From the Wall Street Journal on March 16th:

“We cannot miss this moment because we are distracted by side quests,” Simo told staff last week, according to remarks reviewed by The Wall Street Journal. “We really have to nail productivity in general and particularly productivity on the business front.”

According to the Wall Street Journal, OpenAI’s leadership, including CEO Sam Altman and chief research officer Mark Chen, is actively determining which product lines to deprioritize. The era of launching Sora, Atlas, an eCommerce layer, and a Jony Ive hardware device simultaneously is over. OpenAI is going after the enterprise.

The obvious question is: why? OpenAI has more than 800 million weekly active users. It is, by almost any measure, one of the fastest-growing consumer products in history. Why would a company with that kind of distribution pivot away from consumers and toward selling software to businesses?

The answer has less to do with OpenAI’s strengths than with the fundamental reality that monetizing AI for consumers and for enterprises requires two entirely different businesses.

For most users today, AI acts as a productivity tool. It drafts emails, writes code, summarizes documents, and analyzes data. OpenAI’s enterprise report confirms this, reporting that AI has improved the speed or quality of output for 75% of surveyed enterprise workers, with average time savings of 40-60 minutes per day.

Historically, enterprise customers have been the largest customer segment for productivity tools. This pattern is remarkably consistent. Notion started as a note-taking app for consumers. But today, it generates $600 million in annual revenue, with approximately 90% of its business coming from team and enterprise usage. Similarly, both Dropbox and Figma started as tools for individual consumers, but eventually pivoted into the enterprise. For productivity tools, consumer adoption only creates awareness, but it is enterprises that write the checks.

The reason is straightforward: enterprises are more rational than consumers. A company can put a precise dollar value on the time a productivity tool saves each user. Let’s say there is a productivity tool that saves each user one hour per day. For a company that pays its employees $30 an hour across 20 days per month, the hour saved per day is worth $600 per user per month in recaptured labor. So if the tool costs less than $600 PUPM, the purchase justifies itself for the company.

Consumers, by contrast, rarely pay for productivity tools because for consumers, productivity isn’t measured in monetary terms. Imagine the same tool that saves each user an hour per day. How much would you pay for that hour? For most consumers, the answer is $0. Fundamentally, consumers are not businesses that need to optimize for productivity; therefore, they have little incentive to pay for productivity.

This fundamental asymmetry in consumers’ and enterprises’ willingness to pay for productivity tools is the most important dynamic for understanding where a significant portion of AI revenue will come from, and it points squarely at the enterprise.

Consumers and AI

None of this means consumers don’t matter. They are a massive segment of AI users. Consumers are rapidly integrating AI into their daily lives, using it for everything from drafting emails to planning meals to asking for life advice. After all, ChatGPT alone has 800 million weekly active users, clear evidence of the size of the AI consumer market.

But the way to monetize those 800 million consumers is fundamentally different from the way you monetize enterprise seats. Out of ChatGPT’s 800 million weekly active users, only 5% actually pay for a subscription. Consumers simply don’t pay for productivity tools with money. They pay with their attention, which means they pay through ads.

The analogy here is Google. Imagine if, in 2002, Google had launched as a $20-per-month subscription search engine. It would never have become the company it is today. Search can be thought of as a kind of productivity tool that helps the user find information faster. However, consumers don’t value productivity monetarily. Instead, Google monetized those consumers by building the most lucrative ad business in history. In FY 2025, Google generated over $264 billion in total ad revenue, with search accounting for roughly $200 billion.

The reason this works is that consumers, at the moment of a Google search, are expressing intent. They’re close to a decision: what to buy, where to eat, or which product to choose. That intent is extraordinarily valuable to advertisers, valuable enough that advertisers will effectively subsidize the consumer’s access to Google in exchange for reaching them at their moment of decision.

AI can be an even better version of this. When a user asks ChatGPT what the best running shoe for a marathon is, they are expressing higher-quality intent than in a typical Google search. A Google search for ‘best marathon running shoes’ tells an advertiser that this user is interested in running shoes. A conversation with a chatbot might reveal that the user is a beginner training for their first marathon, has flat feet, runs on pavement, and has a $150 budget. Essentially, ChatGPT can assemble a customer profile in real time through natural dialogue. This massively increases the advertisers’ ability to target their ideal customer profile, resulting in ads with far better returns.

OpenAI’s Dilemma

For OpenAI, it is caught between two enormous opportunities that require two entirely different capabilities, and in the short term, it has to choose.

On the consumer side, the TAM is staggering. Google’s search ads alone generate over $200 billion in annual revenue. If OpenAI can capture that consumer market, it would dwarf enterprise revenue. Truist estimates OpenAI will generate under $1 billion in ad revenue this year and scale to over $30 billion by 2030, a 134% CAGR. And search advertising is built around a powerful flywheel: as people use the product, they generate data that can be used to improve responses as well as ad targeting. This attracts both more users and more advertisers, which generates more data, and the flywheel continues. This is the exact mechanism that allowed Google to own more than 90% of the search market for over two decades.

However, building and scaling an ad business is extraordinarily difficult. Google launched AdWords in 2000. And it took years of building auction systems, self-serve advertiser tools, and an entire organizational apparatus dedicated to ad tech for AdWords to become the core of a $200 billion business. ChatGPT launched ads in February 2026 with a $200,000 minimum buy-in and PCMs around $60, roughly three times Meta’s average rates. While the high buy-in amount already deters some advertisers, early reports also suggest low click-through rates and glitches, further disincentivizing advertisers. Ultimately, building an ad engine is a years-long organizational build and not a quarter-over-quarter revenue story.

On the enterprise side, the opportunity is more immediate, but the competition is also more intense. Unlike running advertisements, serving enterprises does not require building an entirely new revenue engine that needs to be optimized over multiple years. Instead, it’s about implementation, integration, and go-to-market execution. But the competition is fierce. OpenAI generates approximately $10 billion of its roughly $25 billion annualized revenue from enterprise customers. But its enterprise market share has been falling, dropping from around 50% in 2023 to approximately 27%. Meanwhile, Anthropic’s enterprise market share grew from 12% in 2023 to 40% in 2025. Out of Anthropic’s $19 billion of ARR, roughly 80% came from enterprise clients.

Unfortunately for OpenAI, it probably cannot pursue both the consumer market and the enterprise at full intensity simultaneously. There are real limitations to the amount of compute available, which means OpenAI must make real tradeoffs when considering where to allocate that compute. However, the deeper issue for OpenAI is organizational focus. An advertising engine and an enterprise platform require fundamentally different teams, different product priorities, different sales motions, and different cultures, especially at the scale OpenAI hopes to operate at. Ultimately, this is the fundamental reason OpenAI has decided to deprioritize side projects and refocus on a single business model.

