The next trillion-dollar AI track, the context graph, is the real opportunity for AI entrepreneurship.

In Silicon Valley, the debate around a question is heating up:

Will AI, especially Agents, replace SaaS?

The first to give a clear judgment was Jamin Ball, a well - known columnist in the SaaS field.

In his article "Long Live Systems of Record", he bluntly opposed the claim that "Agents will kill all old systems".

In Ball's view, the more powerful the Agent, the higher the requirement for the accuracy of underlying data. Therefore, as the "gatekeeper" of data, the traditional Systems of Record not only haven't lost their barriers but have become more valuable due to their control over the right of interpretation.

However, Jaya Gupta, a partner at Foundation Capital, believes that Ball only sees one side of the coin.

In her latest article "The Trillion - Dollar Opportunity of Artificial Intelligence: Context Graph", she points out that the blind spot of traditional systems lies not in "data" but in the lack of "context".

The real operating logic of an enterprise is often not recorded in the standardized forms of CRM. Instead, it is hidden in exceptional approvals, temporary adjustments, and cross - departmental Slack communications.

Gupta defines these implicit processes as "Decision Traces".

When these decision traces are continuously recorded and connected between time and business objects, a new structure - the Context Graph - will be formed.

This is not just a pile of data but a reproduction of the enterprise's "reasoning process". The opportunity for the next trillion - dollar platform is not to install AI on old systems. Instead, it lies in who can seize the gray area between "data" and "actions". This is the real opportunity that AI startups need to seize.

Today, let's break down the core logic of this super - track.

01 Context Graph: The Most Valuable "Second Asset" of Enterprises in the AI Era

The previous generation of enterprise software created a trillion - dollar ecosystem by becoming "Systems of Record (SoR)". Salesforce manages customer data, Workday manages employee data, and SAP manages operational data.

Their logic is: control the authoritative data, master the workflow, and thus achieve customer lock - in.

The current focus of the debate is: Can these old systems survive the transformation to Agents (artificial agents)?

Jamin Ball's recent article "Long Live Systems of Record" has touched many people's nerves.

He refuted the argument that "Agents will kill everything" and believes that Agents will not replace the Systems of Record. Instead, they will raise the standard for an excellent System of Record.

This view is correct. Agents are cross - system and action - oriented. The user experience (UX) of work is being separated from the underlying data layer. Agents have become the interaction interface, but the underlying layer still needs something authoritative to support it.

However, it should be added that Ball's view assumes that the data required by Agents already exists somewhere, and Agents only need better access rights, better governance, semantic contracts, and clear rules.

This is only half of the picture. The other half is the currently missing layer that truly drives the operation of the enterprise: Decision Traces.

These decision traces include exceptions, overrides, previous cases, and cross - system context. Currently, they are scattered in Slack discussions, Deal Desk conversations, escalation conference calls, and people's minds.

This leads to a crucial difference:

Rules tell Agents what should happen in general cases (for example: "Use official ARR data for reporting").

Decision Traces record what actually happened in specific cases (for example: "We used the X definition, according to Policy v3.2, with special approval from the VP, based on the Z precedent, and we made the following modifications...").

Agents not only need rules but also need access to Decision Traces to understand how rules were executed in the past, where exceptions were granted, how conflicts were resolved, who approved what, and which precedents actually dominated reality.

This is where startups in the "Agent system" category have a structural advantage.

They are on the execution path. At the moment when a decision is made, they can see the whole picture: what inputs were collected from various systems, what policies were evaluated, what exception paths were invoked, who gave the approval, and what state was written.

If you persistently save these traces, you get something that most enterprises don't have today: a queryable record of how decisions were made.

We call the structure formed by accumulating these traces the Context Graph: It is not the "Chain - of - Thought" of a model but a living record that connects decision traces across entities and time, making "precedents" searchable.

Over time, this Context Graph will become the real Source of Truth for automation - because it not only explains what happened but also why it happened.

The core question is not whether the existing Systems of Record will survive. Instead, it is whether brand - new systems will emerge, not just systems for recording objects but systems for recording decisions, and whether these systems will become the next trillion - dollar platform.

02 What the Systems of Record Failed to Capture

When Agents are being deployed in real - world workflows, such as contract review, Quote - to - Cash, and customer service solutions, teams often hit a "wall" first.

