The science of business is scaling
The art of business is doing things that don’t scale.
An MIT report went viral this week, claiming that 95% of enterprise AI implementations got zero return on investment.
We know that B2C scales one way, and B2B another way.
Do AI companies have their own scaling model?
This Friday, I will show a scaling model for AI.
Intro
Most cancer patients don’t benefit from the medicines they take.
Most molecules fail in clinical trials.
I recently hosted Miha Štajdohar, founder and CTO of Genialis.
Genialis uses an AI that understands cancer biology to develop novel RNA biomarkers to match the right patient with the right treatment.
Listen to the podcast here
Genialis began in 2015 as a software company developing AI/ML tools for gene expression analysis. Five years later, they pivoted into an RNA biomarker company.
What excited me was their approach: an AI model that understands cancer at the molecular level. It explains mechanisms of response and resistance and even suggests potential combination therapies.
If an LLM is a foundation model for language, the Genialis™ Supermodel is a foundation model for cancer biology.
Today, one of their biomarkers is close to FDA approval for clinical trials. Their goals for the next 12 months are ambitious:
Increase RNA biomarker adoption in clinical trials
First commercial therapy launch using their biomarker
Expand beyond oncology into immunology and rare diseases
In TechBio - AI scaling is more than business scaling, it impacts life and death for patients.
Let’s start by recalling how traditional B2C and B2B companies scale.
B2C Tech Company Scaling
B2C companies like Uber scale through viral growth and network effects. They focus heavily on user acquisition metrics like cost per acquisition (CPA) and lifetime value (LTV). The playbook often involves:
Building a product that spreads organically through word-of-mouth or built-in sharing mechanisms
Heavy investment in performance marketing across multiple channels (social media ads, search, influencer partnerships)
Optimizing conversion funnels and user onboarding to maximize retention
Leveraging data analytics to personalize experiences and improve engagement
Scaling operations to handle massive user volumes while maintaining low per-user costs
The key challenge is balancing growth spending with unit economics. Companies like Instagram, TikTok, or Spotify exemplify this model - they need millions of users to achieve profitability but can scale exponentially once they hit product-market fit.
B2B Tech Company Scaling
B2B companies like Snowflake scale with a more methodical and relationship-driven approach.
This typically follows a more predictable pattern:
Start with a strong sales team that can handle complex, high-value deals
Develop repeatable sales processes and customer success frameworks
Build partnerships with other B2B vendors to expand reach
Create content marketing and thought leadership to generate qualified leads
Invest in customer success to drive expansion revenue and reduce churn
Gradually move upmarket to larger enterprise clients with higher contract values
B2B companies often scale more predictably but slower than B2C. They focus on metrics like annual recurring revenue (ARR), net revenue retention, and sales efficiency ratios.
Companies like Snowflake use this model - fewer customers but much higher revenue per customer.
Neither B2C nor B2B playbooks fit AI in high-stakes domains like medicine or law. What they need is a trust-first scaling model.
AI Company Scaling
The traditional advice to "do things that don't scale" (famously from Paul Graham) was meant for early-stage B2C companies to find product-market fit.
What I’m seeing with TechBio and AI companies in general is different.
We see this with companies like Flatiron Health and Picnic Health that work very hard, doing non-scalable things with pharma and healthcare providers.
We see this in wildly successful AI companies like Harvey that work very hard, doing non-scalable things with the largest law firms to design and implement their AI solutions in order to earn trust. Once they had trust from the biggest law firms - smaller firms would trust them.
The Harvey example is illuminating because legal work requires trust:
High-stakes decisions where errors have massive consequences
Complex, context-dependent reasoning that's hard to validate
Professional liability and regulatory requirements
Institutional buyers who need extensive proof before adoption
Techbio companies have similar constraints for trust:
Drug discovery and development where mistakes can be deadly
Regulatory bodies (FDA, EMA) that require extensive validation
Hospital systems and pharma companies with rigorous procurement processes
Scientists and clinicians who are inherently skeptical of black-box AI
This leads me to what I call the “Trust-First Scaling Model.”
In medicine and law, adoption requires trust.
Trust demands non-scalable work.
Once trust is won, the network effects can begin.
The Trust-First Scaling Model
Two guests on my podcast, Flatiron Health and Picnic Health did massive manual labeling in order to develop their models.
Flatiron and Picnic Health both use network effects to grow after developing models that achieved the highest performance for their customers and exceeded competing alternatives. Flatiron has a community oncology EHR and Picnic has PicnicAI that can collect medical records from everywhere in the US and transform them into meaningful data for patients, providers, and researchers.
These were long non-trivial processes that took 5-7 years.
B2C scales through virality.
B2B scales through predictable sales motions.
AI scales through trust.
First: system development and creating trust with customers.
Then: the network effects flywheel
Model development
Both Flatiron and Picnic made massive upfront investments in human-in-the-loop data processing that traditional tech companies would never consider. Flatiron had teams of medical professionals manually extracting and structuring oncology data from unstructured EHRs, creating the largest real-world cancer dataset. Picnic had a small army of people who generated over 300 million clinical annotations from clinicians .
This is the hidden cost of trust-first scaling—upfront manual work that looks insane to a Silicon Valley playbook, but creates an unbeatable moat.
They both made a calculated bet that the resulting AI models will create such superior performance that they'll eventually generate network effects.
And that bet paid off.
Genialis has spent the past 5 years developing their Supermodel large molecular model that understands cancer biology and that can explain tumor response and resistance. They recently signed a partnership with Tempus (a company that provides precision oncology services).
As partners start to use Genialis-developed RNA biomarkers, the flywheel will start to turn, and network effects can kick into effect for them as well.
Network Effects
AI companies can leverage classic network effects after they’ve done the hard work of data preparation, and model building and validation.
Flatiron: More oncology practices → better data → better insights → more valuable to existing practices → attracts more practices
Picnic: More medical record sources → better AI → faster/more complete record retrieval → more valuable to patients/researchers → more sources willing to integrate
This suggests that successful techbio companies need to do 3 things
Identify a data-moat opportunity where manual data preparation and R&D can create a sustainable competitive advantage.
Sustain 5-7 years of investment in data preparation and model development
Build out channels of partners using their models.
Outro
If we step back, the scaling path for AI in TechBio looks different from B2C or B2B.
AI companies must first earn trust in their models and systems. Only then can network effects take hold.
Genialis is at the inflection point; FDA approval could turn their years of investment into a compounding data flywheel.
But not every AI company follows a biology-driven trajectory.
Some, like PhaseV, take a very different path. I’ll have more to say about that soon.
For now, if you’re a TechBio founder, ask yourself:
Are you doing hard manual work to clean and prepare data, or building on top of existing structured datasets?
How much of your work is still “non-scalable” custom implementation with customers?
What can you do to scale faster?
About Me
I’m a former pharma-tech founder who bootstrapped to exit.
Now I run a private community with 900+ life science leaders helping them maximize their revenue with the right partners.
I hear these insights first-hand every week from the founders building the future of TechBio.
If you want to get them before your competitors do, join my private network for techbio entrepreneurs.
If you are a techbio leader contact me here to be a guest on the Life Sciences Today podcast.
About Genialis
Genialis is a precision oncology company helping the most promising molecules become life-saving medicines. They develop RNA biomarkers using their AI Supermodel that understands cancer biology at a molecular level and can explain tumor response and resistance.
Visit Genialis here