San Francisco is where many of the AI labs, the best-known accelerators and a large share of AI startups sit within a few miles of each other. That makes it a useful place to watch how founders actually work with AI, as opposed to how it is described in launch posts. The public record from Y Combinator, its partners and the labs themselves gives a clearer picture than the hype does.
The short version: AI changed how products are built first, and it is now changing how companies are run. The founders who benefit most treat AI as staff with a job description, not as a clever search box. This article walks through what the numbers say, the jobs founders hand to AI, where it breaks, and a one-week way to start. There is also a short film on this topic in the GeniOS video library.
Key takeaways
- AI is now the default way many young startups write code, and small teams are reaching revenue that once took far larger ones.
- The next wave is operations: drafting, research, support, follow-ups and investor reporting.
- The limit is rarely the model. It is context and follow-through: what was promised, to whom, by when.
- Founders who do this well give each AI job an owner, limits and a check that it actually happened.
What do founders actually use AI for?
What founders use AI for falls into two groups. The first is building: writing code, generating tests, prototyping interfaces and debugging. The second is running the company: drafting emails and documents, researching customers and markets, answering support, preparing for meetings, following up, and writing investor updates.
The first group came first and is now routine for many new technical startups. The second is where most founders are still experimenting, and where the gap between a demo and a working habit is widest. Building with AI produces something you can test: the code runs or it does not. Running a company with AI is harder to check, because the output is a decision, a promise or a follow-up that has to actually happen in the world.
Stage changes the mix. A two-person team before product-market fit mostly uses AI to build and to research customers. Once there are paying customers and a team of ten or more, the weight moves to the operating jobs: support, follow-ups, hiring loops and reporting. That is also the point where the founder's own memory stops being enough to hold the company together.
That difference explains most of what follows. AI is excellent at producing a draft. A company runs on what happens after the draft.
What do the numbers show about founders and AI?
The clearest numbers come from Y Combinator's own batches.
- In March 2025, YC managing partner Jared Friedman said that "A quarter of the W25 startup batch have 95% of their codebases generated by AI," as TechCrunch reported. YC's Garry Tan added: "This isn't a fad. This isn't going away. This is the dominant way to code."
- A week later, Tan told CNBC that the Winter 2025 group was growing faster than past cohorts: "the whole batch is growing 10% week on week," he said. "That's never happened before in early-stage venture." CNBC reported that about 80% of the companies presenting at that week's demo day were AI-focused.
- On team size, Tan said "you don't need a team of 50 or 100 engineers," and on funding, "You don't have to raise as much. The capital goes much longer."
The labs report similar patterns inside their own walls. In a December 2025 study, Anthropic surveyed 132 of its engineers and researchers. They self-reported using Claude in 59% of their work, with a self-reported productivity gain of about 50%, and said "27% of Claude-assisted work consists of tasks that wouldn't have been done otherwise." The authors caution that productivity is hard to measure, which is fair. The 27% figure is the interesting one for founders: AI does not only speed up existing work, it makes work possible that a small team would have skipped.
Across businesses more broadly, Anthropic's Economic Index report from September 2025 found that "77% of API transcripts show automation patterns (especially full task delegation)," compared with "just 12% for augmentation." Companies building on the models are handing over whole tasks.
Six jobs founders hand to AI
Six jobs cover most of what founders hand to AI. The pattern in each is the same: AI does the first draft and the legwork, and the founder keeps the judgement.
| Job | What AI does | What stays with the founder |
|---|---|---|
| Building product | Writes most of the code, tests and prototypes | What to build, and whether it holds up at scale |
| Research | Summarises markets, competitors, prospects and calls | Which questions matter, and what to believe |
| Writing | Drafts emails, proposals, posts and documents | Tone, promises made, and what gets sent |
| Customer support | Answers routine questions and drafts harder replies | Exceptions, refunds and angry customers |
| Follow-up | Tracks who was promised what and drafts the nudge | Priorities, and relationships that need a call |
| Reporting | Assembles investor updates, board decks and weekly numbers | The story, the asks and the honest bad news |
The top three rows are where most founders already are. The bottom three are where the gains now sit, because they are the jobs that eat a founder's evenings. The investor update in an hour playbook shows the reporting row in practice, and AI for founders covers the wider set of uses.
Garry Tan's own question about AI-written code applies to every row: "Does it fall over or not?" A fast draft that breaks at scale, or a follow-up that never gets sent, is not a gain.
Where does AI break when founders use it to run the company?
It rarely breaks because the model is not smart enough. It breaks because the model does not know what the founder knows: which customer is upset, what was promised on a call, which deadline moved, which investor asked for a number last week. Anthropic's economic research says this plainly. Its September 2025 report found that "deploying AI for complex tasks might be constrained more by access to information" than by model capabilities, and that fixing this "may require firms to restructure their organization, invest in new data infrastructure, and centralize information for effective model deployment."
The second break is follow-through. A founder can generate ten follow-up drafts in a minute. Whether those follow-ups were sent, answered and acted on is a different problem, and chat tools do not track it. The Anthropic staff study also flags a human cost worth noting for a growing team: some employees reported "fewer mentorship and collaboration opportunities as a result." For why this context gap matters across the whole company, see why a company brain is not enough.
Common mistakes founders make with AI
- Ten tools, no system. A chat assistant here, a meeting note-taker there, a coding agent elsewhere. Each holds a slice of context, and none knows what the company is trying to do this quarter.
