Starting an AI Startup: What Three Sacramento-Area Founders Have Learned Building With AI
AI has made it ridiculously easy to build something.
That may actually be one of the biggest problems facing AI founders.
You can describe an idea to an AI coding tool today and have something resembling a working application before dinner. A few years ago, that might have required a developer, a budget, and several weeks of work.
But a working demo isn't necessarily a company.
That was one of the themes that emerged during our recent AI for Business: Starting Your AI Startup panel at Roseville Venture Lab.
Startup Folsom brought together three founders who aren't simply experimenting with ChatGPT. They're building companies where AI is part of the actual product:
Shweta Gandhi, CEO and co-founder of Strived.io, an AI-powered data analytics company working with school districts.
Carolyn Peer, CEO and co-founder of Humaxa, which uses AI and data processing to help organizations in regulated industries manage compliance and complex workflows.
Mauro Sica, founder of Core Envision, who is applying AI to architecture and construction, including the difficult problem of translating two-dimensional construction drawings into usable three-dimensional information.
The discussion covered everything from large language models and vibe coding to cybersecurity, infrastructure costs, customer procurement and intellectual property.
But I think the most useful lessons were actually about startups.
1. Start With the Problem, Not the AI
One of the questions I asked the panel was pretty simple:
Did you have a problem and discover AI could help solve it, or did you have AI and go looking for a problem?
All three essentially came back to the problem first.
Shweta had already spent years working in education and had seen firsthand how difficult it was for schools to make sense of enormous amounts of academic and non-academic student data. In a previous organization, she had engineers doing much of that data work manually.
AI eventually provided a way to do that work much more efficiently and at scale.
As she explained during the panel, AI helped Strived scale the solution, but AI wasn't the reason she started the company. The problem already existed.
Carolyn's story was similar, although Humaxa's problem changed over time.
Humaxa originally focused on employee retention. The company discovered that while the problem was real, the people responsible for solving it didn't always have enough organizational power or budget to buy the solution.
That's a painful startup lesson: a real problem isn't necessarily a good business opportunity if nobody can—or will—pay to solve it.
Humaxa eventually shifted toward regulatory compliance, where the underlying technology could address a problem organizations were much more motivated to solve. When large language models took off in 2023, they became another layer in the product rather than the reason for the company's existence.
Mauro came at it from architecture and construction. He had spent years dealing with inefficiencies in construction and prefabrication before AI became capable enough to tackle some of the industry's messiest information problems.
That's probably lesson number one for anyone thinking about launching an AI startup:
Don't ask, "What can I build with AI?"
Ask:
"What painful problem do I understand unusually well, and can AI now help me solve it better, faster or cheaper?"
That's a much better place to start.
2. AI Becomes Interesting When It Changes the Economics
Shweta gave one of my favorite examples of the evening.
School districts use many different software vendors, which means Strived needs to ingest data from many different systems.
Traditionally, connecting another data source could require substantial engineering work. Strived now uses AI agents to extract the specific information it needs from vendor systems.
According to Shweta, something that previously took weeks can now sometimes be accomplished in under an hour.
That's more interesting than simply saying, "We use AI."
The technology changes the economics of the business.
If integrating another customer's data source costs thousands of dollars and takes weeks of engineering time, you have one business model.
If it takes an hour, you potentially have a completely different one.
Carolyn described another example involving automotive testing.
Companies can generate millions of individual test data points. Reviewing all of them manually to determine where additional testing is needed isn't practical.
AI can analyze those results and help identify the areas where additional testing may deserve attention. Since physical tests can be expensive, better targeting can potentially reduce unnecessary testing costs.
For founders, this is an important distinction.
The strongest AI use cases aren't necessarily the ones with the coolest chatbot.
They're often the ones where you can say:
This used to take three weeks. Now it takes an hour.
Or:
This used to require ten people. Now it requires two.
Or:
We used to test everything. Now we can focus our resources on the places most likely to matter.
That's where AI starts becoming a business model rather than a feature.
3. Your AI Model Probably Isn't Your Competitive Advantage
This was another interesting point from Carolyn.
AI itself is increasingly becoming a commodity.
Every founder can access sophisticated models. Your competitors can access them too.
So saying, "We use AI" isn't much of a moat.
What matters is what you've built around it.
For Strived, part of that advantage comes from its knowledge of education, customer relationships, data integrations and the work it has done around educational standards and pedagogical practices.
Shweta explained that Strived works closely with customers and tunes its system around knowledge that would be difficult for someone to reproduce simply by vibe coding another dashboard.
Mauro made a similar point about construction.
There is specialized knowledge about framing and construction practices that isn't necessarily sitting conveniently on the public internet waiting for an AI model to retrieve it. That domain knowledge becomes part of the system.
This is where I think founders need to ask a harder question:
If my competitor has access to the same AI models tomorrow morning, what do I still have that they don't?
