From SaaS to AI: Why the Old Playbook Is Breaking — and What Replaces It
Source: Upekkha / Paddle Podcast — Rajan, Co-founder & Managing Partner, Upekkha
YouTube ID: nEvZIYe7yiU
Who Is Rajan?
Rajan is the co-founder and managing partner of Upekkha, a B2B accelerator that has backed roughly 180 companies. His thesis is deliberately contrarian: fund what he calls “interesting founders but boring industries” — vertical SaaS and AI applications that Silicon Valley giants (think Satya Nadella or Sam Altman) would never staff a team to explore. Upekkha has been remote from day one, working with founders largely from India building Delaware C-corps for the US market.
How SaaS Happened: Assisted Buying Plus Search Distribution
To understand where things go next, Rajan starts with how SaaS came to be. The first wave was enabled by two simultaneous shifts. First, the delivery model changed: software could be built and shipped over the web instead of installed on-premise. Second, the buying model changed — Google search (and later social media) made it possible for buyers to find software without being prospected.
Before SaaS, an enterprise salesperson walked into an office, pitched a multi-year contract, started with a $100K pilot, and expanded to a $1M deal. SaaS flipped that to assisted buying: users searched "CRM," landed on a homepage, tried the product, and pulled out a credit card. That meant any founder anywhere could access the US market as long as they could rank on Google and deliver a good trial experience.
AI has now broken both halves of that engine.
Traffic Collapse and the End of Blue-Links GTM
When users search today, Google often summarizes the answer inline with AI — it no longer sends them to your website. Rajan points out that even HubSpot, a category-defining SaaS company, has seen traffic crater by roughly 40%. Stack Overflow has suffered similar declines. If your GTM relied on search-led inbound, the ground has shifted beneath you.
“The buying motion changed from selling to assisted buying. That is what enabled SaaS. What AI has done is changed the equation of how traffic flows.”
Rajan's advice is blunt: go look at your own website analytics. If you are not watching demand-shape change, you are operating in a vacuum.
From “Wow Moments” to “Magical Experiences”
Product designers have always talked about wow moments — the instant a user sees value. AI raises the bar. Rajan calls it magical experience: the user tries something and thinks, “I never could imagine something like this was even possible.”
If you are still building the five features everyone expects, while a competitor delivers magic, your customers will ask why you are not giving them the same. He cites Claude Code and Opus 4.5 as near-magical for developers — one-shot app builds that include testing and security checks. Meanwhile, AI SDRs in sales are still struggling because the technology in that domain sits lower on the capability ladder.
The AI Ladder
Rajan maps the capability curve as a ladder:
- Pre-GPT level AI
- Assistant — like an intern (e.g., AI email help for VP Sales)
- Agent — like a junior developer
- Auto-agent — like a peer you can outsource work to (e.g., Cursor, Claude Code, Devin)
Every six months, the cost of intelligence drops roughly 10× and there is a ~30% improvement in capability. The strategic question is not just where your use case sits today, but where the puck is going in the next cycle.
AI PLG: Distribution-First Thinking for a Post-Search World
With search traffic gone, how do you get distribution? Rajan identifies two patterns that are working.
Pattern One: AI PLG
He calls it AIPLG — AI-powered product-led growth. The mechanics are different from old-school PLG:
- Build a magical experience in a narrow use case.
- Let organic word-of-mouth happen.
- Fan the flame with influencer marketing (TikTok, X, YouTube).
These are tiny teams — often 3–5 people — raising $200K or less, yet hitting $1M ARR in under 12 months. Companies like Jenni AI, OpenArt, and Oliv are on this path. The founders think like consumer product managers, not traditional enterprise SaaS founders. Capital efficiency is their source of control.
Pattern Two: Top-Down Enterprise
The other working playbook is the classic enterprise sale, supercharged by board-level FOMO. The board asks, “What is our agentic strategy?” That question filters down to the CEO, then the CXO, then line-of-business heads. If you can sell to the board, you can build a massive enterprise GTM. But this requires serious capital — $5–10M to get started — and is nearly impossible to execute remotely.
The Missing Middle
The segment Rajan does not have an answer for is the midmarket ($2K–$10K ACV) that was historically the playground for global SaaS founders enabled by Google inbound. That playbook is in flux, and nobody has figured out what replaces it.
The Slot Machine Psychology of AI Products
One of the more counterintuitive observations Rajan makes is that AI tools have a slot-machine psychology. You prompt the tool, it almost works, and you think, “Maybe it is me,” so you pay again and try again.
“Whether you are using Lovable or any other AI tool, it almost works, it does not, and then you are like, 'Maybe it is me, it's not working,' so try again, try again. Oh money ran out, credit ran out, let me put more money.”
This is why developer tools — long considered impossible to monetize because “developers don't buy anything” — are now massive businesses. The magical promise keeps users inserting credits. But there is a flip side: annualized cohort churn is brutal. Experimentation is high. Rajan argues founders should design for this psychology intentionally, not ignore it.
Why SaaS Metrics Are Dying — and What Replaces Them
The golden age of SaaS was built on predictability: low churn, annuity revenue, clean NRR benchmarks. AI products look more like consumer products:
- High trial, high exploration, high natural churn.
- Users are “tourists” — they come, experiment, leave.
- Recurring subscriptions may not be the dominant model.
