The Named Owner Problem: Who Is Accountable When Your AI Agent Gets It Wrong

Your company is about to run more decision-makers than it employs. Machine identities already outnumber human ones by somewhere between 45 and 109 to one, depending on whose count you use¹²²¹, and a growing share of those non-human identities are AI agents that do not just read data but act on it. Every one of those actions is, legally and operationally, a decision your company made.

Here is the problem nobody provisioned for: intelligence scales at near-zero marginal cost, and accountability does not. You can spin up a hundred agents this quarter. You cannot spin up a hundred people willing to sign their name to what those agents decide. A landmark 2026 report from Accenture and Wharton put it in one line: “Intelligence may be scalable, but accountability is not.”¹

This article is for the CTO, operations lead, or Geschaeftsfuehrer who is moving AI agents into production and has quietly realised that “IT owns it” is not an answer that survives a bad outcome. The fix is not more dashboards. It is a named human owner for every agent, a tiered model of which decisions the agent makes alone, and a record of the reasoning behind each action so that owner can actually answer for it.

The Accountability Gap Nobody Provisioned For

Most AI governance conversations in 2026 are about identity, access, and monitoring: who the agent is, what it can touch, and what it did. Those matter. But they answer the wrong question. When an agent issues a wrong refund, sends a customer the wrong price, or files the wrong compliance evidence, the question in the room is not “what did it do”. It is “who answers for this”. And in most organisations, nobody has been named.

• Decisions everywhere, owners nowhere - As agents move from advising to acting, high-stakes decisions get made with no specific human attached to them. Industry analysts now describe this directly as “decisions everywhere, but owners nowhere”³.
• Governance is not keeping up with deployment - Adoption is racing ahead of control. McKinsey finds that while roughly three quarters of organisations plan to adopt agentic AI within two years, only about a fifth have a mature governance model for the agents they already run⁷.
• Fewer than 1 percent are fully mature - A large cross-industry survey found that fewer than 1 percent of organisations have fully operationalised responsible AI, and 81 percent are still in the earliest stages⁷.
• The identity flood hides orphaned agents - With machine identities outnumbering humans by up to 109 to one¹², agents accumulate faster than anyone assigns owners to them, and orphaned agents keep acting after the person who set them up has moved on.
• Shadow AI makes it worse - Nearly three quarters of knowledge workers now use AI, much of it through unsanctioned tools¹, which means agents and actions exist that no owner has ever formally claimed.

This is not a monitoring problem you can dashboard your way out of. It is a structural gap between how fast you can deploy intelligence and how fast you can assign the human responsibility that has to sit behind it. Closing it starts with being precise about what accountability actually means.

What Accountability Actually Means for an AI Agent

Accountability, observability, and identity get used interchangeably, and that confusion is a large part of why the gap exists. They are different jobs. Identity says who the agent is. Observability says what it did. Accountability says who answers for it, can explain it, and carries the consequence. You can have perfect logs and still have nobody accountable.

The three layers of accountability

• Liability - Who bears the consequence when the agent is wrong. This is the legal and financial layer, and it always lands on a human or the company, never on the model³⁵.
• Explainability - Whether the decision can be reconstructed and justified to an auditor, a regulator, or a customer. Without a record of reasoning and sources, the owner is liable but mute³.
• Organisational alignment - Whether the action reflected the company’s actual rules and values, not a plausible-sounding invention by the model³.

The law is blunt about where liability lands, which is why the other layers exist to serve it. A company cannot use an agent’s autonomy as a shield.

• Vicarious liability applies - Under agency law, a principal answers for the acts of its agents within the scope of their authority, and this doctrine is being applied directly to AI systems⁵.
• Autonomy is not a defence - A California statute bars defendants from arguing that the AI autonomously caused the harm, explicitly rejecting the idea that software can absorb accountability⁵.
• Agents bind the company - In the Air Canada case, a tribunal held the airline to a refund policy its own chatbot had invented, rejecting the claim that the bot was a separate entity¹⁹.
• You must be able to explain it later - Regulators expect companies to know what authority their agents have, what they can access, how actions are supervised, and how they will be explained after the fact⁵.

Why the Named Owner Problem Landed in 2026

The accountability gap is not new in theory, but three shifts turned it from a talking point into an operational emergency this year. The common thread is that agents crossed from advising to acting, and the scale of that action outran the human structures around it.