For OpenAI, with a potential IPO as early as Q4 2026, the pressure to show high-quality revenue growth is immediate, and enterprise revenue is the fastest path to demonstrating high-quality revenue. Enterprise software revenue is predictable, high-margin, and defensible, which means Wall Street is willing to underwrite a much higher multiple on enterprise subscription revenue than on an early-stage ad business with unproven unit economics. And the opportunity for OpenAI to capture that revenue is fleeting. Anthropic’s accelerating enterprise dominance has created a now-or-never urgency. Every day that OpenAI doesn’t aggressively pursue enterprise is a day where Anthropic’s integrations get deeper, switching costs get higher, and its lead gets harder to close.

The Connection Between Enterprise and Consumer

From Ben Thompson on last week’s Sharp Tech podcast:

The size of [the enterprise] market may be larger than Adrian is giving it credit for, and that may be a sufficient flywheel to then go into the consumer market. And there is an analogy for this, which is Microsoft in the 80s and 90s. Microsoft was an enterprise company; they got the Windows flywheel going in enterprise and basically won the consumer market for free

Windows dominated the enterprise in the 1980s. Workers grew familiar with it at the office and saw no reason to pay for a different operating system at home, so Windows became the consumer default by sheer inertia. However, this analogy breaks down when you consider the fundamental economics of AI.

Windows won both markets because the capability it provided was fundamentally the same in both contexts. Whether it was running applications, managing files, or navigating a desktop, the experience was identical whether you were sitting in a corporate office or at your kitchen table. The difference between Windows for the enterprise and Windows for consumers was licensing and IT administration, not the core product. An accountant running spreadsheets at work and a teenager playing games at home were using the same operating system in the same way

AI is not like this. The value AI provides to enterprises and the value it provides to consumers are fundamentally different. For enterprises, AI is a productivity tool that either replaces or augments labor. The product needs deep integrations like Slack, SharePoint, GitHub, Salesforce, plus security compliance, auditability, and enterprise-grade reliability. The go-to-market is a sales-led motion with long cycles and high average contract values.

For consumers, the value AI provides is closer to Google Search than traditional enterprise productivity tools like Microsoft Office or Asana. It’s a search engine, a decision-making tool, and, increasingly, a source of entertainment. The product needs great conversational UX, broad knowledge, ecommerce features, and an ad infrastructure that can monetize intent at scale. The go-to-market is product-led with zero marginal revenue per user, and monetization happens through a third party: the advertiser.

These aren’t two versions of the same business. They’re two different businesses that happen to share an underlying technology, which means winning one does not hand you the other. Therefore, for OpenAI, the choice between enterprise and consumer is a real choice and not just a sequencing question where you win one first, and the other follows. And with a potential IPO on the horizon and an ad business still in its infancy, OpenAI has chosen the enterprise.

Whether that’s the right long-term call is genuinely uncertain. The risk is that OpenAI trades a potentially dominant consumer position, one with hundreds of billions of dollars of TAM, for a position in enterprise where it’s playing catch-up against a competitor that already has a meaningful lead. Meanwhile, Google, a major competitor, already has a mature ad infrastructure and its own bid to win the chatbot market in Gemini. For OpenAI, the consumer window may not stay open forever.

But the IPO clock is ticking, and enterprise revenue is the kind of revenue that bankers and investors are willing to underwrite. Sometimes the right strategic choice isn’t the biggest opportunity; it’s the one you can actually execute on in time.

From Clayton Christensen’s Disruptive Innovation Theory:

Disruptive Innovation describes a process by which a product or service takes root in simple applications at the bottom of the market—typically by being less expensive and more accessible—and then relentlessly moves upmarket, eventually displacing established competitors.

That theory became the foundation of Christensen’s The Innovator’s Dilemma, the defining framework for how great companies are disrupted. And every so often, a company emerges that maps onto that model almost perfectly. Cloudflare is one of them.

When Cloudflare was founded in 2009, the internet’s leading CDN incumbent, Akamai, was optimized around the largest customers. Akamai was built around high-touch enterprise relationships and bespoke implementations. This enabled Akamai to charge premium prices to its customers, leading to high ACVs and LTVs in return for a high CAC, a model that made perfect sense for serving the largest websites in the world.

Cloudflare, by contrast, started with a freemium model, instant onboarding, and a product that any small developer or website owner could adopt in minutes, which allowed Cloudflare to win smaller customers like personal blogs and e-commerce sites. Fundamentally, Akamai was structurally disincentivized to pursue that part of the market. The lower ACVs from small customers meant that Akamai’s long sales cycle and bespoke implementation process could not return the high LTV a Fortune 500 customer could. Therefore, the customers Cloudflare went after were almost immaterial for Akamai. As a result, Akamai largely ignored Cloudflare as it built a massive user base.

And once Cloudflare established itself at the low end of the market, the rest of the story is familiar. Cloudflare layered on additional functionality, expanded into a broader platform, and steadily pushed upmarket. Because Cloudflare had already built a massive user base that loved the product, Cloudflare turned that adoption and credibility into larger enterprise contracts as its customers scaled alongside it.

It’s also worth noting that Cloudflare is a reminder that the software and internet markets have historically been unusually unforgiving because distribution is not a limiting factor, as it is with other industries. As a result, once a product becomes the default choice for developers or operators, it can move through the market far faster than incumbents expect.

Software, the Innovator’s Dilemma, and AI

From Gavin Baker on Invest Like the Best:

If you’re trying to preserve an 80% gross-margin structure, you are guaranteed not to succeed in AI.

SaaS companies are making the exact same mistake that brick-and-mortar retailers did with e-commerce. They clearly saw customer demand, but they didn’t like the margin structure. Now Amazon has higher margins -- margins can change.

SaaS companies have their 70-90% gross margins and are reluctant to accept AI gross margins -- a good AI company might have 40%.

In some ways, this is the same Innovator’s Dilemma Cloudflare exploited during its rise. As Baker argues, SaaS companies might be unwilling to accept the lower gross margin economics that AI agents require because these companies are too married to the high gross margins and exceptional unit economics of the software model, which in turn creates an opening for AI-native startups that are willing to operate with lower margins to disrupt incumbent software vendors.

Under Baker’s model, the most exposed companies are not weak companies; they are some of the strongest software businesses in the world. ServiceNow reported a non-GAAP subscription gross margin of 83.5%. Atlassian reported 87.5% non-GAAP gross margins for Q2 FY 2026. Both ServiceNow and Atlassian are great businesses on a gross margin basis. But that’s precisely the point Baker is making. Namely, the higher the quality of the legacy model, the greater the temptation to defend it rather than reinvent it.

