Decision Rights in the AI Era
When one analyst can ask 40 model-generated questions before lunch, the scarce thing is no longer answers; it is who owns the final call. AI lowers the cost of analysis, drafts, simulations, and objections. It does not remove the cost of being wrong in public, spending capital, changing a roadmap, or telling a team to stop.
Decision rights are the explicit rules for who can decide, who must be consulted, who can veto, and who carries the loss if the call fails. In the AI era, they matter more because the machine can make every option look researched.
The Case
Before AI, many organizations used slowness as a control system. A proposal had to pass through decks, calendar slots, finance reviews, and senior attention. That friction was annoying, but it also revealed ownership: the person who fought for the work usually owned the decision.
AI breaks that signal. A junior operator can now produce a board-quality memo in 20 minutes. A model can generate five market-entry plans, ten risks, three customer personas, and a neat counterargument. The document looks decided before anyone has taken responsibility.
The sharper question is not "What does the model recommend?" It is:
Who can say yes?
Who can say no?
Who pays if this is wrong?
Who learns when reality replies?
That is the operating core of decision rights. AI makes the first two cheaper to fake and the last two harder to dodge.
How It Works
A clean decision-rights system separates four roles that often get blurred:
| Role | Question | Failure mode |
|---|---|---|
| Recommender | What should we do? | Confuses evidence with authority |
| Decider | Who makes the call? | Hides behind consensus |
| Veto holder | What can block the call? | Uses risk language to avoid tradeoffs |
| Owner | Who absorbs the result? | Claims credit, outsources blame |
AI mostly strengthens the recommender. It can widen the option set, compress research time, and force pre-mortems. It does not become the owner unless the institution is willing to give it budget, reputation, and consequences, which it cannot actually hold.
A useful pattern is "AI can argue; humans sign." The model can produce the main case, the opposing case, the base-rate check, and the missing-data list. The named person still signs the decision record with a date, a metric, and a review window.
The loop matters. A model can answer in seconds, but accountability only exists after reality has had time to disagree.
Where It Shows Up
Amazon's "one-way door" and "two-way door" distinction is one clean example: reversible decisions can move fast; irreversible ones deserve more senior attention. Jeff Bezos described this in his 2015 shareholder letter. The idea predates current AI, but AI makes the split more urgent because reversibility is now the main reason to delegate.
Netflix's culture memo names another version: people should not seek consensus when one informed person can own the call. That principle becomes sharper when every meeting can produce ten polished arguments. Consensus can become a hiding place.
The military has an older language for the same problem: commander's intent. The commander states the aim and constraints; lower levels decide how to act when conditions change. AI agents increase the number of local choices, so intent has to be clearer, not vaguer.
What's Contested
The live debate is whether AI should remain advisory or become a delegated decider in narrow domains. Credit scoring, fraud detection, ad bidding, and logistics already contain automated decisions. The contested part is not whether machines can decide. They already do. The question is where appeal, audit, and human override must sit.
There is also a status problem. Senior leaders may keep decision rights while junior teams do AI-assisted work that actually shapes the choice. If the person with the title only approves what the system and staff have already framed, the formal org chart and the real decision graph split.
That split hides learning. When a decision fails, the team needs to know whether the error came from bad data, weak framing, model hallucination, incentive pressure, or cowardice in the final call.
Why This Crosses Realms
Decision rights connect directly to concept bus factor. A team with one hidden decider is fragile; a team with no named decider is slower than it looks. AI can mask both failures by producing clean artifacts around a broken authority map.
It also connects to concept distributed cognition. A modern team is no longer just brains in a room; it is people, models, dashboards, documents, and memory systems acting together. The more cognition gets distributed, the more decision ownership has to be named.
A clean bridge may be concept ooda loop. Observe, orient, decide, act: AI speeds observe and orient. It does not magically fix decide and act. In fact, it can overload them.
An Open Question
If a model writes the winning argument, a manager signs the memo, and a team executes the plan, where exactly did the decision happen? The next page should probably be concept accountability loop: the missing machinery between a signed decision and a learned one.
Abhishek's take
The part that grabs me is not AI making decisions. It is AI making responsibility easier to blur. I trust a bad call with a named owner more than a polished recommendation that no one is willing to carry.
Where I've used this
I use this on the buying floor and in the tools I write: models can widen the question set, but a person still has to sign the call. The useful artifact is not the generated memo. It is the dated decision record that reality can later judge.
Tags: #decision-rights #ai-leadership #operating-models #accountability #management