Enterprise software differs sharply from B2C services such as Netflix, TikTok and Amazon. Services such as Netflix, Meta, Amazon, TikTok and Google have used every action users take — clicking, moving on, pausing and leaving — to capture and analyse behaviour and improve their systems.

According to Foundation Capital partner Jaya Gupta (자야 굽타), the essence of the competitive edge those services have built over the past more than 20 years can be summed up in one phrase: a data compounding loop.

It is a virtuous cycle in which user behaviour data improves the system, and the improved system draws in more users, generating more data. The better this structure works, the harder it becomes for competitors to catch up.

Enterprise software, by contrast, could not use such a loop. It was not because there were few decisions. It was because observing them was difficult, Gupta said, unlike in B2C services.

In a recent post shared on social media platform X (Twitter), Gupta drew attention by stressing that with the rise of LLMs, a path is opening for enterprise software to use a data compounding loop as a competitive advantage.

"A new weapon for enterprise software companies to raise value is a compounding loop that enterprise software never had," he said. "If consumer platforms learned clicks and scrolls, in enterprise software the data compounding loop learns corporate decision-making processes." He said decision-making is beginning to leave digital traces as remote work and asynchronous collaboration become routine, including comments, document revision histories, tickets, approval records and call recordings.

B2C services involve a single user acting within one interface, he said. Corporate decision-making is different. Sales, finance, legal, operations, security and executives push and pull with different interests and authority. Decisions are made through negotiation, not clicks. Until now, companies had no way to measure the reasoning process linking actions and outcomes.

Enterprise systems are designed to record the final state, so they do not show how final numbers are reached. A revised contract shows final clauses, but it is hard to know which alternatives were rejected. The context of decisions was scattered across meeting rooms, someone’s head, email threads, small talk and unconnected systems. There was little reason to store such data. Decision data was treated as a byproduct of process.

But the situation is changing. A flow is emerging that allows decision data to be stored and tracked.

Gupta cited 3 reasons: work has moved into recordable spaces, LLMs have made it possible to compute unstructured data, and agents can automatically record decisions. "As remote and asynchronous work becomes routine, traces related to decisions accumulate in comments, document suggestions, ticket histories, approval flows and call recordings, and LLMs can extract decision traces from transcripts, chat logs and document comment data," he said.

On agents, he said, "If an agent makes a pricing proposal, a sales representative changes the discount rate from 25 percent to 30 percent and adds a memo saying 'Need to respond to competitor X.' That edit is a decision trace. The model’s proposal is a structured prior value the system deems correct, and a person’s change is a judgement signal the model missed."

B2B SaaS companies such as Salesforce, ServiceNow and Workday are not well positioned to create such an environment on their existing platforms, Gupta said.

They are adding agents on top of their existing platforms, but those agents inherit an architecture that stores only the current state. "Once a discount is approved, the context disappears," he said. "Because you cannot recreate the state at the time of the decision, you cannot audit, learn from, or use precedents for the decision. Data warehouses such as Snowflake and Databricks also receive post-decision data through ETL (Extract, Transform, Load). There are only outputs of decisions, and no reasoning leading to those outputs," he said.

By contrast, agent-based startups starting from scratch are in a relatively advantageous position to build systems that can learn from and analyse decisions, he said. "Because agents execute workflows, they can capture context at the moment a decision is finalised. It is stored as an official record in real time, not after ETL," he said.

That can sharply raise the quality of what AI can do.

"Law, insurance, healthcare, finance, procurement, security. Every industry has decades of accumulated expert judgement," he said. "They have never been documented, learned, or organised into a usable form. That is why experts are paid $2,000 per hour." He added, "Top expert levels come from experience and judgement accumulated by organisations over decades. Now we can capture, structure and learn from this data."