Hilbert: a16z $28M Bet on an AI That Can't Do Math

Today I interviewed Nazli Tan, co-founder and CEO of Hilbert, an AI growth engine for Fortune 500 companies that just raised a $28 million Series A led by Andreessen Horowitz.

A mathematician by training, she ran growth at Getir as it scaled across nine countries to a $12 billion valuation. The twist that makes the conversation interesting is not the raise. It is that Hilbert refuses to be a pure LLM product: its agents are not allowed to do the math, its customers get mapped into Hilbert’s schema instead of the other way around, and its pitch to enterprises spending hundreds of millions on growth is that trust beats clever prompts when the alternative is a probabilistic system deciding where the budget goes.

Structure and Key Highlights

Set Theory as a First Question

• When she met Beril, now her chief of staff, one of the first things she asked was where Beril had learned set theory
• Her read: that question instantly filters for the top 1% of minds she wants to work with
• To an engineer, the elemental unit of math is numbers. To a mathematician, it is sets, because numbers themselves are defined by sets
• Around the third year of pure math, she says you start feeling you can understand anything, including ideas like a set that is both open and closed (cloopen)

Why the Company Is Named Hilbert

• In the early 1900s math went through a foundational crisis: paradoxes and contradictions shook trust in the practice itself
• David Hilbert’s response was Hilbert’s Program, an attempt to put math back on secure foundations
• That was the original name when they raised the pre-seed. She sees growth the same way: not “grow three points faster,” but the foundational reason a company exists
• Left alone, a company diminishes. Retention falls, competition arrives, ad inventory gets more expensive. Stabilizing is already an effort

Would You Bet One Year of Revenue that You Know What Works?

• Her favorite recent question on calls: would you bet your company’s one-year revenue that you know what works and what does not
• She never gets a yes
• Modern business, in her framing, turned into meetings where each team reports its metric, targets get hit, and the company still does not grow sustainably
• Releasing agents into that mess multiplies the mistakes. Wrong definitions scale faster than humans ever could

Logic as a Business Skill

• The skill she uses most at work is logic: proof by contradiction, induction, counterexamples that kill a bad argument in seconds
• She has Russell’s paradox and Cantor’s diagonal argument tattooed on her head, and treats growth as the first place she found the same joy as math
• Growth is users behaving in a space called a company, producing unit economics and behavioral movement. A mind trained for rigor notices when the analysis is wonky

From Citibank and PwC to an Alex Schultz Video in Turkey

• Career start: credit analytics at Citibank, then buying and selling companies at PwC
• She found both boring. Still in Turkey in her twenties, she watched an Alex Schultz video on how Facebook got users and grew
• Schultz is a physicist. The way he talked about growth clicked the same way math had. Years later she did a full-circle episode with him on Startupeable

Quizzes, Cohorts, and Feeling the Company

• At Getir she trained almost everyone who joined her teams. Few arrived already trained
• Method: Alex Schultz videos, quizzes, making people read cohorts until they really got it
• After enough time in growth you start feeling how a company will react to paid budgets, discounts, product changes. Measurement still matters. The feeling is the extra layer she wanted her team to earn

What Bothered Her Most About How Growth Is Practiced

• It took her a long time to believe she was right. Even when she was theoretically sure, she assumed she was missing something
• What frustrated her most: how misaligned people are, and how much expensive C-level and investor time goes into aligning on models and assumptions
• Second: insisting on a wrong analysis. If you can measure something, measure it to the maximum and measure it right. If you cannot, use logic. Not measuring is not the same as trash
• People bullied her for caring about five percent. At gigantic companies those five percents matter a lot

Acquiring a Customer Is More Serious Than Fundraising

• Raising money is often called more serious than marriage. She ups it: acquiring a customer is more serious than raising
• That data point lives forever on your cohorts and retention. She has explained first-year cohorts at twenty-year-old companies
• A badly acquired non-ICP customer is a long, sad investment. Customers look the same at month one and diverge by month three, year one, year five
• Growth decisions are made faster than the journey deserves. Hilbert’s bet is a metric system that gives full visibility without waiting six or twelve months

Closing a Fortune 10 Customer Before the Product Was Finished

• They started bold from day one with four co-founders she would put in her own seat in an exam
• Getir credibility helped, including US expansion, but she does not believe strategies transfer cleanly across markets. Tabula rasa: start with secure tests
• As a growth person she is unimpressed by two-day ARR spikes. They optimized for learning: chase powerful brands and the biggest ones
• Co-founder Ceyda, who she calls a strategy genius, gets credit for early reach-outs. The first big customer clicked with an executive as growth-obsessed and as frustrated as she was, with product roots at Shazam, finishing each other’s sentences

Mid-Market Versus Complex Enterprises

• For mid-market customers, Hilbert sells something they could almost never build internally: too much talent, alignment, and reinventing math systems while flying at full speed
• She pushes back when people ask whether Hilbert replaces the data science department. Foundations should already exist so that team can work on domain problems, not rebuild the kitchen
• Culture line: an intern with a better argument wins. She leads with logic and counterexamples, and changes her mind when beaten

Expert-Led Sales, Then “Give Us an Expensive Problem”