Rajan flips the equation: instead of obsessing over churn, study who retained and why. Focus on cohort retention curves, not churn rate. Expand transaction frequency (weekly to daily) and deepen the product portfolio to make committed users sticky. Instagram is stickier than most SaaS products — it just operates under a different business model.
The Immigrant Founder at Minus Three
For immigrant founders arriving in the Bay Area, Rajan offers a brutal but useful framing: you start at minus three on three dimensions.
| Dimension | The Gap |
|---|---|
| Roots | No local network for warm introductions to employees, partners, or investors. |
| Reputation | Without a Harvard/Stanford brand or ex-FAANG halo, doors do not open as easily. |
| Revenue | Bay Area expectations are now 5× annual growth, not the old T2D3 standard. If you have that slope, the other two barely matter. |
Practical hacks: leverage your cap table for credibility, audit your LinkedIn for former colleagues who can make warm intros, and stop undervaluing your past employers and universities.
Pivot Framework: 20% vs 200%
Rajan distinguishes between two kinds of pivots:
- 20% pivot — change one of: target segment, problem statement, positioning, or product.
- 200% pivot — change all of them simultaneously.
The 200% pivots are emotionally devastating because founders tie identity to their vision. But if you have spent ~18 months and growth is stuck at 2–3%, you may be climbing the wrong mountain. Funded founders sometimes have an advantage here: investors apply external pressure that forces re-evaluation. Bootstrapped founders can stay emotionally attached too long.
“Vision is something that is most beautifully visible in the rearview mirror.”
Why Rajan Changed His Mind on Models
Eighteen months ago, Rajan argued that models did not matter and that startups should focus on the wrapper. He has reversed that position.
His reasoning: if you prove a use case has product-market fit but do not own the model layer, a well-resourced incumbent will assign a PM and an engineer and swallow the opportunity. Cursor made $1B in revenue plausible for AI coding — so Anthropic built Claude Code. Without control over the model layer, you cannot retain the intelligence your product learns from usage, and your unit economics will collapse under API costs.
He now believes no consumer AI startup survives without building and deploying its own model. The three durable pieces are:
- Interface innovation — new UX paradigms (voice, terminal, etc.).
- Integrations — especially critical in B2B.
- Intelligent product — the product learns from usage and improves itself over time.
Wartime CEOs and the Six-Month Horizon
Rajan borrows Satya Nadella's framing of peacetime vs wartime leadership. Adobe and Microsoft both executed brilliant transitions from desktop to cloud. In the AI era, Satya is playing the infrastructure game, not the model game. But Rajan is bearish on incumbents like Salesforce and Adobe if they do not move faster.
He points to the Workday CEO as an example of wartime execution: fired 1,600 people in February 2025, then spent ~$3B on three acquisitions by October 2025 to shift from “system of record” to “role-based agents” — all executed in three quarters.
Rajan refuses to look three years ahead. Scaling laws mean the landscape reinvents itself every six months. The only safe horizon is six months.
Key Lessons
- Search traffic is gone. If your GTM still assumes inbound via blue links, you are operating in a past era.
- Product expectations have shifted from “wow” to “magic.” If a competitor is delivering magical experiences, expected features are table stakes at best.
- AI PLG is the new capital-efficient playbook: tiny team, magical UX, influencer fanning. $1M ARR in 12 months on $200K capital is real.
- Metrics need to match the business model: SaaS metrics may not apply. Study retention curves, expand transaction frequency, and deepen the product for committed users.
- First-principles thinking beats playbooks right now. There is no replicated formula yet; the founders winning are the ones reasoning from fundamentals.
- Own your model layer. If you do not, your PMF is a roadmap for better-funded competitors, and your unit economics will not survive.
- Wartime execution velocity wins. Peacetime leaders who navigated cloud transitions may not survive the AI shift.
Why This Matters for Diffie
Diffie is an AI-powered browser testing tool for frontend engineers. Rajan's frameworks map directly:
Distribution is social, not search. Engineers discover tools on X, GitHub, and through peer recommendations. If Diffie has not invested in developer influencers and community credibility, that is the unlock.
Magical experience is the wedge. Testing tools are notoriously unloved. The move from “check my UI for regressions” (expected) to an autonomous result the engineer did not think possible (e.g., predictive failure detection or automatic fix suggestions) is the gap between a vitamin and a painkiller.
Climb the AI ladder deliberately. Browser testing sits between assistant and agent today. With intelligence costs dropping 10× every six months, Diffie should build toward auto-agent capability, not polish assistant-level UIs.
Metrics are consumer-grade. Engineer tools have high trial and high natural churn. If Diffie is charging SaaS-style annual subscriptions by seat, consider whether the product psychology is closer to a consumer tool. Design for retention curves and transaction expansion, not just low churn.
Model ownership is a strategic choice. If the only moat is the interface, a competitor can replicate it fast. Diffie should decide explicitly whether fine-tuned models on DOM patterns and test-failure data are part of the roadmap. Without that, you are proving the market for someone better-funded.
Execute at wartime speed. Legacy testing suites (Selenium, slow enterprise platforms) are incumbents with high switching costs but low velocity. This is the window to own the terminal/developer UX, integrate deeply into CI/CD, and make the product learn from every test run so it gets smarter the more it is used.