What you own when your team uses AI
There is a common assumption that AI tools diffuse accountability — that because the output was generated by a model, the responsibility for that output is shared with the tool, or the vendor, or whoever configured the system.
It is not. The accountability sits exactly where it always has: with the people who directed the work, reviewed the output, and authorised its use.
Why this matters now
When your team uses AI tools, they are producing output faster than before. More drafts, more reports, more summaries, more first versions of things that previously took longer. That volume is useful. It is also a risk if the review discipline does not scale with it.
The output looks finished. Well-formatted documents, confident language, structured arguments. It is easy to approve something that looks right without verifying that it is right. And at the point of approval, the accountability transfers to the approver — not to the model that produced it.
The three accountability levels
Not every task has the same consequence, and the required level of review should reflect that.
Solo work. Notes you take for yourself, drafts you use as a starting point, research aggregation you will revise. At this level, your own judgment is the check. If you use AI output and it is wrong, you carry that. The bar for verification is set by the consequence to you.
Mission-level work. Work that goes to your team, informs team decisions, or is used in internal processes. At this level, individual review is necessary but the output affects others. Someone with relevant context should verify the key claims. An AI-drafted briefing that is wrong shapes a team's work before the error surfaces.
Company delivery. Work that goes outside your organisation — client deliverables, financial data, regulatory documents, partner communications. At this level, a defined process matters more than individual good intentions. There should be a named person who owns each category of output, a defined review step that is not optional, and a clear path for correction when something is wrong.
What review actually means
Review means you can stand behind the content. Not that it looked correct, not that you read it quickly, not that the formatting was clean. It means you can independently verify the claims that matter and you would be willing to defend them if questioned.
A practical test: if you handed the document to a colleague and they found a significant error, could you explain where in your review process that error should have been caught? If the answer is "it should not have passed my review," that is the right answer. If the answer is "I was not really checking at that level," the review was not a review.
This is not about being slow. It is about the review being real rather than nominal.
What to ask before approving AI-assisted output
Before an AI-assisted deliverable goes to a client, a regulator, or a leadership decision:
Who produced this? Not which tool — who directed the work, what context did they provide, what did they verify before passing it on?
What are the claims that matter? Every document has a few key statements that carry the weight of the rest. A financial summary has numbers. A project update has status claims. A proposal has scope and timeline. Have those been verified independently, or only checked for consistency within the document?
Who can defend this if questioned? That person is accountable for it. If no one can answer that question, the document is not ready to send.
What is the recovery path? If an error is found after it is sent, what happens? Who is responsible for the correction, how quickly, and to whom?
These are not bureaucratic questions. They are the questions that define whether your team is actually accountable for its work or just producing volume.
The Swedish regulatory context
In the Swedish professional context, a specific category of documents carries legal or contractual weight: procurement responses, financial statements, compliance reports, employment decisions. These documents have signature requirements and legal accountability that exist regardless of how they were produced.
AI tools can draft these documents. The accountability structure for them does not change because AI was involved in the drafting. The person who signs is accountable. The process that led to signature should reflect that.
What comes next
Part 3 covers how to measure whether this accountability structure is actually working: what signals to track, what to stop measuring, and how to structure a review that catches what matters.
Next in this series: Part 3 — Measuring outcomes, not activity