This wall is not the lack of data but the lack of Decision Traces. Agents encounter the ambiguity problems that humans solve every day with judgment and organizational memory. However, the input information for these judgments has not been stored as a persistent asset. Specifically:

1. Exceptional logic existing in people's minds. "We always give healthcare companies an additional 10% discount because their procurement cycle is so tough." This statement is not in the CRM (Customer Relationship Management system). It is "Tribal Knowledge" passed on through onboarding training and private conversations.

2. Precedents of past decisions. "We designed a similar deal structure for Company X last quarter - we should keep it consistent." No system links these two deals, and there is no record of why this structure was chosen.

3. Comprehensive judgment across systems. A customer service supervisor checks a customer's ARR (Annual Recurring Revenue) in Salesforce, sees two unresolved escalated complaints in Zendesk, reads a Slack message marking the risk of churn, and then decides to escalate the handling. This comprehensive judgment happens in his mind. And the work order only says: "Escalated to Tier 3".

4. Approval chains outside the system. A VP approves a discount in a Zoom call or a Slack private message. The Opportunity Record only shows the final price, not who approved the deviation or why.

This is what "never captured" means. It's not that the data is dirty or isolated but that the reasoning process connecting data and actions has never been treated as data.

03 Permanently Preserve the "Decision - Making Scene"

When startups deploy at the Agent Orchestration Layer and let each run generate Decision Traces, they get something that enterprises almost never have:

A structured, replayable history that records how context is transformed into action.

What does this look like in practice?

A renewal Agent proposes to give a 20% discount. Company policy stipulates that the maximum renewal discount is 10% unless a "Service Impact Exception" is approved.

The Agent pulls three SEV - 1 incidents from PagerDuty (operations monitoring), an unresolved "Cancel if not fixed" escalated complaint from Zendesk, and retrieves the renewal communication record of a similar exception approved by a VP last quarter.

It submits the special application to the finance department, and the finance department approves it. Finally, only a result remains in the CRM: "20% discount".

Once you have a decision record, "why" becomes the top - priority data.

Over time, these records naturally form a Context Graph: The existing entities in the enterprise (accounts, renewals, work orders, incidents, policies, approvers, Agent run records) are connected through decision events (critical moments) and "why" links.

The company can now audit and debug the automation process and turn exceptions into cases instead of relearning the same edge cases in Slack every quarter.

The feedback loop is the key to generating compound interest effects. The captured Decision Traces become searchable precedents. Each automated decision adds a new trace to the graph.

All of this doesn't need to be fully automated from the first day. It starts with "Human - in - the - loop":

The Agent proposes, collects context, routes approvals, and records traces. Over time, as similar cases repeat, more and more paths can be automated because the system has a structured library of past decisions and exceptions.

Even if humans still make decisions, the graph keeps growing because the workflow layer captures inputs, approvals, and reasons as persistent precedents instead of letting them dissipate in Slack.

04 Why Existing Giants Can't Build a Context Graph

Ball optimistically believes that existing players will evolve into this architecture. According to this scenario, existing giants only need to attach an AI brain to their large data assets to smoothly transition to the next era.

Salesforce, ServiceNow, and Workday also believe this. They are all selling the same story: "We have the data, and now we add intelligence."

However, there is a flaw in this logic: Their underlying architecture is designed for the 'Current State'.

Take Salesforce as an example. It is essentially a huge and complex ledger. It precisely knows what a sales lead looks like now, but it cannot trace back what the world was like at the moment when the decision was made.

For example, when a 20% non - compliant discount is approved, Salesforce records "Discount approved".

But the context that justifies the discount, such as the service outage alarm just triggered by PagerDuty, the customer's angry complaints in Zendesk, and the VP's temporary authorization in the Slack group, is all lost at the moment of being written into Salesforce.

The inability to replay the world state at the time of the decision means the inability to audit the decision, let alone turn it into an "precedent" that AI can learn from.

The escalation of a customer service issue often depends on the customer's level in the CRM, the SLA terms in the billing system, and even the gossip in Slack. None of the existing SaaS giants can see the whole picture because their vision is limited to their own "walled gardens".

What about Snowflake and Databricks at the underlying layer? They are also highly expected and regarded as the cornerstones of AI Agents.

Indeed, data warehouses have time - based snapshots, which seem to have a "God's - eye view". However, the problem is that they are on the 'Read Path' of data, not the 'Write Path'.

Data enters the data warehouse usually after the decision is made, after being transported through a long ETL (Extract, Transform, Load) pipeline. It's like a data warehouse is just a recorder cleaning up the battlefield after a fierce battle.

