The system of record is back: why AI opens the door to new B2B software giants

Last week the shares of large software companies rebounded strongly, erasing (and exceeding) the losses suffered during the so-called "SaaSpocalypse," the sell-off that had hit the sector months earlier on fears that traditional software would be replaced by homegrown AI-built alternatives ("vibe-coded"). According to the authors of this Battery Ventures analysis, Aaron Neil and Michael Brown, that debate over whether companies can use AI internally to stand up their own mission-critical applications overshadows a far more important phenomenon: AI is driving a new generation of "AI-first" software vendors capable of attacking enormous markets that until now looked impregnable.

The article revisits the debate that dominated the tech conversation last year: whether the "system of record" (the software that stores and manages a company's core data, such as CRM, ERP or HCM) was doomed to disappear. Microsoft's Satya Nadella went so far as to say that traditional SaaS applications would be reduced to mere "dumb databases." The conventional wisdom at the time held that those systems of record were too large and too entrenched to be replaced, and that the real opportunity for startups lay in building a lighter "system of action" operating on top of them, at the user interaction layer.

The authors argue that reality is turning out to be different and more nuanced: in many sectors and workflows, this would be precisely the best moment in a long time to build an entirely new system of record from scratch.

**The missing link in AI adoption: data integrity**

The central argument starts from a field observation: outside Silicon Valley, both "white-collar" professionals (accountants, lawyers, doctors) and "blue-collar" workers (contractors, logistics, retail) are far more eager to adopt AI than is generally acknowledged. But there is an obstacle that Silicon Valley underestimates: data integrity. It is not enough to point an AI agent at a company's data and expect it to work; for an agent to actually perform the work people do today, it needs the critical context that comes from domain expertise in a role, a workflow or an industry, and that context is usually buried in fragmented data repositories and antiquated systems from which quality information is hard to extract.

This question of context is forcing many companies to rethink their core software systems, because they cannot deploy AI effectively while their data remains trapped in obsolete systems. Many technology buyers conclude that, to adopt AI seriously, they need a new system of record. That redraws the map of where the next great software companies should be built: precisely in the category —systems of record— that had been off-limits to new entrants for a decade.

**The "TAM trap"**

The article explains why markets such as ERP, HCM or CRM, despite representing billions in software spend, have for years been less attractive than their size (TAM) suggests. The key is the gross retention rate: if a market retains 90% of its customers, by definition only one in ten switches vendors each year, implying a decade-long replacement cycle. So a $5 billion market with 90% retention only puts $500 million of ARR "in play" each year. If a challenger sees only 60% of those opportunities and wins 33% of the ones it sees, the ceiling on incremental growth comes to about $100 million a year: a company ("Business A") reaching $500 million in ARR in that context would grow only 20% annually. The formula they propose is: Gross added ARR potential = TAM x (1 – Gross retention) x % of deals seen x % win rate.

**When markets move: AI breaks the inertia**

According to the authors, they are now seeing markets previously locked up by high retention rates open again, because companies want to adopt AI and often cannot do so within their existing systems of record. There are even markets skipping the SaaS era entirely and moving straight from on-premise to AI, because the cloud was never reason enough to change systems of record, but AI is. Moreover, AI reduces switching friction, because migrating data to a new system has never been so simple.

Back to the example: the same $5 billion market, but now with four switching opportunities a year instead of one (60% gross retention instead of 90%). Using the same formula, "Business B" would get $5B x 40% x 60% x 33% = $396 million of new ARR per year, which means growing close to 80% at $500 million in ARR, roughly four times faster than Business A at the same scale. The authors say they have observed this phenomenon —AI-based systems of record displacing deeply entrenched competitors— within Battery's own portfolio, in both horizontal and vertical markets.

**What today's systems of record look like: the TAM grows too**

Beyond the churn rate, the other factor changing is the size of the market itself, which expands as these systems perform tasks previously done by people. Owning the data lets a system of record automate what the "systems of action" school only theorized about: executing real tasks in order to monetize inference or outcomes. For example, an agent could close month-end accounting accruals instead of an accountant, or replenish inventory before the store owner checks the shelves. When a task's inputs, its execution and its outcome all live in the same system, that system captures exactly what AI needs to keep learning the more the product is used: in the short term, better context engineering; over time, improvements at the level of the model weights themselves. This makes it possible to expand the TAM by shifting from selling "seats" to selling "tokens," moving captured value from data ownership toward outcome delivery (something the authors consider complementary, not mutually exclusive).