And just like with Cloudflare back in the 2010s, software is a winner-take-most game due to the lack of distribution limitations. If a software solution can win credibility with users, it can move from niche to dominant far faster than in other industries. We are already seeing evidence that this transition can happen at an absurd pace. In November 2025, Sierra reported $100 million ARR only seven quarters after launch, one of the fastest growth trajectories in software history.

There is, however, an important distinction between the disruption that software is experiencing today and the CDN market Cloudflare disrupted in 2010. As Baker notes, there is already clear customer demand for AI agents. It is not as if agentic solutions from companies like Sierra only appeal to the low end of the market. In fact, Sierra’s customer base already includes Fortune 500 customers like CDW and SiriusXM.

However, while the margin-side mechanism of the Innovator’s Dilemma still applies, the demand-side mechanism is weaker. While Akamai’s core customers had little reason to care about the lower-end product Cloudflare initially offered, the customers of incumbent software companies like ServiceNow and Atlassian have clearly shown demand for AI solutions. Therefore, while the mechanisms of the Innovator’s Dilemma still apply to today’s incumbent software companies, as McKinsey notes in their article, The AI-centric imperative: Navigating the next software frontier:

For incumbent software companies, the imperative is clear: becoming AI-centric is no longer optional—it is essential to remain competitive. Those that can successfully adapt and thrive will help define the next era of software.

Cloudflare and AI

What makes Cloudflare so interesting, though, is that it may be one of the rare incumbent software companies for which this exact dilemma applies even less. For many software incumbents, AI is threatening because it may abstract away the interface, collapse seats, or reduce the importance of the legacy application layer. For Cloudflare, its value is derived not from seats but from internet activity. And AI is not removing activity from the internet; it’s adding more of it.

Cloudflare’s own recent data makes that clear. In its 2025 Radar year-in-review, Cloudflare reported that AI crawlers accounted for ~20% of verified bot traffic. In a separate analysis, Cloudflare said AI and search crawler traffic rose 18% from May 2024 to May 2025, with GPTBot traffic up 305% over that period. In other words, AI is only increasing internet activity, and Cloudflare is still the software securing it.

In that sense, Cloudflare is not a classic application incumbent defending a high-margin seat-based workflow against a smarter interface. It sits lower in the stack, in a position where it benefits as long as there is more internet activity, regardless of whether that activity is generated by humans or by AI agents and crawlers.

Just as importantly, Cloudflare is not complacent; it is actively building to meet the AI shift. During AI Week 2025, Cloudflare introduced Web Bot Auth so sites can distinguish agent traffic from other verified bots and decide what to allow or block. Cloudflare has also seen success in its products around AI security, crawler controls, and agent authentication. In July 2025, Cloudflare announced that over a million customers had enabled its one-click AI crawler blocking option since its launch in September 2024. Rather than defending a legacy model against AI, Cloudflare is actively trying to become the infrastructure layer that governs how AI interacts with the web.

In a sense, it is quite ironic. Cloudflare was once the disruptor that exploited the Innovator’s Dilemma against incumbents like Akamai. Today, as Cloudflare has become the incumbent, AI has created a new version of that dilemma for large software companies. However, instead of being disrupted by it, Cloudflare looks more like a beneficiary of the AI transition.

Ultimately, the Innovator’s Dilemma punishes incumbent companies whose existing economics make it irrational to fully embrace new technology, which is why AI is so dangerous for traditional software incumbents built around beautiful SaaS margins. And that is why Cloudflare is so unusual: the company’s existing strategic position is already aligned with a world where AI creates more internet activity, not less. In 2010, Cloudflare won because CDN incumbents could not justify becoming more like Cloudflare. Now, Cloudflare may win because the next internet looks even more like Cloudflare’s world than the last one did.

From Alfred Lin, Sequoia Capital partner and co-steward, in an interview with Bloomberg last week:

“Let’s not forget AI is a lot of software,” Lin said in a Bloomberg Television interview on Wednesday. “Legacy software companies of yesteryear like Oracle still exist today.”

Lin added that he believes AI ultimately will improve most companies. “The impacts of AI are real,” he said. “They’re going to allow us to do a lot more than we used to be able to do.”

Lin’s point is hard to argue with. AI clearly allows software companies to “do a lot more than [they] used to be able to do,” whether that’s automating work that once required teams of operators or compressing end-to-end workflows into a single system. If you treat AI-native products as part of the broader software market, it’s easy to see why this isn’t just a feature cycle. AI is actively expanding what software can address, and, in doing so, broadening software’s total addressable market.

That expansion shows up in three main ways: AI can solve previously intractable pain points, it can penetrate strategic positions that software has historically struggled to reach, and it can actually replace human labor TAM. Those three modes of expansion are perhaps onto three AI-native companies, namely Clay, OpenEvidence, and Sierra.

Solving Previously Intractable Pain Points: Clay

Clay is an AI-native GTM platform that was most recently valued at $5 billion in an employee tender offer. If Clay is doing ~$100M in ARR (a milestone it hit in December 2025), that implies a 50x ARR multiple, underpinned by 250% revenue growth and a staggering enterprise net retention rate north of 200%. Clay has also built a deep roster of credible early customers, including OpenAI, Anthropic, and Rippling.

Clay fundamentally operates as an orchestration layer on top of its data-provider integrations, namely tools like Wiza, Apollo.io, and RocketReach. When a user searches for a lead, Clay queries multiple providers in parallel to reliably surface the right contact. If one source comes up empty, another often has the missing record, so coverage and accuracy can improve simply by aggregating across the full set of providers Clay can tap.

However, as powerful as that position is on its own, Clay’s real value proposition is what AI enables on top of it. With AI, Clay can go beyond pulling records from integrations. Clay can automatically search the web, synthesize context across sources, enrich each lead with the details that actually matter, and then generate outreach tailored to the specific person and company at a level of personalization that would be impossible to do manually at scale. For example, a rep can feed Clay a list of 500 target accounts. Clay can then automatically find the right persona from each account, pull and verify contact details across providers, scrape the open web for relevant signals, and draft a tailored first-touch email that connects to a targeted pitch.

This automation of personalization is what ultimately underwrites Clay’s rapid growth and the staggering valuation multiples that come with it. Clay attacks a core GTM bottleneck: the traditional tradeoff between high-effort, customized outreach and high-volume, generic outreach. By making tailored prospecting scalable, Clay lets teams run far more efficient GTM motions than previously possible. Importantly, GTM efficiency ties directly to a company’s upside. Better targeting and messaging translate into a better pipeline and more revenue, which means customers are willing to pay for a tool that meaningfully expands that ceiling. And that jump in efficiency is only possible with AI.