By the time the data finally lands in Snowflake, the "decision - making context" full of games, trade - offs, and unexpected situations has evaporated. That is to say, it can only tell you what happened, not why.

Although Databricks is desperately integrating fragments, there is still an insurmountable gap between "the place where data is stored" and "the execution path where decisions are made".

Compared with these large companies, startups in the Agent system category have a structural advantage: They are on the 'Orchestration Path'.

When an Agent is routing work orders, responding to an incident, or approving a quote, it is not only calling tools but also executing the workflow.

It is at the center of the storm, pulling information from multiple systems, evaluating rules, resolving conflicts, and then taking action.

Because it is on the "execution path", it has a privilege that giants cannot reach: At the 'Commit Time', it can completely "freeze" all inputs, logic, exceptions, and reasons.

This is the Context Graph, and it is also the most valuable single asset of a company in the AI era.

Of course, existing giants will fight back. They will try to add orchestration capabilities through acquisitions. They will lock APIs and charge egress fees to make data extraction expensive.

This is the same scenario used by hyperscale cloud providers. They will build their own Agent frameworks and promote the narrative of "keeping everything in our ecosystem".

However, capturing Decision Traces requires being on the execution path at the Commit Time, not imposing governance afterwards. Giants can make data extraction more difficult, but they cannot insert themselves into an orchestration layer they have never participated in.

05 Three Paths for Startups

Startups in the Agent system category will take different paths, each with its own trade - offs.

1. Replace the existing Systems of Record from day one.

Reconstruct CRM or ERP around Agent execution, making "Event - sourced state" and "Policy capture" native features of the architecture. This is difficult because giants are deeply entrenched, but at the turning point of technological generations, it is not impossible.

Among the many startups pursuing the AI SDR (Sales Development Representative) category, Regie chose to build an AI - native sales engagement platform to replace traditional platforms like Outreach/Salesloft (the latter was designed for humans to execute sequences in a fragmented toolchain).

Regie is designed for hybrid teams, where Agents are first - class citizens: It can prospect, generate outreach, follow up, handle routing, and escalate to humans.

2. Replace modules instead of the entire system.

These startups target specific sub - workflows with special cases and approval - intensive processes, become the Systems of Record for these decisions, and synchronize the final state back to existing giant systems.

Maximor is implementing this logic in the financial field. It automates cash flow, closing management, and core accounting workflows while keeping ERP as the underlying General Ledger (GL).

In other words, ERP is still the "ledger" for accounting, but Maximor has become the "brain" that controls the reconciliation logic.

3. Create a brand - new System of Record.

These companies start from the orchestration layer and capture something that enterprises have never systematically stored - Decision Traces. Over time, this replayable relational data becomes a new authoritative asset.

At this point, Agents are no longer just automation tools but become the enterprise's archive for answering "why we do things this way".

PlayerZero is a typical example of this model. Production Engineering has long been at the intersection of SRE, QA, and development. It is a typical "glue function" that relies on human experience to carry the context that software cannot capture.

PlayerZero has built a "Context Graph" about the interaction between code, configuration, and customer behavior. When there is a problem in the production environment, it can answer "why it broke" and "what consequences this change will bring" - questions that no existing system can answer.

Above these paths, a new infrastructure is taking shape: Agent Observability.

As Decision Traces stack up, enterprises need to monitor Agent behavior just like they monitor code.

Arize is trying to become the Datadog in this new stack. It allows teams to see how Agents reason, where they fail, and evaluate the quality of their decisions. In the era of autonomous decision - making, this is not just a tool but a sense of security.

06 Two Key Signals for Entrepreneurs

For entrepreneurs, where should they focus? The signals released by the market overlap but point to different opportunities.

First, there are two general signals: high human input and high accident rate.

First, high human input. If a company still uses 50 people to manually route work orders or reconcile data, this is the most direct signal. The existence of a large amount of labor just proves that the decision - making logic is too complex for traditional tools to automate.

Second, the need to handle a large number of "unexpected" situations. Those transaction approval and compliance review processes full of "it depends" are also the best soil for Agents to establish decision - making lineage because the logic is complex and precedents are important.

Another signal specifically points to the birth of a "new System of Record": The "glue function" at the intersection of systems.

RevOps (Revenue Operations) exists because no one can handle sales, finance, and marketing systems simultaneously; DevOps exists because there is a deep gap between development and operations; and Security Operations (SecOps) is stuck between IT and compliance.

The emergence of these roles is actually a satire of the existing software ecosystem -