1. Agents got write access - The move from read-only copilots to agents that create purchase orders, issue refunds, and update records means every action now has real-world consequences a human has to own⁶.
2. Deployment outran governance - With roughly 74 percent of organisations planning agentic AI within two years but only 21 percent having mature governance, the gap between what is running and what is controlled widened sharply⁷.
3. The identity count exploded - Non-human identities now outnumber humans by dozens to one, and analysts flag non-human identity sprawl as a primary enterprise security risk for 2026¹¹¹².
4. Regulation put a clock on it - The EU AI Act became fully applicable in August 2026, with human oversight, AI literacy, and transparency obligations that assume a responsible human behind each system¹⁵.
5. The failures got expensive and public - Court decisions like Air Canada, and penalty caps up to EUR 35 million or 7 percent of global turnover, moved accountability from a governance nicety to a board-level exposure¹⁴¹⁹.

Put together, these shifts mean the old answer - “the platform team manages our AI” - now leaves a company legally exposed and operationally blind. The response is a specific, assignable model of ownership.

The Named Owner Model

The core move is simple to state and hard to dodge: every agent in production has exactly one named human owner. Not “IT”. Not “the AI team”. Not a committee that meets monthly. One person, identified by name and role, who is accountable for what that agent does. Research on governance maturity is consistent that named accountability outperforms diffuse ownership³¹⁰.

What the named owner is on the hook for

• Scope - Defines what the agent is allowed to do and, more importantly, what it is not.
• Boundaries - Sets the autonomy tier for each action class, deciding what runs alone and what needs sign-off.
• Review - Approves or rejects the high-impact actions the agent escalates.
• Explanation - Can reconstruct and justify any action after the fact using the recorded reasoning.
• Consequence - Is the person who answers to leadership, auditors, or customers when the agent is wrong.
• Lifecycle - Owns the agent through changes, and hands ownership to a named successor before leaving.

“IT owns it” fails for a specific reason: the platform team can run the infrastructure but cannot judge whether a given refund, price, or contract clause was the right business decision. Accountability has to sit with someone who understands the process the agent is acting inside.

Tiered Autonomy: Which Decisions the Agent Makes Alone

A named owner can only stay accountable if they are not drowning in approvals. The mechanism that makes ownership workable is tiered autonomy: you match how freely an agent acts to how much damage it can do if it is wrong. Low-impact and reversible actions run alone; high-impact and irreversible ones stay human-led. This is the practical version of bounded autonomy that governance frameworks now recommend³⁶.

The three tiers

The tier is a property of the action, not the agent. A single support agent might tag tickets at Tier 1, issue small refunds at Tier 2, and escalate a contract dispute to Tier 3. The owner sets those thresholds and can tighten them after any incident.

How the tiers play out in practice

• Finance - An accounts payable agent matches and posts invoices under EUR 5,000 alone (Tier 1), routes anything above that for a controller’s approval (Tier 2), and never releases a payment run without a human (Tier 3).
• Customer service - A support agent answers routine questions and updates cases alone, offers a goodwill credit up to a set limit with async approval, and hands any legal or safety complaint to a person.
• Sales - A sales agent drafts and sends standard quotes at list price alone, escalates any non-standard discount for approval, and never signs a framework agreement.
• Procurement - A procurement agent reorders standard consumables against agreed contracts alone, flags a new supplier for review, and leaves capital purchases to a human.
• HR - A recruiting agent schedules interviews and sends status updates alone, but any screening decision that filters candidates stays human-led, because it is a high-risk use under the EU AI Act¹⁵.

The Company Brain: Recording the Reasoning Behind Every Action

Naming an owner and setting tiers is worthless if the owner cannot reconstruct why the agent acted. Accountability without a record is a liability with no defence. This is where a Company Brain - a governed memory layer that stores your rules, context, and the provenance of every answer - turns the named owner model from a policy into something operational²⁰.

What the record has to contain

• The reasoning - The decision path the agent followed, not just the final action, so the owner can see how it got there³.
• The provenance - Which source, rule, or document the action was based on, so every answer points back to something the owner can check²⁰.
• The confidence - How certain the agent was, so low-confidence actions can be routed to a human by default³.
• The authority - Which tier the action fell into and who, if anyone, approved it.
• The context - The company-specific rules and exceptions in play, so the action reflects your policy and not the model’s guess²⁰.

Analysts describe this as a “digital flight recorder” or chain-of-thought audit trail: force the agent to log its reasoning, sources, and confidence so that when a failure happens, you can assign responsibility precisely instead of guessing³. A Company Brain is where that record lives and, crucially, where it stays after individuals leave.

