Vocap
← Back to Insights
Emery Waddell

The Four-Part Moat: Revisited, Part 2

2026-11-17

Part 2 of 3: Compounding context and the right to act

The short version

  • The most durable form of cornered data is compounding context: what the product learns about how each customer actually runs, one completed workflow at a time.
  • The moat is not capturing data. It is grading outcomes and improving from the grade.
  • A system of action needs more than the ability to act. It needs the customer's trust to act, and a seat close to where goals are set.
  • Products that only search, monitor and report are exposed. That is exactly what general-purpose agents do well.

Cornered data: grading, not capturing

In Part II we described cornered data as the "what and why" of how a business operates. We would now put it more precisely.

The moat is not capturing context. It is grading it, inside each customer.

Inside one customer, the product learns how that business actually runs: the exceptions, the trade-offs it prefers, where it rightly departs from the default, and the judgment of the veteran planner who will someday retire. The product should know far more about a customer on day 400 than on day 1. A competitor cannot buy that knowledge, and a model cannot absorb it, because it was never written down. It has to be earned, and the switching cost grows every month.

Earning it takes a closed loop. The product makes a call, reality scores it, and the score flows back into sharper policies and better defaults.

Consider maintenance in multifamily property management. A new product arrives with sensible defaults: how to triage a work order, which vendor to send, what a fair price looks like. Inside one owner's portfolio, every ticket closes with a score: what it cost, how long it took, whether it reopened, whether the resident renewed. Over time the product learns that slow drains in this owner's 1980s garden-style buildings need a plumber, not a handyman; that one vendor's "fixed" tickets keep reopening; and that residents who wait more than two days for an HVAC fix rarely renew. None of that exists anywhere a model could read it. A competitor that wins the account tomorrow starts at day 1.

A useful test: if a customer's general-purpose agent had full access to your data, would your product become more valuable or less?

If the workflow memory, preferences and grading live in your system, the agent makes you more useful. If they don't, the agent can replace you.

Lessons learned across customers matter too, but they are shorter-lived. Competitors serving similar customers learn the same lessons, and a widely known best practice is no longer an advantage.

Cross-customer insight gets you in the door. Depth inside each customer keeps you there.

System of action: does it tell you, or does it do it?

Part II called a system of action "the transition from recording the work to performing the work." The simplest version of that test:

Is it an information tool or a doing tool?

Most promising vertical startups follow the same two phases. Phase one is a search and monitoring tool that finds, tracks and summarizes what matters in the industry and produces answers and reports. Phase two is embedding into the workflows downstream of those answers, where decisions get made and work gets done. The second step is hard. The downstream workflow usually sits on someone else's system of record and demands write access, deeper integration, change management and a customer willing to be accountable for the result.

A company that stays in phase one is vulnerable.

Search, monitoring and reporting are what horizontal agents do best. Once one can connect to the same sources, a standalone answer engine becomes a feature.

Any agent framework can act. Getting to phase two takes two harder things.

The right to act. The customer trusts the product with consequential decisions, not just recommendations. That trust is earned in stages: a person reviews everything, then only the exceptions, then only sets policy. Stripe Radar decides whether a payment goes through. Sierra resolves customer conversations end to end and charges per resolution, a price that only works because the software is trusted to finish the job. The outcome-based pricing shift we flagged in Part I follows directly from this.

Proximity to intent. The strongest position is where the user states the goal. Think of an org chart. At the bottom you see a small piece of the picture and hear about decisions last. Near the top you hear the intent firsthand and decide how to mobilize resources. The top is a much harder position to replace.

Two questions now anchor our first conversation with founders. What does your product know about a customer on day 400 that it didn't on day 1? And does it tell the customer what to do, or does it do it? A company can have one without the other. The strongest have both, and each makes the other stronger. The final post shows how, using Vocap portfolio company ketteQ.

Further reading

From Vocap: Forget the SaaSpocalypse. Here's How Vertical AI Actually Wins; Part II: Building the Moat; What the Cerebras Debate Tells Vertical AI Founders (why value is moving to orchestration).

Elsewhere: Bessemer Venture Partners, Roadmap: AI systems of action; Foundation Capital, AI's trillion-dollar opportunity: Context graphs; Martin Casado and Peter Lauten (a16z), The Empty Promise of Data Moats; Ben Thompson, Aggregation Theory; Bret Taylor on outcome-based pricing, Cheeky Pint interview.