- Pasting context by hand. If every request starts with the founder explaining the situation again, the founder is still the system, and still the bottleneck.
- Drafts with no owner. AI writes the follow-up, it sits in a tab, and the promise slips. A draft is not a done item.
- Letting agents send without approval. Early on, an AI that emails a customer or an investor on its own creates risk that one bad message can turn into a lost deal.
- Measuring hours saved. Hours saved are hard to count and easy to inflate. Count promises kept and items closed on time.
- Skipping the team. A founder who uses AI well but leaves the team on old habits has sped up one person. The company runs at the speed of its slowest handoff.
How to start: a founder's first week of running the company with AI
- Day 1: list your recurring jobs.
Write down what you do every week that is not product or selling: investor updates, hiring follow-ups, vendor chasing, support escalations. Mark the ones you dread.
- Day 2: put context in one place.
Connect, read-only, the places where commitments are made: email, calendar and chat. The goal is that AI starts from the same facts you do, without you pasting them in.
- Day 3: pick one job and write its rules.
Choose one recurring job, such as follow-ups after calls. Write what good looks like, what the AI may draft, and what needs your approval.
- Days 4 and 5: run it with approval on.
Let AI prepare the work and route it to you or the right owner with a date. Approve every outbound message. Note every correction.
- Weekend: review.
Count what closed on time, what slipped and what you rewrote. Keep the job, adjust the rules, and add a second job next week.
One rule makes the week work: approval stays on for anything that leaves the company. It is tempting to switch it off on day three because the drafts look good. Keep it on until the corrections log is short and boring. Approval is also how the AI learns your standards fastest, because every edit teaches something about tone, price or priority that a written rule would have missed.
Founders who want the whole weekly cadence should read running an AI-native startup without status meetings.
How to measure whether AI is helping you run the company
| Measure | Why it matters |
|---|---|
| Promises kept on time | The plainest test of whether follow-through improved |
| Items that needed you to unblock them | Shows whether you are still the hidden bottleneck |
| Time from a customer request to an owner | How long things wait before someone owns them |
| Drafts approved without edits | Whether the AI is learning your standards |
| Founder hours on status and chasing | Should fall first, and fall steadily |
If the second row is not falling, the AI is speeding you up without freeing you. That pattern is the subject of the founder bottleneck.
How GeniOS approaches founders using AI
Most founders already have tools that draft. What they lack is something that knows what the company is trying to do and checks the work against it. GeniOS reads where a company's work happens, such as email, calendar, chat and the business tools already in use, read-only to start, and forms the company's goals from its own emails, decks and meetings.
It checks every team's pending work against those goals. What reaches the founder, or whoever owns the item, is one situation, what to do, why, one owner and a date. The owner answers Delegate, Done, Later or Not relevant. Delegate hands work to a person or an AI agent, and nothing is sent until a human approves.
An item closes only on proof, such as the email that went out, and GeniOS learns from each outcome. The founder stops being the place where context lives, which is the step toward what we call the autonomous company.
Worked example: a founder's Tuesday at a safety-shoe maker
The names, dates and numbers in this example are made up to explain the idea. Vaultex makes safety footwear in Greater Noida. This week, Bata wants 12,000 pairs of S3 safety shoes at $12 a pair, about $144,000, with a vendor review on Friday. Reebok samples are running on Line 2. Adidas has a vendor audit coming and wants IS 15298 test reports. Payment terms are 60 days, and the steel toe-cap supplier has just raised prices 8%.
Before, the founder used a chat assistant for drafts and kept the rest in their head. On Tuesday they would ask it to "write a quote email to Bata," paste in last month's price, and send it late at night. Nobody would notice that the $12 price was set before the toe-cap rise, or that Harsh had already promised Bata a reply on WhatsApp.
With context in one place and a rule for each job, Tuesday looks different:
- Quote: one item to Harsh in sales, due Wednesday: revise the Bata quote for the 8% toe-cap rise before Friday, with the supplier email attached. Harsh delegates the draft to an agent and approves the final version.
- Capacity: one item to Anushka in production: confirm whether Line 2 can take Bata once the Reebok samples finish, because the quote names a delivery date.
- Audit: one item to the quality lead: the Adidas pack is missing two IS 15298 reports. The agent lists which ones; a person chases the lab.
- Cash: one decision for the founder: whether 60-day terms still work at the new margin.
The founder makes one decision instead of four relays. Each item closes when its proof exists: the sent quote, the slot in the plan, the reports in the pack.
On Friday morning, before the Bata review, the founder opens one page. The revised quote went out on Thursday at the new price. Line 2 has a slot after the Reebok samples. One IS 15298 report is still at the lab, with a named person chasing it and a date. Nobody had to call a status meeting to find any of this out, and nothing went to a buyer without a person approving it.
Conclusion
The founders at the centre of the AI wave use it first to build and then to run their companies. The public numbers from YC show how far the building side has gone. The running side is where the next gains sit, and it depends less on the model than on context and follow-through. Put context in one place, give each AI job an owner and limits, keep approval on what goes out, and count promises kept. To see how this works with your own company's data, see pricing or book a call.
Sources
- TechCrunch, A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated, 6 March 2025
- CNBC, Y Combinator startups are fastest growing, most profitable in fund history because of AI, 15 March 2025
- Anthropic, How AI is transforming work at Anthropic, 2 December 2025
- Anthropic, Anthropic Economic Index report, 15 September 2025