Fat Wrappers: The Unbundling of ChatGPT
Could consumer vertical AI wrappers unbundle the horizontal AI winner?
eBay owned everything, then it got unbundled: Uber took rides. Airbnb took stays. Etsy took craft. Could the same thing happen to ChatGPT? Could consumer vertical AI wrappers unbundle the generalist horizontal AI winner?
Well, I know what you’re thinking. Wrappers are easy to build, each ChatGPT release kills hundreds of them. But all wrappers aren’t created equal. There’s 2 types of wrappers: Thin and Fat. Thin Wrappers are like burritos: simple AI wrapped in a thin tortilla-like UI, most often used as a tool, which Sam Altman promised to steamroll. Fat Wrappers are like burgers: AI in the middle, with human buns on both sides to perform an hybrid end-to-end service for the user.
Thin Wrappers won’t unbundle ChatGPT, but Fat Wrappers could. Here’s why:
Thin Wrappers: AI tools
When LLMs happened, builders shipped the obvious thing: thin wrappers. Clever prompts wrapped in apps, maybe some workflow and a paywall. That works for a moment. But when code becomes as easy to produce as content, it follows content rules: trends emerge, everyone copies, trends die. When your product is mostly a prompt and a feed, copying is trivial. Everyone converges on the same tricks. Competition flattens. Most hit a $30M ceiling and spend their days A/B-testing paywalls and doing great collabs e.g. the Thin Wrapper GOATs Cal AI. And that’s awesome.
But consumers now behave like kings: They DEMAND (agentic) services, not (AI) tools. ChatGPT gave them personalized content on demand. Now they want the next level: personalized action. “Do it for me, my way, now.” Where users paid $10/month for Thin Wrapper tools, they’ll pay $100/month for Fat Wrapper services. At that price point, hybrid human+agent services can work today with real margins, where pre-AI tech-enabled services couldn’t.
10x experiences are now about magical outcomes, and a few (portfolio) companies have started to deliver: Taxes filed by Deduction, trips booked by Miso, skin checked by Thea. Welcome to the age of the Fat Wrappers, where most of the value will accrue.
Fat Wrappers: Agentic services
Not so fast, though. That’s a nice theory, but in practice things are way messier. The issue is that AI is middle-to-middle, not end-to-end. It’s great at the workflow center - drafting, planning, reconciling, synthesizing. It’s clumsy at the operational edges - messy goals going in, real-world execution coming out. As AI speeds up the middle, excellence migrates to the new bottlenecks: prompting and verifying.
Luckily, Fat wrappers are built for that reality. They’re like an AI sandwich: humans handle the edges, agents work the middle. The top bun handles intake, the middle does the heavy lifting, the bottom bun handles verification and real-world action. This is the only way to compete with ChatGPT: Go where OpenAI doesn’t want to go: the complex human-AI hybridization that actually gets things done.
But fully autonomous agentic is hard. Worse, today it’s impossible.
The hard things about agentic things
Indeed, full agentic autonomy isn’t here yet. Think self-driving: the last 1% takes longer than the first 99%. Most products today are level 2 or 3 - agents under human oversight. Level 4 (agents handle most tasks, humans manage exceptions) might happen in 18 months. Level 5 (true independence) could take 5 to 7 years.
But builders aren’t waiting. They’re designing the handoff. Starting 80% human, 20% AI, then pushing boundaries inward until they hit 20% human, 80% AI. They treat reliability as a product feature: auto-approve above X confidence, route to human below. Turn human judgment into policy, then tools, then agent skills. Every solved edge case becomes data and IP. The key is building systems designed to evolve from day one and let the bun get thinner as reliability rises.
Take Deduction: their CPA dashboard lets humans review and edit AI-generated tax responses, building a proprietary dataset with every correction. Thea will route prescription decisions to dermatologists while auto-approving routine consultations. Miso handles simple cash bookings automatically but sends complex award redemptions to ops teams. They’re systematically retiring edge cases.
You know what that sounds a lot like? Ops.
Ops teams will win the Fat Wrapper era
You thought AI killed Ops - well, they’re back. Software is now a world where you need to hold 2 truths in mind simultaneously:
Product is key, but not enough. In the fat wrapper era, great product means great ops. While everyone chases the latest AI models, the companies with the best operations will take dominant positions. Great ops enable magical experiences today, not 5 years from now.
Distribution is key, but not enough. Even distribution becomes ops-heavy. The best teams now run UGC campaigns at massive scale, testing thousands of creatives per week, building systems to make sense of the operational chaos. You wouldn’t believe the intensity that a company like Thea applied to distribution/content Ops to take over TikTok.
As software commoditizes, nailing both product and distribution via operational excellence is (one of) the only way to build defensibility. The best teams will excel at both. Let’s not forget that one of AI’s biggest exit is an ops company: Scale AI.
The Fat Wrapper playbook
Bundling and unbundling hasn’t gone anywhere. ChatGPT rebundled the web; the unbundling will come from vertical services that own the edges. Thin wrappers flattened the market and taught us about attention. Fat wrappers will build trust by finishing the job.
If you’re building: pick a consumer job that’s high friction, high consequence, and recurring. Define “done” so it’s verifiable. Make the buns excellent - intake that never drops context; verification that never waves risk through. Price the outcome. Miso charges €200-300/year for trip booking outcomes, not search queries. Thea started as a thin wrapper, reached $1M in collected revenue, and is now building agentic features. Deduction targets $1M ARR at 2,000 clients. When you own the full job completion, you can price like a service, not a tool.
Measure agentic coverage, cycle time, defect rates, human minutes per outcome. Celebrate every edge case you retire. Start with a thick bun. Make it thinner on purpose. Keep the service invisible. March toward level 4, keep level 5 in sight, and let the system upgrade itself as the frontier moves.
In the end the metric that matters hasn’t changed: Did the thing get done, my way, on time, without me?
And of course, if you’re building a Fat Wrapper, you know where to find me.
Love this!
Great read