Maybe it's proprietary data.
Maybe it's workflow knowledge.
Maybe it's customer relationships.
Maybe it's regulatory expertise.
Maybe it's patents.
Maybe it's years of accumulated industry experience.
Whatever it is, "we use GPT" isn't enough.
4. Vibe Coding Is Powerful. It Isn't Magic.
We recently hosted an entire Startup Folsom panel on vibe coding, so naturally I asked this panel about it too.
Yes, these companies are using AI-assisted coding.
Shweta noted that even her CTO, who has a Ph.D. in AI, uses it.
Mauro described using AI heavily in software architecture and development, but he made an important distinction: AI works much better when the human has already broken the problem into clearly defined pieces.
His advice was essentially to use human intelligence to decompose the problem, establish what success looks like and then let AI attack the smaller tasks.
That's very different from typing:
"Build me a SaaS company."
Carolyn used a good analogy during the discussion.
AI might be able to help you manufacture individual parts of a car. But somebody still needs to understand how the engine, transmission, brakes, electronics and everything else fit together.
Otherwise you've built a pile of impressive parts rather than a car.
5. A Demo Is Not a Product
This may be the most important warning for today's AI founders.
AI makes prototypes incredibly easy to create.
Production software is another story.
Carolyn pointed out that vibe coding creates the feeling that you can build almost anything. But if you can build anything, the business question becomes even more important:
What should you build?
She also distinguished between a demo that is good enough to show an investor or potential customer and software that can reliably operate in production.
Mauro pushed the point further.
A production system needs architecture, documentation, maintainable code, testing and the ability to handle scenarios the original developer didn't necessarily anticipate.
AI can write a lot of code.
That doesn't mean the code automatically becomes a product.
I think we'll see a lot of AI startups learn this lesson over the next couple of years. Building the first 80% of an application is becoming dramatically easier.
The last 20%—security, reliability, integrations, compliance, customer support and all the weird edge cases customers inevitably discover—may still be where most of the real work lives.
6. Security and Compliance Become Part of the Product
This is especially important when selling into schools, healthcare, automotive, government or other regulated industries.
Shweta described the procurement and data privacy requirements Strived encounters when working with school districts. The company handles student-level information, which means security isn't something that can be bolted on later.
She also talked about the legal expense involved in navigating district procurement and privacy requirements.
For founders selling B2B AI products, that's worth remembering.
Your product isn't just your software.
Your product may also include:
Security practices
Data governance
Privacy policies
Compliance
Contracts
Procurement
Insurance
Audit readiness
Documentation
None of those things look particularly exciting in a pitch deck.
Customers still care about them.
Sometimes they care about them more than the AI.
7. Watch Your AI Costs
Another audience question focused on something founders don't always think about while building prototypes: compute costs.
Early-stage companies can often receive generous cloud and AI credits through startup programs. Shweta described how Strived benefited from credits from major technology providers during its early development.
That's great while the credits last.
Eventually, however, somebody has to pay the bill.
Strived reached a point where the team spent considerable R&D effort figuring out how to reduce compute costs, including using smaller models for tasks that didn't require the capabilities of the largest frontier models.
There's a broader startup lesson here.
Don't optimize only for whether the technology works.
Eventually you have to ask whether the economics work.
A product that costs $20 in AI compute every time a customer pays you $10 isn't a business. It's a very sophisticated way of losing money.
The Bigger Lesson: AI Makes Building Easier, Not Entrepreneurship Easier
This was my biggest takeaway from the evening.
AI is removing some of the traditional barriers to starting a technology company.
A founder can prototype faster.
A small team can write more software.
Companies can analyze datasets that would have previously required large teams.
AI agents can automate work that once required significant manual effort.
That's all real.
But AI doesn't eliminate the hard parts of entrepreneurship.
You still need to find customers.
You still need to understand their problems.
You still need to convince them to pay you.
You still need to protect their data.
You still need to build something reliable.
You still need to understand your costs.
And you still need some reason why customers should buy from you instead of the 50 other people who now have access to exactly the same AI tools.
In some ways, AI may actually make those things more important.
When everyone can build, knowing what to build becomes the advantage.
About the AI for Business Series
The Starting Your AI Startup panel was held September 2, 2026, at the Roseville Venture Lab as part of Startup Folsom's AI for Business series. The event featured Shweta Gandhi of Strived.io, Carolyn Peer of Humaxa and Mauro Sica of Core Envision. The free monthly series is supported by the City of Roseville and focuses on practical ways entrepreneurs and business owners can use artificial intelligence.
Startup Folsom will continue bringing founders, AI practitioners and business owners together to talk about what's actually working—not just what's getting attention.
Because the interesting question isn't whether AI is going to change how startups are built.
It already has.
The more useful question for founders is:
What can you build now that wasn't economically or technically practical before?
