Decision Rights in AI Teams
When one model can draft 20 options in a minute, the scarce question is who owns the final call, not who produced the first draft. AI makes ideation cheap, but it does not make accountability cheap. A team without clear decision rights starts mistaking volume for judgment.
The old bottleneck was often output: who could write the memo, build the model, make the deck, pull the cases. In AI teams, the bottleneck moves to authority: who can say yes, who can say no, who carries the cost when the answer is wrong.
The case
Herbert Simon's 1947 point still bites: organizations exist partly because human attention is limited. AI changes the shape of that limit. It can expand the number of options a team can inspect, but the human system still has to choose under time, politics, risk, and incomplete facts.
Decision rights name that system before the meeting starts. They answer four questions:
| Question | Bad version | Better version |
|---|---|---|
| Who frames the decision? | "Everyone contributes." | One named owner writes the decision statement. |
| Who supplies evidence? | "Ask the model." | People and tools provide cited inputs. |
| Who decides? | "Leadership aligns." | One accountable person has the D. |
| Who audits the result? | "We will monitor it." | One review date and one metric are named. |
The AI-specific trap is that generated work looks finished. A model can produce a pricing argument, a hiring rubric, a policy draft, or five product narratives with confident syntax. That polish can hide the fact that nobody has accepted risk. A signed bad decision is easier to learn from than an unsigned polished one.
Where it shows up
RACI gives one older vocabulary: responsible, accountable, consulted, informed. Rogers and Blenko's 2006 HBR piece sharpened the language with RAPID, especially the question of who has the "D." AI teams need the same clarity, with one added line: what work was machine-generated, what was verified by a person, and what remains an assumption.
A useful decision log for an AI-assisted team can be small:
| Field | Example |
|---|---|
| Decision | Ship variant B to 10% of users on 2026-07-20 |
| Owner | Named person, not a group |
| AI role | Drafted test plan and edge cases |
| Human checks | Legal, data quality, customer support load |
| Reversal point | Error rate above 2.5% for 48 hours |
| Review date | 2026-08-03 |
The point is not paperwork. The point is memory. Without a record, the team cannot separate a bad prompt, bad evidence, bad judgment, and bad luck.
What's contested
The live debate is how much autonomy to give AI agents before the decision right has to return to a human. NIST's 2023 AI Risk Management Framework uses governance language because high-impact systems fail socially before they fail technically: nobody knows who approved the behavior, who saw the warning, or who could stop it.
There is also a speed tradeoff. Too much approval turns AI work into theater. Too little approval creates a hidden bureaucracy where the model acts, the team reacts, and the accountable person only appears after damage.
Cross-realm bridge
Decision rights rhyme with concept bus factor because both ask the same unglamorous question: where does the system break if one named person disappears? They also touch concept ooda loop. AI can compress observe and orient, but decide and act still need ownership.
This is why concept information theory matters here. More generated text is not more signal unless the team has a filter. The decision owner is the final compression function: many possible futures become one committed path.
Key sources
- Herbert A. Simon, Administrative Behavior (1947) - the classic account of bounded rationality inside organizations.
- Paul Rogers and Marcia Blenko, "Who Has the D?" Harvard Business Review (2006) - a useful popular frame for decision ownership.
- Andrew S. Grove, High Output Management (1983) - useful for output, meetings, and managerial accountability.
- NIST, AI Risk Management Framework 1.0 (2023) - governance vocabulary for AI systems with human consequences.
Abhishek's take
The mistake I see in AI work is treating the model as the new center of the team. It is not. The center is the person willing to write their name next to the call, define the reversal point, and come back after 14 days with the evidence.
Where I've used this
I use this pattern on the buying floor and in the tools I write: the machine can propose, rank, and explain, but a person owns the call. The log matters because memory beats vibes when a decision has to be defended 30 days later.
Tags: #decision-rights #ai-teams #management #accountability #operating-models