for revenue, and it accumulates ten years of edge cases. His closing line at 64.02 s, "it's really important to not lose focus on what your company does," is the correct frame, and he reaches it without needing any of the numbers he could not support.

Key Takeaways

  • Unverified: The video's central credential, "I grew my company from zero to 10 million ARR, exit at the end of 24" (69.06 s), has no public source I could find. No company is named in the audio or in any of the 17 frames I read. The nearest late-2024 software exit involving a Joe Koenig is Embrace Education's sale to Everway, announced 30 January 2025, but that release names Gus Brecht as Embrace's CEO and the advisor's page dates the company's founding to 1999, so I do not connect the two.
  • Partial correction: He calls the end of 2024 "right at the very beginning of the AI stage" (4.04 s). Anthropic's Claude 3.5 Sonnet announcement of 21 June 2024, six months before his stated exit, already described the model as able to "independently write, edit, and execute code" and reported 64% on an internal agentic coding evaluation. The framing is wrong for AI generally.
  • Clarification: The same framing is close to right for the capability his argument actually turns on. Claude Code launched as a limited research preview on 24 February 2025, about two months after his stated exit, so the agentic coding tooling that makes "vibe coding your own CRM" a real option did arrive after he left.
  • Context: "Vibe coding" (24.18 s) was coined by Andrej Karpathy in February 2025, listed by Merriam-Webster as slang and trending in March 2025, and named Collins English Dictionary's Word of the Year for 2025. The practice he is warning against entirely post-dates the company he is drawing on.
  • Verified: The premise that companies are replacing purchased software with internal builds holds. Retool's 17 February 2026 report, 817 respondents, found 35% have already replaced at least one SaaS tool and 78% expect to build more in 2026, with CRM and form builders at 25%. The sample is Retool's own customer base and should be read as high.
  • Unverified: "We were paying HubSpot about 30 grand a year" (33.42 s) is a private contract figure, hedged in the clip with "I think." It is consistent with published list prices: Marketing Hub Professional is $800 per month billed annually and Sales Hub Professional is $90 per seat per month billed annually, so roughly $9,600 plus nineteen sales seats, or twenty-eight sales seats alone, reaches $30,000.
  • Correction: Anyone citing METR's 19% slowdown finding to back the "don't build it yourself" case should stop. METR's own post now carries a banner reading "These results are out of date" and points to results current as of early 2026, stating the historical findings no longer reflect current AI impact on developer productivity.
  • Context: No frame I read contains a screen recording, CRM interface, invoice, code, terminal, chart, company name or logo. Every quantitative claim in this video is spoken only.

Resources

Sources I could not reach: OpenAI's ChatGPT introduction post returned HTTP 403, and McKinsey's State of AI 2026 survey page timed out on two attempts. Neither is cited above.

Published October 26, 2026. Writeup generated from a favorited TikTok.