Gamma passed $100M ARR with a team of around 50 people. Profitable, on just $23M raised.
Then this year they did something most founders would never dare: they tore apart a product 70 million people use and rebuilt it as an agent.
One finished presentation now runs through roughly 20 AI models and 200 calls before you ever see it.
Jon Noronha, Gamma's co-founder and chief product officer, sat down with Greg Isenberg to break down the layer that makes this work: the agent use. He also explained why uses, not models, are where the next wave of AI startups gets built.
Here's the full episode. My notes below.
What is an agent use?
Jon's definition is the cleanest I've heard.
An agent use is the set of tools an agent uses to act in its environment, inspect its own work, and ask you what you actually want.
Three jobs:
- → Act: browse, write files, call APIs, generate images
- → Check: look at the output and catch its own mistakes
- → Ask: pull your preferences instead of guessing
Claude Code is the best-known example. The model underneath is the same one everyone can rent. The use around it (the terminal access, the file edits, the self-testing loop) is what makes it feel like a coworker instead of a chatbot.
Greg's framing on the episode: uses are the new wrappers.
Remember when "GPT wrapper" was an insult? Jon's response is that the wrapping turned out to be most of the work. Even OpenAI and Anthropic pour serious effort into use development for their own products.
The model is the engine. The use is the car. Nobody commutes on a bare engine.
Inside one Gamma deck: 20 models, 200 calls
This is the part of the episode worth replaying twice.
You paste messy notes into Gamma and ask for 20 slides. Here's what Jon says happens behind the scenes:
- Gamma spins up a cloud sandbox for your deck
- It asks you a few questions through a quick form
- It pulls context from your Notion, Slack, or Linear
- It writes an outline, then picks a template using vision models and embedding search
- It generates text, layouts, images, and charts in parallel, in about a minute
- A vision model then inspects every single slide for misaligned text, cropped heads, and weak contrast
- It runs rounds of fixes until the deck passes
Total damage: about 20 different models and 200 calls for one presentation.
That last QA step is the detail most builders miss. The agent looks at its own work the way a picky designer would, then fixes what smells off before you see it.
Gamma started in 2020, nearly stalled out, then broke out in 2023 when they bet the company on AI. This year they went further and rebuilt the core product around the agent instead of bolting one onto the side.
Where vertical uses win
So where does that leave you, the founder without 70 million users?
Jon's filter: look for subjective work. Problems where people disagree on what good looks like, care deeply about the specifics of the output, and want a human in the loop.
Coding agents are crowded because code has a built-in use: tests either pass or they don't. Subjective work (design, video, physical spaces, games) is wide open because the use has to encode taste. That's hard. Hard is good for you.
Areas Jon calls active right now: physical services, legal tech, medical tech. He also loves personal video games built for non-technical players. Greg pushed back on legal tech (too crowded for his taste), which tells you even the obvious lanes still get debated by people at the top of this game.
Jon names three places your edge can come from:
- → A distinct vision you move toward faster than anyone else
- → Proprietary data nobody can scrape
- → A business that mixes AI with hardware or operational services
Notice what's not on the list: a better model. You're not going to out-train Anthropic. You can absolutely out-use a lazy incumbent in one niche.
I wrote about how fast this shift is moving in the 9 stages of AI agent adoption. Uses are the layer that pulls a business up those stages.
Taste beats evals
My favorite stretch of the episode is Jon on "AI smell."
You know AI smell. The output is technically correct and still feels generated. The gradient background. The bullet points that say nothing. In creative work, that smell kills the product even when the evals score fine.
Jon's position: in subjective domains, taste carries more weight than evals. Gamma still measures, but a founder with strong opinions about what good looks like beats a dashboard of benchmark scores.
Some specifics from how they operate:
- When a new model drops, Gamma tests it within the hour
- About 98% of Gamma's value comes from prompts; fine-tuning is just a cost and quality optimization at the margins
- They run a broad but shallow stack: many models, swapped often, none load-bearing enough to hurt when it changes
And then there's the experiment that surprised me most.
Gamma tested a fast mode that stripped steps on the way to the finished deck. It lost. The version that made you answer a short form first converted better to paid.
Good friction beat speed. People trust the result more when the agent asked first. (How many of us are busy deleting every click from onboarding right now?)
Jon's team treats experiments as a numbers game: volume of tests, not elegance of any single one.
The uses Jon would build next
Greg asked Jon what he'd build if he weren't building Gamma. The answers are basically a free idea list.
A Slack agent that learns your voice. Not canned replies. A use that reads how you write, drafts responses the way you would, and gets corrected until it nails it. Jon wants this for himself, today.
A couch shopping use. Render the couch in your actual room, measure the space in pixels so it knows a 58-inch fits where a 68-inch won't, then suggest a better option. Durable buyers already exist: interior designers and pros at companies like Restoration Hardware.
Anything built on AI image models. Jon calls image models the most underhyped tool in the stack. Most people stop at logos and menu mockups. Gamma runs about 10 different image models in production for diagrams, textures, and icons.
His closing advice for picking: choose work you could dogfood eight hours a day. Greg's version: marry the niche, date the product.
What this means if you're bootstrapping
You don't need $23M or a research team to play here.
A use is product work, not model work. Context gathering, self-checking loops, the right questions at the right moment. That's founder-brain stuff, and it's exactly the kind of thing a solo builder can ship in a niche an incumbent ignores.
The other half of Gamma's story is distribution, and they're just as deliberate there. I broke down how Lovable and Gamma run B2B influencer marketing if you want the growth side of the same company.
Pick a subjective niche and build the use. Let the model labs fight over the engine.
FAQ
What is an agent use?
An agent use is the set of tools wrapped around an AI model that lets it act in an environment, inspect its own output, and ask the user for preferences. Jon Noronha, Gamma's CPO, points to Claude Code as the best-known example: same models everyone has, with a use that makes them useful for real work.
What's the difference between an agent use and a wrapper?
A wrapper passes your input to a model and returns the answer. A use gives the model tools, context, and a self-correction loop. Gamma's use runs about 20 models and 200 calls per presentation, including a vision model that QAs every slide. That loop is the difference between a demo and a product.
Do I need to fine-tune my own model to build an AI product?
Probably not. Jon says about 98% of Gamma's value comes from prompts, and they treat fine-tuning as a cost and quality optimization, not a moat. They also swap models constantly, testing new releases within the hour. Your edge is the use and your taste, not a custom model.
How do I pick a vertical AI agent idea?
Jon's filter: subjective work where people disagree on what good looks like and care about the details of the output. Then check your edge: a distinct vision, proprietary data, or a mix of AI and the physical world. Finally, pick something you could dogfood eight hours a day.
Steal playbooks like this every week
This is what I do: find founders with real numbers and pull their playbooks apart so you can run them yourself.
On the Profitable Founder Podcast I interview bootstrapped founders making $100K to $10M a year and get them to show the actual math behind what's working.
