Last week major software stocks rallied, more than erasing their losses from the “SaaSpocalypse” selloff earlier this year. It finally seems fears that traditional software will get replaced by vibe-coded alternatives have abated. Unfortunately, this debate around the viability of companies using AI internally to build critical, B2B software applications misses the larger point of the actual impact AI is having on software—namely, that it is ushering in a new age of innovative, AI-first software vendors attacking large markets.

Just rewind to last year, when the debate du jour in tech was whether the software “system of record” was dead. Microsoft’s Satya Nadella famously claimed traditional SaaS apps would be relegated to dumb databases. There was blog post after blog post about how the new opportunity for software was to become a lighter-weight “system of action” that would sit on top of big, traditional systems of record, like customer-relationship management (CRM), enterprise-resource planning (ERP), and human-capital management (HCM) products. The conventional wisdom was that these huge software systems are too difficult to replace and wouldn’t capture much value in world of AI anyway—so startups should just focus on building at the engagement layer.

We think something more nuanced and very different is happening as the AI wave washes over the B2B tech landscape. In many industries and workflows, we believe now is the best time in a very, very long time to build a robust, brand-new system of record. Here’s why.

Missing Link in AI Proliferation

It is no secret that everyone wants to adopt AI today. But in our experience, the desire from folks outside of Silicon Valley—people working in professions other than technology—to adopt AI is much stronger than people in Silicon Valley give them credit for. This is true in both B2B and consumer use cases, but our focus here is just on B2B.

However, when you actually talk to these AI-curious people, both “white collar” professionals (accountants, lawyers, doctors) and “blue collar” folks (contractors, logistics, retailers), we’ve found there is one big issue preventing their widespread and successful use of AI that is often underdiscussed in Silicon Valley: data integrity.

Simply put, you cannot just point an AI agent at your data and expect it to work. To have an agent that can actually do the work humans do, that agent needs the critical context that comes from domain experience (of a job, workflow, industry, etc.) that’s hidden deep within a business’s data repositories. The challenge is that business data is generally fragmented across multiple systems, many of which are extremely clunky. So it’s often hard for harnesses to extract quality data from these systems to make an agent useful.

It is this exact context issue, interestingly, that is forcing people to reconsider their core business software systems, because they can’t effectively deploy AI with data trapped inside their software. To adopt AI, many technology buyers believe they need a new system of record. This mindset is important because it re-writes the narrative around where new AI businesses should be built, which is around a category of software—that is, systems of record—that has long been off limits for new startup entrants. There simply haven’t been many compelling reasons for businesses to switch out their systems of record since the shift from on-premise to cloud software years ago.

The TAM Trap

The problem this has caused for founders and investors alike over the last decade is that many of these markets (such as ERP, HCM, CRM) represent billions in software spending, so they’re tempting for new startup entrants to target. However, historically, only a tiny fraction of that ARR actually changes hands from one vendor to another in these huge markets each year.

For example, if the gross retention of a market is 90%, by definition only one in 10 buyers switches vendors per year, meaning that market is on a 10-year replacement cycle. The implication is that even a $5 billion market, as an example, may not be that attractive: At 90% gross retention, that means only $500 million of “jump ball” ARR comes up each year for new vendors to capture.

Further, if you factor in how many of those jump balls you, as a challenger vendor, actually see (perhaps 60%) and win (maybe 33%), suddenly the ceiling on your growth becomes around $100 million of additional ARR per year. This concept of market turnover makes it really difficult to compound growth at scale no matter how large the absolute TAM of a particular market might seem. In this example, the business we’re talking about—let’s call it Business A—would be growing only 20% at $500 million ARR scale (assuming no churn).

Ultimately, the amount of gross ARR a software business can add per year, and therefore its growth rate is constrained by this equation:

Gross ARR Add Potential = TAM x (1 – Gross Retention) x % Market Deals Seen x % Win Rate

Moving Markets

In today’s world, we’re seeing software markets that had been locked up from high gross retention rates break back open as enterprises are yearning to adopt AI but often can’t do it within their existing systems of record.

In fact, we’re even seeing some markets skip the SaaS era entirely and go straight from on-prem to AI because cloud wasn’t a good enough reason to switch systems of record… but AI is! Moreover, AI also decreases the friction traditionally associated with technology switching costs because migrating data into a new system has never been easier.

Going back to our example above: Take that same $5 billion market, but now, instead of one jump ball per year, imagine there are four (i.e., market gross retention falls to 60%). Using the same math but for Business B:

$5B x 40% x 60% x 33% = $396M

Business B is adding almost $400M of ARR a year, meaning the business is growing around 80% at a scale of $500 million in ARR, or roughly 4x as fast as Business A at the same size.

Within the Battery portfolio this year, we’ve seen this phenomenon of AI systems of record displacing deeply entrenched incumbents across both horizontal and vertical markets.

