Insights
How should decision rights change when AI agents take action?
Updated September 2026
Decision rights have to be written down for each agent and each type of decision before the agent acts. For every one, name what the agent may recommend, decide, and execute. Also name what makes it stop, who can override it, and which person owns the outcome. "Human in the loop" is too vague to govern with: a person who approves every action and a person who reads a report after 10,000 actions are running completely different control models.
A value statement is not a control
A speed-limit sign is not a braking system. Statements like "humans remain accountable" describe what should happen. Decision rights determine what actually happens when an agent is doing the work at full speed.
The gap is widening. 30% of workers say AI agents are already integrated into their workflows, up from 13% a year earlier, while governance for oversight and accountability lags far behind (BCG, 2026). Among large organizations, 40% now report scaling AI agents, up from 27% (McKinsey, 2026).
Four levels of agent authority
| Agent authority | What the human does | Fits decisions that are | Example |
|---|---|---|---|
| Recommend | Makes every decision | High-stakes, rare, hard to reverse | Approving a credit exception |
| Decide with approval | Approves or rejects each action | Consequential and frequent | Issuing a refund outside policy |
| Act within limits | Reviews exceptions only | Routine, bounded, reversible | Rescheduling service appointments |
| Act and report | Sets boundaries, reviews results in aggregate | High volume, low risk per action, easy to undo | Tagging and routing inbound email |
Authority should move down this table only as two things grow: how easily a mistake can be reversed, and how much evidence the agent has earned.
Seven questions every agent's decision rights must answer
- Can it recommend, decide, or execute?
- Can it spend money?
- Can it change a customer record or send a communication?
- Can it modify production systems?
- What causes it to stop?
- What requires human approval, and who can override it?
- When something goes wrong, who owns the outcome?
Where decision rights should live
- With the business owner of the process, not with IT alone. The person who owns the result owns the rules.
- In a registry people actually use, next to the agent's owner, scope, and stop conditions.
- Revisited when anything shifts: a wider scope, higher volume, or a change in error rates.
- Sized to capacity. Every team has an Oversight Ceiling. Granting authority past it turns review into rubber-stamping.
What our field data shows
The perspective most often missing in our field data is the one decision rights depend on. 55% of respondents show the Conductor lens, which asks "will this actually work in practice?", as their weakest. Teams without it tend to write rules that read well and break under load.
Source: Neurocollective n.print field data, 464 respondents from 182 email domains, as of September 2026. Self-reported. No single organization exceeds about 30% of responses.
Where this fits in the discipline
AI Adoption Engineering, developed by Dr. Lisa Palmer and Neurocollective, converts AI investment into measurable business results by building human and organizational capability. Clear decision rights put two of its Four Guiding Principles into daily practice: Speed with Rigor and Human-Centricity.
Writing decision rights is the easy part. Using them under pressure is the capability. Decision Rehearsal, powered by n.sim, puts a leadership table inside a realistic agent scenario where conditions change mid-round, so teams find the gaps in their rules before production does.
Related answers: What is shadow AI? · What skills do people need to supervise, validate, and override AI? · What is decision simulation for AI adoption?
Rehearse the calls before your agents make them.
Decision Rehearsal shows how differently your leaders read the same agent decision, and builds one plan together.
Questions leaders ask
What does "human in the loop" actually mean?
It can mean four different things: a person who approves every action, one who reviews exceptions, one who sets boundaries but never sees individual actions, or one who reads a report afterward. Name which one applies to each agent.
Who should own an AI agent's decisions?
The business owner of the process the agent works in. IT can own the platform, but the person accountable for the business result owns the agent's decision rights.
How often should decision rights be reviewed?
Whenever an agent's scope, volume, or error rate changes, and at least quarterly for any agent that can spend money, change records, or contact customers.
Dr. Lisa Palmer
Founder of Neurocollective. Her doctorate produced the IRB-approved research behind AI Adoption Engineering, validated by external reviewers, and brought to life in Show AI—Don't Tell It: Build Buy-In with Visual Storytelling (Wiley, 2026). drlisa.ai