When Agents Call Agents: Accountability in Chains

The named owner model gets harder, and more important, the moment agents start calling other agents. A single request now passes through several agents, each acting on the output of the last, and a wrong step early in the chain propagates silently until it surfaces as a real-world action nobody chose. This is where accountability most often collapses in practice⁹.

• Diffuse causation - When five agents each did part of a task, it is tempting to say no single one is at fault. That instinct is exactly the trap; the chain still has one named owner accountable for its outcome.
• Silent propagation - A low-confidence guess by an upstream agent becomes an input the downstream agent treats as fact, so a small error compounds into a confident wrong action⁹.
• Cross-team chains - A sales agent hands to a finance agent that hands to a fulfilment agent, crossing team boundaries where no single person sees the whole path.
• Missing checkpoints - Without a tier boundary somewhere in the chain, the whole sequence can run to an irreversible action with no human ever in the lead.
• Attribution after the fact - When it goes wrong, you need the record to show which agent, on which input, made the decision that tipped it - which only a shared reasoning trail provides.

How to Put a Named Owner Model in Place

This is a weeks-long exercise, not a year-long programme, and most of the work is inventory and boundary-setting rather than technology. The goal is to reach a state where no agent runs in production without an owner, a tier map, and a record. Here is the sequence.

1. Inventory every agent - List every agent already acting in production, including the shadow ones. You cannot assign owners to agents you have not found, and the identity flood means there are more than you think¹¹.
2. Assign one named owner each - Attach a specific person and role to every agent. If nobody will own an agent, that is your signal to pause it, not to leave it running unowned.
3. Classify actions into tiers - For each agent, sort its action types into Tier 1, 2, or 3 by impact and reversibility. Start conservative and loosen as trust builds.
4. Wire the record - Route the agent’s reasoning, sources, and confidence into a Company Brain so every action is explainable by default, not reconstructable only in a crisis³.
5. Set escalation paths - Define exactly how Tier 2 approvals reach the owner and how Tier 3 decisions get made, so nothing waits in a queue nobody watches.
6. Handle the leaver case - Make owner departure a formal event: reassign ownership, re-review permissions, and pause any agent with no named successor.
7. Review after every incident - When an agent gets something wrong, the owner tightens a tier or a rule. The system gets safer with each correction rather than accumulating silent risk.

How Superkind Fits

Superkind builds AI employees that live inside your systems and run on a Company Brain - a governed memory layer that holds your company’s knowledge, rules, and the provenance behind every answer²⁰. That combination is what makes the named owner model workable in practice rather than a policy on a slide.

• Company Brain provenance - Answers point back to a source by default through citations and provenance, so the named owner can explain and defend every action after the fact²⁰.
• Learns your company, not the internet - AI employees act on your documented rules and exceptions, so actions reflect your policy rather than a plausible invention.
• Human oversight built in - AI employees take over recurring routine work while people focus on what genuinely needs human judgement, which is exactly the tiered split.
• Explainability and access controls - Explainability reporting and access controls are built into what gets deployed, so you get audit-ready evidence without a separate tooling stack.
• Sits on top of your systems - AI employees connect to your existing tools rather than replacing them, so ownership maps to processes your team already runs.
• EU-hosted or on-premise - Personal data can stay within your perimeter, and deletion means removing a source rather than retraining a model²⁰.
• Knowledge that outlives people - Because rules and reasoning live in the Company Brain, ownership can pass to a successor without the reasoning walking out the door.
• Live in two weeks - The first use case goes live in two weeks, so accountability structures get tested on real work quickly rather than in theory.

Decision Framework: Is Your Agent Ready for Production?

Before an agent handles real decisions, run it through this. If any row fails, the agent is not ready, no matter how capable it is.

Frequently Asked Questions

Related Articles

• AI Agent Identity: How to Authenticate and Authorise the Non-Human Workforce
• AI Agent Observability: Seeing What Your Agents Actually Do
• Why AI Agents Need Write Access: From Read-Only Copilots to AI Employees That Act
• Shadow AI Governance: Finding the AI Nobody Approved
• Agent Sprawl: When You Run 40 AI Tools and No Company Brain Connects Them
• AI Agents for the Mittelstand: Deploying AI Without Losing What Makes You Great

Henri Jung

Co-founder of Superkind, where he helps SMEs and enterprises deploy custom AI agents that actually fit how their teams work. Henri is passionate about closing the gap between what AI can do and the value it creates in real companies. He believes the Mittelstand has everything it needs to lead in AI - it just needs the right approach.