For the past decade-plus, the startup world has operated around a tidy narrative of how great software companies get built. Digitize a core workflow for a discrete buyer and call that product-market fit; land small, expand fast, and then maximize gross margins by avoiding services revenue but sell more seats with a traditional sales motion.
That playbook created generational companies — and it will lead you astray if you’re building an AI company today.
Not because the underlying principles are wrong, but because the SaaS era produced a set of intermediate instincts that many founders and investors now carry as received wisdom. Instincts about what good margins look like, what a clean go-to-market motion looks like, and what product-market fit feels like once you have it.
Those instincts were earned honestly, but they were trained on a different world.
The first principles haven’t changed: solve a real problem, deliver more value than you capture, build something defensible, find repeatable demand. But the rules of thumb that became axiomatic during the SaaS era are now actively misleading in many cases, and the hard part is that they still feel right. They’re pattern-matched into how founders, boards, and investors evaluate progress.
What follows are the SaaS-era instincts we think AI founders most need to unlearn, organized around three questions every company has to answer: what are you building, how do you build it, and who are you building it for.
What You Build
Your product doesn’t fit a clean category. That’s probably correct.
The SaaS instinct: Map your product to a department and own a category. Salesforce maps to sales. Workday maps to HR. Zendesk maps to support.
The AI reality: AI agents don’t respect organizational boundaries. Agentic workflows traverse systems, functions, and historically siloed roles, and they don’t optimize one department so much as operate across departments.
That creates real challenges for founders. Budget ownership becomes ambiguous when your product replaces slices of multiple roles and nobody is sure who should buy it. Category clarity erodes because you aren’t “a better CRM” so much as a partial replacement for sales ops, RevOps, analytics, and a junior SDR all at once.
If this describes your company, resist the urge to prematurely simplify. AI products often win in the seams, in the friction between systems. But seams are politically messy, and your job is not just to build the agent but to help customers reorganize around it.
This is also why architectural clarity matters more in AI than it ever did in SaaS. You need to get clear, earlier than feels comfortable, on where you sit in the emerging stack. Are you the system of record, the system of action, or the API layer? Are you the context layer that sits above foundation models and below systems of action, which we believe will become the true “foundation model” for enterprise work? In SaaS, you could defer this question for a while. In AI, it determines your defensibility, your pricing, and your competitive exposure from the start.
Your product includes services. That’s not a failure.
The SaaS instinct: Services are a smell: low margin, unscalable, a distraction from “real software.”
The AI reality: In AI, services are frequently the delivery layer of the product itself. The human-in-the-loop work, the workflow configuration, the fine-tuning, the QA layer: these aren’t add-ons but a core and deliberate choice made to accelerate product delivery, adoption, and value creation.
SaaS sold access to tools and workflows, while AI increasingly sells outcomes. That shift collapses the old boundary between product and services and creates a direct correlation between the outcome delivered and the value captured.
In many verticals, the sophistication of the user still lags the sophistication of the product. The mistake is trying to force premature purity by pushing self-serve when the product still requires adaptation, fine-tuning, and personalization to deliver consistent outcomes.
Your early “services” motion is likely where your defensibility is born. Your learning velocity will come from deeply embedded, sometimes messy customer engagements where your team is close enough to the problem to understand what the product actually needs to do. In AI, services are often how the system learns, and the system learning is the moat.
How You Build It
Your margins look ugly. That might be a good sign.
The SaaS instinct: Gross margins should trend toward 80-90%, and anything less raises questions about the business.
The AI reality: Compute reintroduces real economic constraints. Every inference costs something, every training and inference run costs something, and every heavy workflow consumes real dollars.
But here’s the counterintuitive part: the prevalence of low gross margins in early-stage AI companies often denotes product defensibility, reflecting the costs needed to build a potential moat. Meanwhile, a high-margin AI company in the early days may actually be a thin wrapper on someone else’s model and data.
This doesn’t mean margins don’t matter. It means the signal from margins is inverted relative to what a decade of SaaS taught us to expect. The first-principles question of whether you’re building something defensible still applies. The SaaS-era shorthand for answering it doesn’t.
Your company feels like a system, not a product. Good.
The SaaS instinct: Build features, ship them, and iterate on usability and adoption.
The AI reality: AI companies are systems companies. An AI product includes models, prompting layers, guardrails, data pipelines, evaluation frameworks, human oversight, and UX, and these components interact dynamically. When outputs degrade, it’s rarely one bug but rather system drift across multiple interacting layers.