Using the same math but doubling the TAM by moving from software to human labor: $10B x 40% x 60% x 33% = $800 million. This "Business C" could grow 80% at a scale of $1 billion in ARR, keeping the same growth rate Business B had at half the size. AI-based systems of record, they conclude, can literally "defy gravity" compared with businesses trapped in locked-up TAMs.

**Entirely new ("greenfield") markets**

The article introduces an important nuance: there are processes and workflows that simply did not exist before AI, such as natural-language software engineering (they cite Cursor and Cognition), the deconstruction of text and unstructured data in a single instruction (they cite Harvey and Legora), or the analysis of engagement with sales prospects (they cite Gong, a Battery portfolio company). In these cases there are no incumbents to displace, so the concept of market churn does not apply; what matters is the speed at which the market itself expands, which today usually stems from improvements at the model layer that let agents complete longer, more complex tasks. The formula is adjusted accordingly: Gross added ARR potential = TAM x Market growth factor x % of deals seen x % win rate.

With a hypothetical example, "Company D," operating in a market also worth $10 billion but doubling every year as task horizons expand: $10B x 2x x 60% x 33% = $4 billion of new gross ARR per year, a figure that —the authors say— "would break a lot of models."

**Why these new companies are also "systems of record"**

According to the analysis, these AI-native companies are, at bottom, systems of record because they are embedded in their users' workflows and collect highly valuable data: Cursor/Cognition capture rules files, indexed embeddings, learned conventions and memory of the agent's past decisions; Harvey/Legora accumulate firm-specific playbooks, precedent libraries, matter histories, clause-level drafting patterns, and the corrections lawyers accept or reject on AI-generated drafts; Gong gathers call recordings and transcripts, objection and sales-pitch patterns, coaching benchmarks and conversation data tied to deal outcomes.

This gives them a structural advantage: they own both the context agents need at runtime and the outcome data that serves as the reward signal for training, because actions and their results live in the same system. That combination compounds with use: workflows generate action trajectories whose outcomes are observable within the same system, post-training raw material that competitors integrating only via API cannot capture. As the AI application layer integrates vertically into the model layer (thanks to open-weight models and increasingly mature evaluation and reinforcement tooling), the AI system of record can turn that data into weight-level improvements for a specific task, generating durable differentiation.

The authors also note that, just as AI applications today depend mostly on a limited number of external systems, as they gain power and influence over their users' workflows they will end up pushing and pulling data from more adjacent software systems, just as traditional systems of record once did. They cite Clay as an early example: it began by pulling contact and company data from a handful of data enrichment providers, and today already ingests information from its customers' entire go-to-market stack —CRMs such as Salesforce and HubSpot, sales tools such as Outreach and Salesloft, product usage and intent signals, and more than a hundred third-party data providers— sending enriched records, scores and AI-generated research back into the customer's CRM. In this way, Clay is becoming one of the software tools sales teams use every day.

**No absolutes**

The authors qualify that not every sector will need or want an entirely new AI-based system of record; they explicitly mention extremely complex verticals, such as insurance, where they see more difficulty replicating the phenomenon. They do, however, see great potential in horizontal markets such as sales, engineering and finance, and in vertical industries such as manufacturing, logistics and construction.

**The window is open now**

The article concludes that the future may look more like the past than the current tech narrative suggests: creating lasting value still rests on compelling customer experiences and on competitive differentiation that compounds with scale, and the history of software shows that data ownership is a key enabler of both. What has changed is what software can do: it can now perform the work itself, thanks to AI. That is why, the authors say, the systems of record of the AI era will look like very different businesses from those of the previous generation, but they will share the same foundation on which every great software business has been built. The text closes with a warning to founders and investors: we are at a unique moment in which customers are curious about AI but largely underserved by their current software vendors, and that window will not stay open forever.

The article itself includes a disclaimer: the views are the authors' own, are for informational and educational purposes, do not constitute investment, legal or tax advice, nor an offer to sell or a solicitation to buy an interest in any fund managed by Battery Ventures, and the anecdotal examples cited are not recommendations or endorsements of any particular business; it further clarifies that Gong is a Battery portfolio company.