Unlocking New Revenue Pools: OpenEvidence

Another way AI expands software’s TAM is by unlocking previously untapped revenue pools through embedding itself in important strategic positions that legacy software couldn’t reach effectively. There is arguably no better example of this than OpenEvidence. Positioned as an AI copilot for clinicians at the point of care, OpenEvidence effectively functions like a “ChatGPT for doctors,” optimized for clinical questions where accuracy and sourcing matter. OpenEvidence’s most recent funding round was a $250M Series D in January 2026 at a $12B valuation, co-led by Thrive Capital and DST Global. This doubles the $6B valuation it set in its $200M Series C in October 2025, led by Google Ventures.

That meteoric rise is driven by many factors. For example, OpenEvidence has a proprietary data advantage. Unlike general-purpose copilots trained on the open internet, OpenEvidence is trained on licensed, peer-reviewed medical literature. OpenEvidence has inked exclusive deals with publishers and clinical evidence providers such as the JAMA network, the New England Journal of Medicine, and Wiley to access trusted research. As a result, when a doctor uses OpenEvidence, they get a level of confidence that cannot be achieved with a general-purpose copilot. This, paired with OpenEvidence’s first-mover advantage at the point of care, that data moat has helped it build a meaningful lead in clinical Q&A, a lead that becomes harder to close as usage, feedback loops, and workflow lock-in compound over time.

However, OpenEvidence’s biggest value add is commercial. OpenEvidence has opened a new on-ramp into healthcare’s largest marketing spend, pharma. By sitting inside a clinician’s workflow at the point of care, OpenEvidence creates premium real estate for pharmaceutical companies to reach doctors in context, adjacent to the exact clinical questions and treatment decisions that matter.

Pharma has long spent heavily to influence prescribing, but most channels are either indirect, such as advertising through patient-facing media, or expensive and operationally intensive, such as advertising through field reps, conferences, and legacy publications. OpenEvidence offers targeted, measurable exposure to clinicians at the moment of decision, with tighter feedback loops and clearer attribution. In that sense, AI isn’t just improving the product; it’s enabling a strategically advantaged distribution surface that unlocks a revenue pool traditional software couldn’t credibly access.

It’s also worth noting that OpenEvidence only unlocks this revenue pool because it first solves a previously intractable pain point. Before AI, there was no scalable way for physicians to retrieve and synthesize the right medical evidence quickly enough to be useful in the exam room. OpenEvidence uses AI to make that kind of point-of-care clinical lookup fast and practical, and the advertising surface is downstream of that utility. In other words, the ads work because the product first earned its place at the moment decisions get made, and OpenEvidence was only able to earn that position due to AI.

Replacing Human Labor: Sierra

It is important to note that for both Clay and OpenEvidence, AI is augmenting human labor. Clay makes GTM teams much more efficient, and OpenEvidence helps clinicians become faster and more accurate at the point of care. But the third, and arguably most TAM-expansive, mode is when AI replaces human labor wholesale. A good example of this is Sierra.

Sierra is an AI-native customer experience agent platform that was last valued at $10B after raising $350M in September 2025. If you benchmark that against Sierra’s reported $100M ARR milestone, the headline valuation implies an eye-popping ~100x ARR multiple. Sierra has already signed numerous blue-chip customers, including Discord, Ramp, Rivian, SoFi, and SiriusXM.

Sierra’s core value proposition is simple: it can replace customer service labor end-to-end, not just assist it. Sierra’s AgentOS platform can handle the customer interaction, execute the downstream action, and then write back to the systems of record. For example, if a customer initiates a conversation asking for a refund, the agent can resolve the conversation, issue the refund, and log the outcome in the relevant back-office ERP system, all without a human in the loop. And that no human in the loop promise shows up in Sierra’s monetization. Sierra prices on outcomes, meaning Sierra only gets paid when the interaction is completed without a human from start to finish.

This is where the TAM expansion becomes obvious. With AI, Sierra isn’t just selling a better CX software; it’s selling into the entire customer service labor spend. And that’s a big reason investors are comfortable backing multiple top players in the category. You can see the same investor conviction in other agentic CRM solutions like Parloa (which raised a $350M Series D at a $3B valuation) and Decagon (which announced a $250M raise at a $4.5B valuation). The prize for AI-native CRM solutions is not a marginal slice of SaaS budget, but a direct claim on a massive labor pool.

AI Limitations

However, even if AI can expand software’s TAM in meaningful ways, it still has real limitations, especially around reliability. A common thread across Clay, OpenEvidence, and Sierra is that each operates in a domain with a built-in tolerance for failure. With Clay, the downside of an imperfect output is usually limited. If the agent misses a lead or gets a detail wrong, the worst case is a lost outreach attempt. It’s frustrating, but rarely catastrophic. OpenEvidence has a different safety valve. A trained clinician remains the ultimate decision-maker. If the model is wrong, a knowledgeable doctor can often catch it. And if they can’t immediately validate the answer, the doctor can still flag uncertainty rather than treat the output as ground truth. Sierra, meanwhile, can handle failure by simply escalating edge cases. If a request is too complex, sensitive, or ambiguous, the system can hand off to a human agent before anything irreversible happens. In other words, these are AI products designed around bounded risk where errors are either non-fatal, human-auditable, or recoverable via escalation.

But there are domains where the downside of failure is catastrophic. Take Jeppesen ForeFlight, an aviation software company that Thoma Bravo spun out of Boeing in a $10.55 billion deal. Jeppesen ForeFlight sits directly on the critical path of flight operations, from flight planning and weather to aeronautical charts and navigation data. In that environment, an error is much larger than a slightly worse conversion rate or an easily corrected suggestion. It can translate into a wrong route, a missed constraint, or a compliance and safety incident with real consequences.

Fundamentally, today’s AI is a poor match for domains like aviation, where the tolerance for error is effectively zero. Because AI systems are probabilistic rather than deterministic, they can produce rare but unpredictable failures, and in safety-critical environments, “rare” is still unacceptable. The problem is compounded by the fact that a human-in-the-loop often can’t reliably catch or correct an AI’s mistake in real time, especially under operational time pressure. In those settings, AI may still be useful at the margins, but broad autonomy is unlikely if reliability, verifiability, and strict guardrails matter more than raw capability.

Overall, AI’s most important macro effect is that it makes the software market bigger. AI is stretching the boundary of what can be productized, capturing new distribution surfaces that weren’t economically reachable before, and converting entire labor pools into addressable “software spend.” However, the tolerance for error in a workflow is still a constraint that will ultimately shape where AI wins. In high-variance, recoverable workflows, probabilistic systems can compound into real leverage; in zero-failure domains, they’ll remain tightly bounded behind verification, controls, and humans.

