a16z drew a line in the sand: for software companies, there are two paths forward.
Build AI-native from the ground up. Or retrofit legacy code with AI.
"There is no middle ground," they said. "One will dominate. One will die."
The a16z thesis
a16z looked at 200+ software companies and found a clear divergence in outcomes.
The a16z thesis
- Architecture assumes AI as a core component.
- Retrofitted APIs to legacy code.
- Bolted on LLMs to existing features.
Path 2: Added AI to existing software.
- Retrofitted APIs to legacy code.
- Bolted on LLMs to existing features.
- Tried to make old product competitive with new.
The performance gap
After 18 months:
- AI-native companies: 3x faster feature development, 50% lower cost per user acquisition.
- Retrofitted companies: stuck in pilot mode, internal debates about which features to augment, costs rising.
Why retrofitting fails
When you try to bolt AI onto legacy code, you hit a wall:
- Your data architecture doesn't feed AI well.
- Your product decisions were made assuming humans do the heavy lifting.
- Your entire cost structure assumed humans, so AI savings are offset by new infrastructure costs.
You can't patch a house that was designed without indoor plumbing.
If this plays out
By 2027, the software market splits clearly:
- AI-native companies capture market share, grow faster, raise more money.
- Retrofitted companies slowly become "legacy." Customers switch. Revenue stalls.
There might be some middle ground for very large platforms (Salesforce, Microsoft, etc.) but for startups and SMBs? Clean separation.
Who wins
Founders starting new companies now. They get to build correctly from day one.
Who loses
Companies with 10+ years of legacy code trying to compete with fresh startups built for AI.
Want more on managing in the AI era?
The Little Book for New Managers has helped 30,000+ managers figure out what actually works.
Get the book ($19) →