She Was The Customer Before She Was The Founder. That Was The Whole Product Strategy.

In late 2021, Cecilia Ziniti was the general counsel at Replit. An executive there told her she was a great GC and good at business development, but that her head did not seem to be in it because she was so into this AI thing.

She agreed with him. It was true.

Ziniti had spent roughly two decades as a lawyer by then. She started as a paralegal at Yahoo, worked at Morrison and Foerster, then took senior in-house roles at Amazon, Cruise and the robotics company Anki. In 2013 she became the first full-time lawyer assigned to Amazon Alexa. She had never learned to code, never founded a company, and never held the product or engineering roles most venture-backed software founders come from.

She left Replit on November 1, 2023, and incorporated GC AI about a week later. The company has now raised close to $72 million across three rounds, most recently a $60 million Series B co-led by Scale Venture Partners and Northzone at a $555 million valuation, according to Crunchbase News.

The interesting part is not the valuation. It is what she was doing in the months before she quit.

Before founding anything, Ziniti taught classes on using ChatGPT for legal work. Replit had been working with OpenAI on coding products, which gave her access to an early version of GPT before ChatGPT launched publicly. She spent that access on one question: what does this mean for the people who do my job?

In one class she asked the model about the legal considerations of entering the Brazilian market. The answer covered the right issues. It also opened with something about going to Rio and grabbing a surfboard.

A lot of people saw that and concluded the technology was not ready. Ziniti concluded something narrower. The gap between what the model produced and what a lawyer could hand to a business partner was not a mystery. It was a list. Citations had to be accurate. The software had to show where in a document it found its support. The tone had to match how in-house counsel writes.

As she put it, she would teach in the morning and tell her co-founder in the afternoon what the software needed to do.

That is the mechanic worth stealing. She did not survey the market. She put a general-purpose tool in front of the exact job she knew cold, watched where it failed, and wrote the failures down as requirements.

Ziniti described her advantage plainly. She said she is a better founder for GC AI than she thinks anyone could be, because she was the ICP, the ideal customer profile.

Worth taking seriously and worth bounding. Being your own customer removes a specific cost: the months a founder spends learning what a workflow looks like on a bad Tuesday, which requests are urgent for political rather than legal reasons, what an answer must look like before someone forwards it to a CFO. That knowledge is expensive to acquire and easy to fake badly.

What it does not do is tell you whether the problem is worth money, whether enough companies have it, or whether you can reach them. Plenty of practitioners solve their own version of a problem perfectly and nobody else has quite that version.

The practitioner advantage is compression on discovery. The rest of the company is still the rest of the company.

Ziniti did not build the product. She teamed up with Bardia Pourvakil, an engineer from Replit with experience on earlier generations of GPT. He had been admitted to Stanford Law School and picked engineering instead, partly because he expected AI to absorb some legal work.

That is a specific kind of technical co-founder. Not just someone who can ship, but someone who had already formed a view about this market before the two of them met. The division of knowledge was clean. He knew how to build it. She knew what the work it supports actually is.

For a practitioner turned first-time founder, this is the search that matters most, and the one most people rush. The instinct is to find any engineer willing to say yes. The better filter is someone interested in your industry for reasons that predate you, because they keep making good calls when you are not in the room.

Foundra was built for exactly this stage, when a founder can describe a problem in painful detail but has not yet turned it into a plan someone else can execute against.

Ziniti said trust is the number one most important thing in selling AI software to corporate legal departments. One longtime customer described the difference between GC AI and general-purpose tools as trust, trust, trust, and then trust again.

What makes that more than a slogan is what the company spent money on. About one third of GC AI’s 125 employees are lawyers. It pursued SOC 2 compliance early, built data isolation protections, and says it does not train models on customer confidential information.

Hiring 40 or so lawyers into a 125 person software company only makes sense if you believe trust is a product property rather than a messaging exercise. It shows up in how the software cites, how it hedges, how it declines.

If you sell into a regulated function, ask which of your costs a competitor cannot copy by rewriting their homepage. Certifications are copyable. A staff that has done the job is not.