Bret Taylor at Sierra has described how, when their product made an error, they didn’t just fix the error but examined the system that produced it in order to minimize the probability of that class of error happening again. They treat the company like a system, constantly learning and adapting.
This changes what “product quality” means, because it’s no longer just about usability but about reliability under probabilistic conditions. It also changes hiring, since you need people who think in feedback loops, strong evaluation culture from day one, and technical leaders who understand both product and model behavior.
If you treat your AI startup like a feature factory, you will accumulate fragility. But if your company feels more like an organism than a machine, you’re probably building it right.
You’re moving fast, and it still doesn’t feel fast enough. Keep going.
The SaaS instinct: Speed is a competitive advantage, but second-mover strategies can work if you execute well.
The AI reality: In SaaS, markets matured slowly enough that features were legible and distribution advantages compounded over time, which meant second- and third-movers could still win. In AI, learning curves compound faster than your planning cycles. The winners are often the first learners, not just the first shippers, and the market punishes hesitation while disproportionately rewarding a bias for execution. If “fast enough” is your standard, you’re already behind.
Who You Build It For
Your pipeline looks amazing. Don’t trust it.
The SaaS instinct: Landing new logos is the hard part. Once you’re in, expansion is cheaper, and renewals are predictable. The simplified mental model in SaaS was that for every $1 in revenue, the cost was roughly $1.50 to acquire a new logo, $0.70 to expand and $0.25 to renew.
The AI reality: That math is inverted. Experimental budgets are abundant right now, driven by enterprises’ fear of missing out and innovation mandates that encourage trying everything. As a result, landing is easy while renewing is hard.
You can accumulate logos faster than you can accumulate durable value, and at Seed and Series A this is intoxicating. Pipeline looks incredible, inbound is strong, and usage spikes. But usage is not value if the outputs aren’t reliable, defensible, or embedded in real workflows.
Founders need to recalibrate which signals matter. In SaaS, logo count was meaningful early on. In AI, retention and repeatable outcome delivery are far stronger indicators of whether you’re on the right track. Renewal is the sale; everything before it is experimentation.
You found product-market fit. You’ll need to find it again.
The SaaS instinct: PMF, once achieved, persists. It was stabilized by switching costs, data lock-in, habit formation, and slower competitive cycles.
The AI reality: PMF decays. Models commoditize, competitors copy capabilities quickly, APIs improve, and user expectations shift quarterly. A product that felt magical six months ago can feel stale today, and this is true even for companies with $50M or $100M in revenue.
If you’re building an AI company, PMF is something you rent, not something you own.
That means roadmaps must assume PMF erosion and, in some cases, deliberately cause it to happen to stay ahead of competitors in a dynamic market. Teams must tolerate constant reinvention. Leadership must separate temporary excitement from durable value. And founders must accept that metrics will lag reality, because churn signals tend to show up after the competitive shift has already begun.
The founders who struggle most in AI are those who believe they have “locked in” fit. The founders who endure are the ones who treat fit as something they must continuously re-earn.
The Enduring Part
The first principles of great company building haven’t changed. Solve a real problem, deliver value that compounds, build something that’s hard to replicate, and earn your customers’ renewal and trust rather than just their attention.
What has changed is the operating environment in which those principles must be applied, one where markets move at model speed, where the half-life of a good idea is shorter than ever, and where the signals you were trained to trust often point in the wrong direction.
The founders who will build enduring AI companies (AI-native or AI-enhanced) aren’t the ones who throw out everything they know. They’re the ones who can tell the difference between a first principle and a SaaS instinct, and have the discipline to hold onto one while letting go of the other.
If there’s a common thread among the next generation of category leaders, it’s this: they’re not starting from precedent, they’re starting from truth. What’s actually possible now? What would this look like if we weren’t constrained by how it’s always been done? Founders who embrace first principles thinking — and treat history as context, not constraint — aren’t just building companies. They’re expanding the size of the categories themselves. I feel privileged to be partnered with many of these daring founders and entrepreneurs who are building the future for all of us.
Karan Mehandru is a managing director at Madrona, where he invests broadly in B2B SaaS and AI-native companies across early and growth stages. He is an investor in Zapier, Outreach, Cohesity, Auth0, Fyxer, GrowthX, Algolia, Klaviyo, Deepgram, Clerk, Charta, Trek Health, and others.