What Today’s Systems of Record Look Like

Importantly, market turnover rates aren’t the only variable changing the market today: The more impactful variable is TAM, which is growing as systems of record perform more tasks that were previously performed by humans

In short, owning data enables a system of record to automate what the “systems of action” camp theorizes: executing actual tasks to monetize inference and/or outcomes. For example, an agent could run month-end accruals instead of an accountant, or reorder inventory before the shop owner ever checks the shelf.

When a task’s inputs, its execution, and its outcome all live in one system, that system captures exactly what AI needs to continuously learn the more the product is used. In the near term, that shows up as better context engineering so agents complete more tasks correctly. Over time, it can be converted into weight-level improvements to train the underlying models themselves, so agents become smarter. This enables TAM expansion from selling seats to tokens, reflecting the idea that value captured over time shifts from owning data to delivering outcomes, which is not mutually exclusive in AI systems-of-record; the first is what enables the second.

Using the same math, but doubling the TAM from software to labor:

$10B x 40% x 60% x 33% = $800M

The result is that this business, Business C, can grow 80% at a $1 billion ARR scale—sustaining the same growth rate Business B had at double the scale! Relative to businesses operating in locked-up TAMs, AI systems of record can truly defy gravity.

Greenfield Markets

The caveat is that in a world with AI, there are processes and workflows that fundamentally did not exist before. Think: natural language software engineering (Cursor/Cognition), single-instruction deconstruction of unstructured text and data (Harvey/Legora), or prospect engagement analysis (Gong*).

That said, this system-of-record framework still applies, but there is a slight adjustment to the annual gross ARR add formula:

Gross ARR Add Potential = TAM x Market Growth Factor x % Market Deals Seen x % Win Rate

Because there are no incumbents to displace in these markets, the concept of market turnover does not exist, so the market’s gross retention is irrelevant. Instead, what is more important is how fast the underlying market is expanding. This can be due to several factors. But today, it commonly stems from improvements at the model layer that enable agents to complete longer and more complex tasks.

For sake of completeness, consider Company D operating in a comparable, $10 billion market. But the market is doubling each year in tandem with expanding task-completion time horizons:

$10B x 2x x 60% x 33% = $4B

In this analysis, Company D can add $4 billion of gross new ARR per year. At this point, you can do the growth-rate math yourself. Safe to say, it would break many models!

Finally, we would assert that these new, AI-native businesses, too, are really systems of record. Why? First, they are embedded in user workflows which enables them to collect the following important data:

• Cursor/Cognition— rules files, indexed embeddings, learned conventions, agent memory of past decisions.
• Harvey/Legora—firm-specific playbooks and precedent libraries, matter histories and work product, clause-level redline patterns, lawyer accept/reject edits on AI-generated drafts, agentic workflow templates.
• Gong*—call recordings and transcripts, objection and talk-track patterns, coaching benchmarks, conversation data linked to deal outcomes.

As such, the AI system of record in the market today has a structural advantage: It owns both the context that data agents need at runtime and the outcome data that serves as reward signal for training, all because the actions and their results live in the same system. That pairing compounds with usage.

Workflows generate action trajectories whose outcomes are observable in the same system, which is the raw material for post-training that competitors integrating via API can’t capture. As the AI app layer increasingly vertically integrates back into the model layer thanks to open- weights models and maturing eval/RL-tooling, the AI system of record can convert that data into weight-level improvements for a given task, which drives durable differentiation.

Lastly, while AI apps may primarily ride on top of a limited number of other systems today, as they gain more power and influence on user workflows, the AI systems of record will inevitably push and pull data into more adjacent software systems, just like traditional systems of record. Clay is already an early example: The company started primarily by pulling contact and company data from a handful of enrichment providers. But Clay now ingests data from across a customer’s entire GTM stack—CRM (Salesforce, HubSpot), sales engagement (Outreach, Salesloft), product-usage and intent signals, and 100-plus third-party data providers—and pushes enriched records, scores and AI-generated research back into a customer’s CRM. By doing this, Clay is becoming one of the primary software tools GTM teams use daily.

There Are No Absolutes

Of course, there are no absolutes; not every industry or process will need or desire a brand-new, AI-native system of record. We’re thinking here about extremely complex verticals like insurance, for example. On the other hand, there are systems of record in some industries that will be easier to replicate than others. We see great potential in horizontal markets like sales, engineering, and finance, as well as vertical industries like manufacturing, logistics, and construction.

The Window is Now

In summary, we think the future might look more like the past than the current tech narrative is giving this technology cycle credit for. The truisms of business haven’t changed: Durable value is built through delightful customer experiences and competitive differentiation that compounds with scale. The history of software would argue that data ownership is an important enabler to both. What has changed is what software can do, which is now the work itself, thanks to AI. As a result, systems of record in the age of AI will look like much different businesses than in the prior generation—but that doesn’t mean they won’t share the foundation of what every great software business has been built upon.

The opportunity to build a new system of record is now. We are at a unique moment in time where customers are AI curious but largely not served by their existing software providers. That window won’t last forever.