Ownership Before Agents: Why Conway's Law Eats Your AI Rollout

Seventy-five percent of data teams have no clear owner for their data products. Eighty-five percent have a clear owner for their infrastructure. That asymmetry comes from Joe Reis’s June 2026 Practical Data Pulse Survey (N=212), and it is the single most useful number I have seen for predicting whether an agent rollout will pay off or backfire.

The framing that ties it together is Conway’s Law. Any system you build ends up mirroring the communication structure of the organization that built it. An agent does not escape that law. It inherits it. The schemas it queries, the permissions it holds, the pipelines it triggers, those artifacts were shaped by who talks to whom and who owns what. When an agent runs, it runs straight into the org chart, because the org chart is already encoded in the substrate it operates on.

The Agent Is a Conway’s Law Amplifier

An autonomous agent does not float above your organization waiting for clean inputs. It works inside the exact structures your teams produced. If two teams never agreed on who owns the customer table, the agent finds two half-owned versions and picks one, usually the one with the loosest permissions. If nobody owns the revenue pipeline end to end, the agent optimizes the slice it can see and breaks the slice it cannot.

This is why the ownership number matters more than the tooling number. An organization where 85% of infrastructure has an owner but only 25% of data products do is unstable ground for automation: strong plumbing sitting on weak accountability. Drop a fast, tireless, permission-holding worker into that structure and you do not get order. You get the existing disorder executed faster.

The infrastructure-owned figure is the trap. Teams read 85% and conclude they are ready, because the machines are managed and the pipelines run. Infrastructure ownership answers “who keeps the lights on.” Data-product ownership answers “who is accountable for whether this number is correct and who is allowed to change how it is produced.” An agent needs the second answer far more than the first, and three out of four teams do not have it.

Anarchy Has a Weekly Cost, and It Is Measurable

Reis’s survey puts a price on the missing accountability. Teams operating in full organizational anarchy, no clear ownership of data products, spend roughly 45% of their week firefighting. Almost half the working week goes to chasing broken numbers, reconciling versions, and answering “which report is right.”

That number describes the human baseline before any agent arrives. It is what the organization already pays to survive its own ambiguity. Now add an agent that can issue thousands of queries, trigger dozens of pipeline runs, and write to tables no single person is accountable for. The firefighting does not shrink. The surface that can catch fire expands, and it expands fastest exactly where ownership is thinnest.

We have argued before that AI is an amplifier of whatever structure it lands in, and that autonomous systems erase accountability when no one owns the outcome. Reis’s data gives those arguments a measured floor. The 45% comes from what teams reported about their current week, in a survey the author ran himself and published openly.

Ownership Is an Upstream Fix, Not a Dashboard

The reflex, when an agent misbehaves, is to add observability. Watch what it does, measure its outputs, build a dashboard, catch the bad writes after they happen. That treats missing ownership as something you can detect downstream. You cannot. A dashboard tells you a number diverged. It cannot tell you who is allowed to decide which version is canonical, because that decision was never assigned to anyone.

Ownership has to be fixed before the agent runs, in the org design itself. The prescription that maps cleanly onto Reis’s finding is Team Topologies, from Matthew Skelton and Manuel Pais. It splits responsibility into two roles that the survey’s two numbers already gesture at. Platform teams own the infrastructure: the pipelines, the compute, the plumbing that keeps data moving. That is the 85% most organizations already staffed. Stream-aligned teams own end-to-end data products: a specific dataset or metric, from source to consumer, including the authority to define what correct means and to approve changes to how it is produced. That is the 75% most organizations left vacant.

The two roles are not interchangeable. A platform team that keeps the customer pipeline running is not the same as a team accountable for whether the customer table is right. The first is a maintenance function. The second is a decision-making function, and it is the one an agent needs a human to hold. Assign a stream-aligned owner to each data product an agent will touch, and you have given the agent an address to route accountability to. Leave it vacant, and the agent becomes the de facto owner of a decision no human agreed to delegate.

The Sequence That Actually Works

Ownership first, agents second. Not because ownership is philosophically prior, but because the reverse sequence is measurably worse. An agent deployed onto unowned data products inherits the anarchy, executes it at machine speed, and hands you a larger version of the 45%-firefighting week you started with. An agent deployed onto owned data products inherits a structure where every table it touches has a human accountable for it, and its mistakes route to someone who can correct them and adjust the pipeline that produced them.

The governance substrate has to exist before autonomous operations run on top of it. Ownership is the first layer of that substrate. It is cheaper to assign than any monitoring stack, and no monitoring stack substitutes for it.

Do This Now

Before you deploy a single agent against your data, run one exercise. List the data products the agent will read from and write to: the specific tables, metrics, and datasets, not the infrastructure. For each one, write a person’s name in an owner column. Not a team, not a platform, a person accountable for whether that product is correct and empowered to approve changes to how it is produced.

Count the blanks. If your organization matches Reis’s survey, three out of four rows will be empty. Each empty row is a place where your agent will become the owner by default, of a decision no human chose to delegate. Fill those rows using the Team Topologies split: platform teams for the infrastructure that moves the data, stream-aligned owners for the products themselves. Then, and only then, deploy the agent.

The teams that win with agents over the next two years are not the ones with the most autonomous agents. They are the ones who assigned ownership before they automated, so that when Conway’s Law does what it always does, it has a clean structure to mirror instead of an anarchy to amplify.

This analysis synthesizes Your Agents Are Stuck In Your Org Chart (Joe Reis, July 2026).

Victorino Group helps teams fix data-product ownership before deploying agents. Let’s talk.