The a16z thesis

a16z partner Seema Amble writes that AI could strengthen systems of record: an agent uses their data, permissions and available actions even if the user no longer opens the familiar interface. Salesforce with Anthropic[1], Docusign Iris, Atlassian Rovo and Klaviyo Composer are cited as examples of the shift from storing records to performing work. [1 · Andreessen Horowitz]

In the proposed hierarchy, agents progress through four levels: finding information and preparing an answer; executing a process; applying organizational rules to an ambiguous situation; and making decisions independently with explicit trade-offs. Incumbent providers move faster through the first three levels where they already control data and permissions. [1 · Andreessen Horowitz]

Why an interface over a model is not enough

A vertical product is vulnerable if it merely adds a new screen on top of an old system. A more defensible position emerges when the product takes a task through multiple systems and participants, from identifying a problem to a verified result, and accumulates knowledge about exceptions that a general-purpose agent lacks. [1 · Andreessen Horowitz]

A learning cycle based on corrections becomes a durable asset. a16z cites Harvey: the company created approximately 1,750 simulated legal tasks, each with an average of 50 quality criteria, to train and evaluate the system before accumulating large volumes of client data. [1 · Andreessen Horowitz]

A practical criterion for selecting products

The work must recur often enough, require expert judgment, allow the result’s correctness to be assessed quickly and expand from a single task into an entire workflow. These four questions help distinguish a standalone product from a feature that an existing provider could quickly incorporate. [1 · Andreessen Horowitz]

For consumer devices, similar protection could arise in after-sales services: setup, compatibility, warranties and support that learns from technicians’ corrections. The device itself is easy to compare on price, while an accumulated history of correct decisions creates a more durable customer relationship. [1 · Andreessen Horowitz]

Expert commentary

It is useful to monitor not only providers’ statements but also which actions agents actually perform without a manual return to the old interface, how often they make errors on exceptions and who owns users’ corrections. Those data will show whether an agent strengthens a system of record or creates an independent new layer. [1 · Andreessen Horowitz]

Sources

  1. Andreessen Horowitz — The Incumbents Are Coming — September 3, 2026