Ultimately, the AI-winners are the companies that can earn the right strategic position in a workflow by providing new capabilities. But no matter who wins, there’s no denying that the new, AI-enabled capabilities are expanding the overall pie that is the software market.

The unemployment rate printed 10.2% this morning, a 0.3% upside surprise. The market sold off 2% on the number, bringing the cumulative drawdown in the S&P to 38% from its October 2026 highs.

That’s the dystopian future Citrini portrays: a near-term “AI takeover” severe enough to push unemployment into double digits and drag equities into a deep bear market. To be fair, Citrini presents this as a hypothetical scenario rather than a point forecast. But even as a thought experiment, the future presented in Citrini’s memo feels like science fiction. And since this chain of causality starts with AI’s impact on enterprise software, the focus of Points and Platforms, that’s where I think I should start as well.

Software Survival

Citrini’s hypothetical began with:

A competent developer working with Claude Code or Codex could now replicate the core functionality of a mid-market SaaS product in weeks. Not perfectly or with every edge case handled, but well enough that the CIO reviewing a $500k annual renewal started asking the question “what if we just built this ourselves?”

Fiscal years mostly line up with calendar years, so 2026 enterprise spend had been set in Q4 2025, when “agentic AI” was still a buzzword. The mid-year review was the first time procurement teams were making decisions with visibility into what these systems could actually do. Some watched their own internal teams spin up prototypes replicating six-figure SaaS contracts in weeks.

Admittedly, it isn’t pure science fiction. A strong developer armed with tools like Claude Code, Codex, or Cursor can already reproduce plenty of straightforward software solutions, especially point solutions with limited edge cases and low regulatory surface area.

But that’s also where the premise starts to break. Software companies don’t just sell code. They sell an operating bundle around the code: ongoing updates, maintenance, security patches, compliance posture, uptime guarantees, debugging when the weird edge cases inevitably show up, integration support, admin tooling, audit trails, data migrations, roadmaps, and all the unglamorous work of making the product behave reliably inside an enterprise stack.

Enterprise customers can absolutely “spin up prototypes replicating six-figure SaaS contracts.” But they will be just that: prototypes. Getting from a prototype to a real product that the business can depend on requires a dedicated team, months (and potentially years) of iteration, and a never-ending stream of maintenance. That journey doesn’t end when Claude generates the first version of an app; that’s when it starts. And for most companies, none of this is what they’re actually in business to do. To quote the Stratechery article, “Microsoft and Software Survival”:

Companies — particularly American ones — are very good at focusing on their core competency, and for most companies in the world, that isn’t software.

To be fair, Citrini’s thought experiment does nod at this dynamic, but it tries to skirt around it by arguing that enterprises won’t build replacements internally. Instead, they’ll outsource the building to foundation model vendors such as OpenAI and Anthropic via forward-deployed teams that use AI tooling to replicate incumbent SaaS vendors. The article offers this story:

The procurement manager told [the software salesperson] he’d been in conversations with OpenAI about having their “forward deployed engineers” use AI tools to replace the vendor entirely. They renewed at a 30% discount. That was a good outcome, he said. The “long-tail of SaaS”, like Monday.com, Zapier, and Asana, had it much worse.

But zoom out for a second: what’s the practical difference between paying OpenAI to “replace the vendor” and paying Monday.com, Zapier, or Asana in the first place? It’s not as if a forward-deployed team shows up, generates a bespoke clone in a month, and walks away. Turning a prototype into something production-grade is still an ongoing commitment. So the real question isn’t whether enterprises will pay for software. It’s who they’ll pay, and what they get in return.

You could argue the foundation model vendor wins because customers prefer consolidating tools under a single supplier. But consolidation pressure isn’t new. “Vendor rationalization” has been a procurement mantra for a decade, long before AI became the headline. But ultimately, the decisive variable is product quality and ROI. If a tool reliably drives outcomes, it sticks. That’s why a company like Monday.com can still post ~110% net revenue retention even as it sits squarely in the “long tail of SaaS.”

Software Pricing

But there is one real, practical difference for customers once foundation model vendors become credible alternatives, namely the “30% discount.” Basic supply-and-demand says that if foundation model vendors (or AI-enabled services teams) become credible substitutes for certain categories of SaaS, incumbents should face more competitive pressure. That competitive pressure might then force software vendors to give back some pricing power. Citrini does make this point more explicitly:

SaaS wasn’t “dead”. There was still a cost-benefit-analysis to running and supporting in-house builds. But in-house was an option, and that factored into pricing negotiations. Perhaps more importantly, the competitive landscape had changed. AI had made it easier to develop and ship new features, so differentiation collapsed. Incumbents were in a race to the bottom on pricing - a knife-fight with both each other and with the new crop of upstart challengers that popped up. Emboldened by the leap in agentic coding capabilities and with no legacy cost structure to protect, these aggressively took share.

It certainly is possible that building in-house (or, more realistically, outsourcing to forward-deployed teams at foundation model vendors) becomes a credible alternative at the margin, and that might chip away at SaaS pricing power. However, it doesn’t follow that AI will make “differentiation collapse.”

For the system-of-record vendors Citrini argues are also vulnerable, the moat isn’t a prettier UI or a slightly better workflow. It’s the fact that they sit on proprietary data that AI-native challengers simply can’t access at scale. To quote this publication’s eponymous article, “Points and Platforms”:

Because platforms like ServiceNow and Salesforce have a wholistic view on core workflows, over time, they have accumulated a depth of workflow data that AI-native startups and standalone point solutions can’t match. With that data, platforms are often better positioned than AI-native startups and point solutions to train and refine agents that behave reliably inside real enterprise processes.

Without that data, challengers have to compensate with heavier R&D, deeper domain modeling, more integration work, and more time spent closing the gap to production-grade reliability. And that investment has consequences. If an AI-native company is spending meaningfully more to build something truly differentiated, it can’t undercut everyone on price long-term. The economics don’t support a permanent “race to the bottom.” Therefore, price pressure may rise at the edges, but the central contest won’t be a knife-fight over who can charge the least; it will be a knife-fight over who can deliver the better product with the most defensible outcomes.

Software and White-Collar Job Losses

According to Citrini, the downstream consequence of AI-driven disruption in software is job loss:

The interconnected nature of these systems weren’t fully appreciated until this print, either. ServiceNow sold seats. When Fortune 500 clients cut 15% of their workforce, they cancelled 15% of their licenses. The same AI-driven headcount reductions that were boosting margins at their customers were mechanically destroying their own revenue base.

