Who is responsible for an AI assistant's answers? A practical model

An AI assistant may answer quickly and use approved sources, yet still run into an ordinary problem: nobody knows who should correct outdated information. The vendor operates the platform, sales knows the offer, support sees complaints, and legal joins only after an incident. The gap between those roles can leave a wrong answer active for too long.

Accountability should not sit with one vaguely defined “AI administrator”. It should follow specific decisions. The NIST AI Risk Management Framework recommends clearly defined human roles, oversight and procedures. The OECD accountability principle highlights traceability, documentation and ongoing risk management throughout the system lifecycle.

Responsible AI operations do not require one person who knows everything. They require a team in which everyone knows which decision they own and when it is due.

Five roles that cover everyday operations

1. The service owner defines purpose and boundaries

This person decides why the assistant exists, who it serves, which subjects it may cover and when a human should take over. The role may belong to a support lead, ecommerce manager, municipal office manager or department head.

2. The content owner is accountable for the source

Every important subject needs a named person or department. Sales owns prices and commercial terms, HR owns careers information, and each municipal department owns its procedures and deadlines. The content owner approves the source, sets a review date and updates it when facts change.

Many failures begin in the source rather than the model. If a page still contains an old price or deadline, the assistant may accurately repeat the wrong fact.

3. The operations owner watches real conversations

Support teams see what visitors misunderstand and where conversations end without a result. The operations owner triages unanswered questions, feedback and human handoffs, then assigns each finding to the relevant content owner or technical administrator.

4. The technical administrator protects access and integrations

This role manages permissions, allowed domains, imports, limits and incidents. It verifies that sources were ingested, internal documents stay out of the public assistant, and human handoff continues to work. It should not decide whether a price list or public procedure is factually correct unless it owns that subject.

5. A subject-matter approver reviews sensitive topics

Not every edit needs legal review. Personal data, health, financial or binding information may need a privacy officer, specialist or lawyer. The approver defines acceptable wording and the cases the assistant should send directly to a person.

Use a simple operating rhythm

Before launch

• Write down the assistant's purpose and excluded subjects.
• Assign an owner and review date to every source.
• Test real, ambiguous, conflicting and sensitive questions.
• Configure human handoff and decide who receives the alert.

Every week

• Review unanswered questions, negative feedback and handed-off conversations.
• Fix the source, instruction or integration according to the real cause.
• Add repeated failures to a small regression test set.

Every month or quarter

• Confirm that owners, contacts and permissions remain current.
• Review public sources, high-risk topics and handoff quality.
• Check that the assistant has not drifted into unapproved use cases.

The NIST AI RMF Playbook asks who maintains, re-verifies, monitors and updates an AI system after deployment. Those questions separate a living service from a launch document that nobody revisits.

Online store example

An assistant promises free delivery after a promotion has ended. The operations owner flags the conversation, the content owner corrects the source, the technical administrator confirms re-ingestion, and the approver repeats the test. The service owner decides whether affected customers need a direct response.

Municipal example

A resident asks about waste collection and the assistant retrieves an old timetable. The responsible officer corrects the official source, operations finds similar conversations, the technical administrator refreshes the knowledge base, and the service owner decides whether the assistant should temporarily link to a current notice.

What to do after a wrong answer

1. Capture: save the question, answer, source and time.
2. Limit impact: disable the affected source or topic and route visitors to a person when the risk is serious.
3. Find the cause: distinguish outdated content, a missing source, a bad instruction, an access error and a technical failure.
4. Fix and verify: change the cause, not only one answer, then test related questions.
5. Close: record the owner, deadline, result and any process change.

Start with fifteen minutes

Open the last five problematic conversations. For each one, write down the source owner, the operations owner and the repair deadline. Any blank field reveals an organisational risk before it becomes an incident.

Informio can keep answers tied to controlled sources, show citations, surface conversations and feedback, and hand a conversation to an operator. The tool cannot decide who in your organisation approves a price list or public procedure. Clear content and operations ownership is therefore part of a sound deployment.

Would you like an AI assistant with a practical accountability model? Arrange a consultation with Informio.

Sources

• NIST AI Risk Management Framework
• NIST AI RMF Playbook: Govern
• OECD AI Principle: Accountability
• OECD AI Principle: Transparency and explainability

This article offers an operational framework, not legal advice. NIST AI RMF is voluntary, and specific duties depend on the use case, contracts and applicable law.