• Early customers arrived pre-product. They started with deep learning models and built the platform on top
• The Hilbert way: the customer is mapped to Hilbert’s schema, not the other way around. Necessary and sufficient vertical schemas, plus the company’s unique sources attached on top
• She had never sold before. Classic discovery felt wrong, so she led with education until Pete Kazanjy’s founder-led sales course told her she was doing everything wrong, even though it somehow worked
• The turn that unlocked deals: ask for the biggest problem they cannot solve, and ask for an expensive one. “Give it to me. Even if I leave, you keep something amazing”

Opinionated Product, Opinion Learned from Their Own Data

• What changes first is the metric system: value, risk, inactivity, personalization inputs
• Hilbert is opinionated, but the opinion comes from a closed learning loop on that company’s history. They ask for a minimum of three to four years of data
• They have never run an engagement where they found nothing wrong in the data, including with teams proud of their data quality. Garbage in, garbage out, and she refuses to let “bad data” become “Hilbert failed”

Not a Pure LLM Product

• First differentiator: they are not a pure LLM product. They use models where each is strongest
• Deep learning predictions are deterministic for the same customer state. LLMs are probabilistic. She asks whether you would deploy hundreds of millions of dollars on a system that answers differently every day
• Agents get a tightly defined problem. Early Hilbert let agents invent risk definitions and they produced composite nonsense that changed run to run
• Context is not enough. Without a harness, agents make junior mistakes, like calling month-to-date shrinkage when ten days are still missing

The AI that Is Not Allowed to Do Math

• Foundational models were bad at basic arithmetic when they arrived. Hilbert does the math on the database and hands results back. Agents may round. They may not add two plus three themselves
• The platform is optimized for trust: absurd test suites, agentic councils, distance to product. Real companies deploying hundreds of millions need that more than clever demos

Causality, Credit, and a Ruthless Coach

• Every engagement starts with an A/B test. If they cannot pass it, they leave
• Causality is hard. Correlation stories like “customers who hit a third order retain forever” are if statements, not if and only if. Inducing the behavior is the real test
• Hilbert labels how the company should be managed and shows quantified wins against how budget is actually deployed. Customers do not always follow the right answer
• When people get offended she asks what they expected from a growth solution: praise, or a list of what to do differently

Pricing Without Seats or Usage Caps

• Appetite splits between success-based fees and fixed fees. She will do either
• What she will not do: seat-based pricing (everyone in the company should use Hilbert) or usage limits that punish heavy use
• Fixed fees reflect business complexity and data volume. Pricing will evolve; right now she optimizes for people actually using the product
• Success raises the baseline, so next year the bar is higher. That is growth: limited resources, best use, compounding loops instead of linear spend

What Silicon Valley Gets Wrong About Growth

• She respects Reforge deeply and learned from it. The critique is about transfer
• PLG playbooks built for self-serve consumer apps do not travel cleanly to grocery, QSR, or retail, where castles are built on Meta and Google
• Paid marketing is good when contribution profit works and journeys are long. Saying “paid does not work” as a universal rule is, in her words, illogical
• Outside the Valley, listing, service quality, and inventory matter as much as funnel. Those questions still lack experts

Joy, Creativity, and Georg Cantor

• With Hilbert, a head of growth stops fighting over whose number is right and starts using creativity the way Alex Schultz meant it: referral mechanics, viral coefficients, hypotheses, not pixels
• Either Hilbert proposes actions or the growth lead sells an idea to Hilbert and they execute together
• Asked who pushed her to think bigger, she thanks Georg Cantor, founder of set theory, bullied for arguing that infinities come in different sizes, and who died in a mental hospital
• Her bar: if someone dared to change the practice of mathematics at that cost, maybe they can change business. She would prefer not to end up in the hospital

Silicon Valley Glossary

People, Companies and References

People Mentioned

• Nazli Tan: co-founder and CEO of Hilbert; mathematician; former growth lead at Getir
• Beril: her chief of staff, also a mathematician, asked about set theory on first meeting
• Ceyda: co-founder she credits for early enterprise reach-outs and strategy
• David Hilbert: mathematician behind Hilbert’s Program, the company’s namesake
• Georg Cantor: founder of set theory; the person she thanks for pushing her to think bigger
• Alex Schultz: Meta’s chief data officer; the YouTube video that pulled her into growth
• Pete Kazanjy: founder-led sales educator who told her she was selling wrong
• Enzo Cavalie: podcast host

Companies and Organizations

• Hilbert: AI growth engine for large consumer companies; $28M Series A led by a16z
• Andreessen Horowitz (a16z): led the Series A
• Getir: the ex-decacorn where she ran growth across nine countries, including the US
• Citibank: credit analytics, her first role
• PwC: valuation and deal work she found boring before growth
• Shazam: product background of the executive who became an early large customer
• Reforge: growth education she respects, while arguing its playbooks do not fully travel
• Meta and Google: the paid castles most large consumer companies still run on

Data and References

• $28 million: Series A led by a16z
• $12 billion: Getir valuation at the scale she describes
• Nine countries: Getir markets she helped grow
• Three to four years: minimum historical data Hilbert wants
• Hundreds of millions: budget scale of the enterprises she sells into
• Fortune 10: size of an early customer closed pre-product
• About 18 months: company age when they spoke