The company that sold workflow automation was being disrupted by better workflow automation, and its response was to cut headcount and use the savings to fund the very technology disrupting it.

There’s no denying that AI is already causing job displacement, and will most likely continue to. Forecasts suggest roughly 6–7% of Americans will face some form of displacement over time. In the memo’s logic, enterprises adopt AI, which reduces the need for human seats, compressing seat-based SaaS revenue. That revenue pressure then forces software vendors to cut costs, primarily by reducing headcount. And those cuts are made possible, ironically, by the vendors themselves shipping AI tools that let them maintain output with fewer people. Once those tools exist, enterprises can also use them to cut headcount, which further reduces seats and compresses revenue, continuing the cycle. As Citrini themselves put it:

AI-threatened companies did the only thing they could. Cut headcount, redeploy the savings into AI tools, use those tools to maintain output with lower costs.

Each company’s individual response was rational. The collective result was catastrophic. Every dollar saved on headcount flowed into AI capability that made the next round of job cuts possible.

However, this mechanism is largely unconvincing. If anything, widespread agent adoption should increase software activity. Agents don’t reduce the number of workflows flowing through ITSM, CRM, HCM, and other software systems; they multiply them. An agent that can execute tasks at machine speed still needs those tasks to be orchestrated, governed, audited, and measured inside platforms like ServiceNow, Zendesk, Salesforce, and their peers.

What likely changes is not demand for software, but pricing. Pure seat-based pricing may fade, but that doesn’t imply shrinking revenue. Vendors can (and increasingly will) monetize usage, outcomes, or agent “seats,” which might be more valuable, and thus expensive, than a human seat because they generate orders of magnitude more throughput.

Therefore, while the job impact is real, they are a function of productivity gains creating surplus capacity in certain roles, rather than collapsing software budgets. That has two important implications:

1. AI adoption isn’t “funded” by a shrinking revenue base in software; it can be funded by rising throughput and value creation.

2. The software economy retains the capacity to reallocate labor. Therefore, if AI is expanding output per worker, the question becomes where people are most valuable next, not whether demand disappears altogether.

This same logic applies across white-collar work. Yes, some people will be displaced as AI boosts productivity, but that doesn’t mean revenues fall. Instead, higher output and faster execution mean more revenue, not less. And if revenues are growing, then companies will still have the capacity to hire. Therefore, the challenge isn’t a lack of demand for labor; it’s the allocation of it. Firms will spend less on certain roles and more on others, and the real question simply is: what will companies actually value? Whatever the answer, it will almost certainly be more interesting than the repetitive, monotonous work that AI is poised to automate away.

A Quick Note on AI Agents and Friction in Society

You might notice I haven’t addressed most of Citrini’s memo, namely the parts that reach beyond software into industries like real estate, payments, and more. Since those industries are not the focus of Points and Platforms, I leave it to other analysts to cover these topics. But there’s one line I do want touch on briefly:

We had overestimated the value of “human relationships”. Turns out that a lot of what people called relationships was simply friction with a friendly face.

I wholeheartedly disagree with this premise. Humans are fundamentally social beings. We don’t live only to know things; we live to share them, argue about them, learn alongside other people, and feel seen in the process.

An AI agent might be able to tell me everything I could ever want to know about the NBA, but that doesn’t replace the 3 AM dorm room GOAT debates. It might explain Plato more clearly than any textbook, but no AI agent can replace sitting in a room with a professor, pushing on ideas, and talking them through with peers. For most of us, we aren’t just looking for the most efficient or optimized solution; we want the human experience around it. If, as Citrini suggests, human relationships are just “friction with a friendly face,” then that friction is not only valuable but also essential to our very being.

That’s why, both in the enterprise and in society as a whole, there will always be space for humans.

From Bloomberg last week:

In the span of two days, hundreds of billions of dollars were wiped off the value of stocks, bonds and loans of companies big and small across Silicon Valley. Software stocks were at the epicenter, plunging so much that the value of those tracked in an iShares ETF has now dropped almost $1 trillion over the past seven days.

On January 30, 2026, Anthropic launched 11 role-specific plug-ins for Claude Cowork aimed at knowledge-work functions such as sales, marketing, finance, and legal. And thus began the “SaaSmageddon”: over the next five days, we saw a massacre of public software companies. The S&P 500 Software & Services index fell around 8%. Salesforce slid around 9% over the week. ServiceNow fell more than 14%.

Frankly, this has been a long time coming for software companies. In 2023, education software giant Chegg lost 48% of its market cap in a single day after warning that ChatGPT was pressuring growth on its Q1 earnings call. In 2024, Klarna launched an AI assistant that handled roughly two-thirds of customer-service chats in its first month, one of the first signals that AI could reduce seats across traditional SaaS. Throughout 2025, sentiment toward software continued to cool as sales cycles stretched and investors grew more anxious that customers would build tools in-house rather than expand SaaS spend. On Sprinklr’s Q1 2026 earnings call, management pointed directly to longer, more cautious deal timelines amid macro uncertainty. And on both Braze’s Q3 FY2026 and monday.com’s Q4 2025 earnings calls, management teams were pressed on whether customers building more tools in-house could push churn higher and weigh on gross retention.

Ultimately, the launch of Anthropic’s Claude Cowork plug-in wasn’t the root cause of the massacre last week. It was simply the catalyst that turned the last two years of skepticism into a full-blown sell-off. Therefore, while many analysts argue the sell-off was an overreaction, a quick rebound may be unlikely. The doubts have been building in the background for years, and it’s unlikely that sentiment will reset.

Unfortunately for public software companies, this is unambiguously bad news. Most legacy vendors have to pivot toward agentic AI offerings. They can build those capabilities through R&D, but that path is often slow and cumbersome. M&A offers a much faster alternative. By buying AI-native companies with proven products and leveraging the incumbent platform’s existing sales motion, customer relationships, and implementation ecosystem to drive upsell and cross-sell, software companies can quickly integrate AI-native solutions without relying on reinventing their R&D around agentic AI.

However, AI-native assets are expensive, and many acquisitions require at least some equity consideration. When public software stocks sell off, that currency weakens immediately. In that sense, the market’s skepticism actively constrains the very M&A that could accelerate the AI pivot. The result is a vicious cycle in which weaker equity currency leads to fewer strategic deals, which in turn slows the AI transition, further eroding investor confidence. Ultimately, it is far more difficult for public software companies to reinvent themselves quickly enough.

Private Markets and Software

This raises a natural question: Is there a way for software companies to execute the AI pivot away from the constant scrutiny of public markets? The clear answer is the private markets. Private equity firms have already started capturing that opportunity. After a post-COVID period of overpaying for assets, many sponsors have faced muted private-market returns as valuations reset while public software multiples have continued to compress. That widening gap set the stage for a surge in take-private activity in 2025, as sponsors step in to buy legacy software companies, reposition them for an AI world, and rebuild outside the public spotlight.