Growth followed: roughly 2,100 customers, up from about 900 a year earlier, and 400% year over year growth, with names including Lockheed Martin, Time, Eventbrite, Vercel and Gusto.

GC AI competes with Harvey and Legora, both better known. Ziniti’s answer is that they serve different buyers. Harvey and Legora started with law firms. GC AI started with corporations and their in-house teams.

That distinction has product consequences. A contract intelligence tool for an in-house team analyzes one company’s own documents. A law firm tool works across materials belonging to thousands of clients, with the permission and isolation problems that implies. Different products wearing similar names.

First-time founders often read a crowded category as closed. More often the category is one word covering three or four buyers with incompatible requirements, and the funded companies have quietly picked one. Write down who each competitor sells to, then ask which adjacent buyer is served badly by a product built for someone else.

Ziniti is the fifth profile in a Crunchbase series on non-technical founders, alongside MagicSchool’s Adeel Khan, Trunk Tools’ Sarah Buchner, Tabs’ Ali Hussain and Craft’s Ilya Levtov. A real pattern, and also a survivorship sample.

Three cautions. First, Ziniti was not a stranger to venture. She had angel invested and served as GC at several venture-backed startups. Her first commitment came from Amber Illig, whom she had worked with at Cruise, and she noted that once you get one commit, the rest is relatively easy. A practitioner without that network faces a harder first month.

Second, early access to GPT before public launch is not a repeatable advantage in 2026.

Third, the funding market rewards concentration. August saw $42 billion go to just over 1,500 startups globally, up 122% year over year, with seven billion-dollar rounds. Five of those seven had last raised less than twelve months earlier. Capital moves fast toward companies already moving fast, a harder environment for a first round than the headline totals suggest.

If you are sitting in a job you know well and suspect there is a company in it, here is what Ziniti did without meaning to.

Week one. Pick the task in your job that eats the most time and produces the least judgment. Run it through the best general-purpose tool you can access, ten times, on real inputs. Write every failure as a requirement sentence, not a complaint. The output should read like a spec.

Week two. Take that list to eight people who hold your job elsewhere. Do not pitch. Ask them to rate each requirement as must-have, nice, or irrelevant. If four or more cross out your top three, you have a personal workflow, not a market.

Then answer two questions in writing. What would this software have to prove before you would have approved it in your old job? Which competitor is serving your buyer with a product built for a different buyer?

That second question is where the wedge usually is.

Do I need a technical co-founder if I am the domain expert? You need someone who owns the build. Ziniti paired with an engineer who already had a view on the market. What works is complementary knowledge, not a contractor relationship, because early product decisions happen weekly and need both halves in the room.

How do I know if my domain knowledge is actually a moat? Ask what a well-funded generalist team would get wrong in six months, and whether those mistakes would cost them customers. If they would ship something clumsy but usable, your knowledge is a head start rather than a moat.

Is a crowded category a reason not to build? Not by itself. Check who the funded competitors actually sell to. GC AI grew alongside Harvey and Legora by targeting in-house teams rather than law firms, which changed the product requirements, not just the pitch.

What if I do not have investor relationships like Ziniti did? Your first commitment takes longer and probably comes from customers or angels in your industry rather than funds. Build the evidence a stranger would need, which usually means paying customers earlier.

How much should compliance work cost at seed? Enough to close your first real customer, and no more. Data isolation and security review are worth doing early if your buyer is regulated. Certifications you cannot yet tie to revenue can wait.

1. A Startup General Counsel Knew What Corporate Lawyers Needed From AI. So She Built It. (Crunchbase News, September 4, 2026)
2. Global Venture Funding Jumps 122% In August As Streak Of Billion-Dollar Deals Continues (Crunchbase News, September 3, 2026)
3. Craft’s Ilya Levtov Never Learned To Code. He Built A Software Company Anyway. (Crunchbase News)
4. How A Teenage Carpenter Became The Founder Of AI Construction Startup Trunk Tools (Crunchbase News)
5. Nobody Wanted To Give A Former Principal Money: How An Educator Built An Edtech AI Startup (Crunchbase News)