One of the largest enterprise software take-privates last year was Thoma Bravo’s $12.6 billion acquisition of Dayforce. Dayforce is a legacy HCM vendor whose valuation had fallen sharply from its post-COVID peak and continued to slide through 2025 as AI-driven skepticism weighed on the broader software sector. So while Thoma Bravo paid a 32% premium to the unaffected share price ahead of the announcement, the deal was still struck at a discount to Dayforce’s 52-week high and only represented an approximately 7x revenue multiple.

While Thoma Bravo was able to acquire a scaled software franchise with a strong revenue profile at an attractive valuation, Dayforce gets something just as valuable: the environment to pursue its AI transformation away from the quarter-to-quarter pressure of the public markets. Or as CEO David Ossip claimed in Dayforce’s deal announcement:

We are partnering with a truly special organization to accelerate our business - with our focus, resources, and product innovation all laser-pointed on leaping forward as the HCM leader for a world of work shaped by AI.

Growth in Private Markets

It isn’t only legacy software companies that are reaping the benefits of private markets. Many high-growth startups are also staying private longer. Private capital has exploded with deeper and more flexible financing options that can now support large, high-growth companies such as Databricks, Anthropic, SpaceX, and OpenAI.

Take Databricks: it just closed its Series L round, raising $5 billion at a $134 billion valuation, which represents a ~25x revenue multiple on ~65% YoY revenue growth. Meanwhile, Databricks’ closest public comp, Snowflake, saw nearly 20% of its market cap wiped out last week despite not being a pure-play SaaS business.

When you compare the two companies, both are free cash flow positive. But Snowflake is “only” growing revenue at roughly 29% YoY and trades around 12.5x revenue. On a growth-adjusted basis, Databricks is actually cheaper. Its last private round implies a roughly 0.38x growth-adjusted revenue multiple, while Snowflake trades at a ~0.43x growth-adjusted revenue multiple. Databricks also appears to win on retention metrics. Snowflake’s 125% NRR is exceptional for enterprise software, but Databricks’ NRR is above 140%. In other words, Databricks’ financial profile is more than ready for the public markets. However, with abundant late-stage financing, Databricks can remain private, sidestepping public-market volatility and quarter-to-quarter scrutiny, while continuing to develop leading products and push deeper into AI-native workloads.

But the explosion of private capital and the trend of late-stage startups staying private come with real risks. Take Rippling: it’s a standout business and likely a long-term AI winner. However, Rippling's last round reportedly valued the company at $16.8 billion on roughly $570 million in revenue, representing a ~29x revenue multiple on ~30% YoY growth. Although I suspect Rippling’s actual revenue and revenue growth are much higher, publicly available figures imply a roughly 0.96x growth-adjusted revenue multiple, which is higher than almost any large public software company. Given the valuation and Rippling’s ambitions to compound into a platform that can compete with names like ServiceNow and Salesforce, an acquisition is quite unlikely. Thus, for late-stage investors, the most realistic path to liquidity is an IPO. However, an IPO at Rippling’s current multiple would certainly face heavy pushback from public markets. So even though Rippling has massive growth potential and a strong, proprietary data stack for building and hosting AI agents, it still needs time to grow into its valuation. That time delays liquidity for late-stage investors, which compresses IRR.

For Rippling, staying private is a clear advantage while public software companies face sustained skepticism. Private software companies can keep investing aggressively in AI, either through R&D or M&A, without watching their equity currency weaken in real time. But for private-market software investors, last week should be a warning. Private markets may offer a place away from the spotlight of quarterly reports and earnings calls to execute an AI pivot. But eventually, most large private software companies, whether late-stage startups or legacy vendors that were taken private, still have to face the public markets, where investors are expecting a proven, monetizable AI transformation. And for late-stage investors in startups like Rippling, the clock is ticking if they want to sustain an attractive IRR.

In ServiceNow’s Q4 FY2025 Earnings Call, CEO Bill McDermott remarked,

Many people ask why our valuation has not kept pace with our results. The short answer is that we have been viewed as a feature-oriented SaaS company. We are not living in a SaaS neighborhood. We are a platform company, executing a long-term platform strategy where AI agents and workflows are harmonious and synonymous, creating sustained advantage, not short-term wins.

The term “feature-oriented SaaS company” can probably describe most software companies. Historically, even companies that ultimately became platforms often won customers through a flagship product that offered a best-in-breed solution for a single high-value workflow. Salesforce is perhaps the most classic example of this. Salesforce landed customers through its CRM product, Sales Cloud, and then expanded across adjacent clouds and capabilities. Even ServiceNow has traditionally followed a similar sales motion. Many customers first adopted ServiceNow through its ITSM software and then expanded usage to other workflows across IT, employee, and customer operations. You’ll notice that in both cases, Salesforce and ServiceNow initially won customers because they offered a single best-in-class point solution that delivered clear, immediate value. Importantly, customers did not buy Salesforce or ServiceNow just because they were a platform. In that way, both Salesforce and ServiceNow were fundamentally feature- and product-led rather than platform-oriented.

That’s not to say platforms, despite being product-led, don’t have real structural advantages over point solutions. Take Zoom, a point solution. Most people think of Zoom today as a low-growth company; after revenue surged more than 300% during COVID, growth has cooled to around ~3% YoY, largely reflecting the normalization of work patterns. But Zoom has another indicator of slowing momentum that isn’t fully explained by the shift back in person: net dollar retention.

According to Zoom’s Q3 FY2026 Earnings Deck, Zoom has a TTM NDR of 98% for its enterprise customers, meaning the average enterprise cohort is shrinking year over year. While it’s understandable that the drop in demand has led to slower revenue growth, it’s less convincing that the shift back to in-person is solely responsible for the lackluster retention rate. Even in a post-COVID environment, organizations still rely on video conferencing for core functions such as cross-office collaboration, customer calls, recruiting, and meetings with external partners. Therefore, since video conferencing remains a durable, must-have utility, you’d expect existing customers to at least hold steady even if growth has stalled.

The lackluster net retention rate is probably best explained by competitive pressure from Microsoft Teams and Google Meet. Even if, in my experience, Teams and Meet are meaningfully worse products on pure video quality and usability, large enterprises don’t buy video conferencing in a vacuum. Because Teams is bundled into Microsoft 365 and Meet is bundled into Google Workspace, Microsoft and Google can sell them at a discount. Therefore, both products only have to be “good enough” to convince procurement teams to standardize and cut a standalone Zoom contract, both saving money and reducing vendor sprawl. In this environment, Zoom can only keep a foothold for premium use cases, but it’s harder to expand within existing accounts. And some customers will rationalize seats or switch entirely simply because the all-in economics favor the bundle, ultimately leading to Zoom’s low NDR.

Software Platforms and AI

Bundling and the pricing leverage it creates are still relevant for software platforms in the AI era, but that isn’t the “sustained advantage” Bill McDermott emphasized on ServiceNow’s earnings call. Instead, he argues the durable edge comes from being a platform where “AI agents and workflows are harmonious and synonymous.” Put simply, McDermott’s claim is that the strongest position in enterprise software is to be the system that hosts AI tools and embeds them directly into core workflows. Of course, that conclusion itself seems quite obvious. If AI is truly transformative for enterprise tech, naturally, you would want to be the platform where AI runs and work gets executed. But why is McDermott so confident that ServiceNow is best positioned to win that role? Or, more generally, what allows the enterprise software platforms of today to defend their positions in a post-AI world?

The overarching advantage software platforms have is that they’re already deeply embedded in some of their customers’ most critical workflows. Take ServiceNow for example. Because its product suite spans incidents, requests, assets, and other core IT processes, a large share of an organization’s day-to-day IT work flows through the platform. This gives ServiceNow a unified and coherent view of how a customer’s IT operations actually run, something an AI-native startup or a legacy point solution typically can’t replicate without years of deployment depth and integration.

This visibility gives legacy platforms a major advantage: data. Because platforms like ServiceNow and Salesforce have a wholistic view on core workflows, over time, they have accumulated a depth of workflow data that AI-native startups and standalone point solutions can’t match. With that data, platforms are often better positioned than AI-native startups and point solutions to train and refine agents that behave reliably inside real enterprise processes.

Ironically, one of the most explicit examples of this platform advantage comes not from an incumbent, but from a late-stage startup: Rippling. In Rippling’s 2024 Investor Memo, CEO Parker Conrad claimed that,

The corollary opportunity for Rippling, then, is to rebuild business software across each software vertical, but to embed an understanding of your company’s employees in the foundations and tissue of each of these products.

Rippling’s core thesis is that its products are built around a shared layer of employee data, namely Rippling’s Employee Graph. Therefore, every application in the suite benefits from the same unified system of record. That gives Rippling an advantage over point solutions that only see a narrow slice of the customer and thus cannot build the same cross-functional context into their products. Just as importantly, Rippling’s point-solution competitors struggle to justify comparable investments in their data infrastructure. For Rippling, that infrastructure is amortized across the entire suite, whereas for a single-product vendor, that same infrastructure can only improve one workflow, making the ROI on building a similarly deep data layer far less attractive.

This advantage arguably applies even more strongly against AI-native point-solution startups. Rippling can leverage the proprietary data from its Employee Graph to train and improve its agentic AI features, because those agents operate inside a system that already knows who employees are, what roles they have, what they’re allowed to do, and how workflows connect across functions. By contrast, AI-native point solutions typically have thinner access to employee and workflow data and even less incentive to build a deep, unified data layer if it only improves a single product. The result is that their agents often have less context, weaker permissions awareness, and fewer reliable integration points, which can translate into a less capable enterprise agent.

You might notice there is another glaring advantage to being a platform that I have yet to mention: the switching costs platforms impose on their customers. And in an AI-driven market, that advantage is hard to overstate. A standalone point solution is relatively easy to replace. You can easily pivot to an AI-native alternative in one team, migrate a narrow dataset, retrain users, and swap one workflow with limited disruption. But ripping out a platform is much more difficult. Let us return to the example of ServiceNow. ServiceNow’s platform is woven into the fabric of their customers’ IT operations. Replacing that entire system means rebuilding integrations across dozens of systems, migrating years of historical data, and retraining multiple teams that depend on it. That depth of entrenchment dramatically raises the bar for any AI-native challenger hoping to displace the platform end-to-end. As a result, if platforms like ServiceNow, Salesforce, and Rippling can embed AI into workflows quickly and credibly, it becomes much harder for new point solutions to gain a foothold, and far less likely that enterprises will rip and replace a system that already sits at the center of their core business workflows.

Notice that all of the advantages above stem from the platform itself rather than any single product. The ability to act as a system of record and leverage workflow data across modules is what a software company actually needs to survive and thrive in a post-AI world. And those advantages can only accrue to companies that operate as true platforms, not standalone point solutions. Fundamentally, this is what Bill McDermott means when he says that ServiceNow’s advantage comes from being a “platform company.”

AI-Native Startups

Of course, it’s important not to ignore the flip side of this dynamic. Large software platforms are often slow to pivot to new business models. They are constrained by massive installed bases, legacy architectures, and the need to maintain reliability, security, and governance across thousands of enterprise deployments. That friction creates an opening for more agile, AI-native startups that are able to iterate faster and ship new workflows quickly to capture value before incumbents fully adapt.

We saw a similar dynamic during the early 2010s as software companies rushed towards the cloud. As Parker Conrad pointed out:

Because [business software vendors like SAP, Oracle, and Microsoft] were poorly positioned to rebuild their technology as cloud software, the shift to the cloud created a moment-in-time opportunity for focused competitors to peel off specific features and products from these mega-vendors and turn them into standalone, point-SaaS companies.

With AI, there is once again a real “moment-in-time” opening for AI-native startups to ship standalone products that can compete with established platforms by simply being better at the AI layer. And it’s hard to deny that some of these companies are breaking out. Serval, for example, is positioning itself as an AI-native IT service management platform and has landed credible early customers such as Perplexity AI, Clay, Mercor, and Notion. It raised a $75M Series B led by Sequoia Capital at a reported $1B valuation, representing over 110x Serval’s ARR on 500% revenue growth since August. This signals of both rapid adoption and investor conviction even while it competes in territory long dominated by ServiceNow.

While AI-native players like Serval are outpacing incumbents in building the agentic layer, incumbent platforms like ServiceNow and Salesforce still own two assets that are nearly impossible to replicate: the data layer, namely the system-of-record visibility into real workflows, and the distribution channel. In that world, the most important thing for platforms to emphasize isn’t any single feature. Instead, it’s having the strongest data stack and workflow visibility, because that’s what lets software companies seamlessly integrate any agentic solution they build or acquire. Then the incumbent’s existing sales engine can immediately drive cross-sell and upsell into its installed base of enterprise customers. Thus, in the AI era, the defensible advantage shifts from the product to the platform itself.

Back in 2011, Marc Andreessen said, “Software is eating the world.” And now, almost 15 years later, enterprise software is still not full; it’s just moving to a new course: AI.