Five decisions that should never disappear inside an agent
Keep outcomes, evidence authority, action boundaries, uncertainty, and ownership visible.
Agentic systems compress work. That is part of their value. A request that once moved through research, comparison, drafting, and coordination can arrive as a single recommendation.
Compression has a cost: the decisions inside the work become easier to miss.
A polished result can hide that the system chose a source, resolved a conflict, interpreted a rule, accepted uncertainty, and selected an action. If those choices matter to the organization, they need explicit owners and visible evidence.
Imagine a fictional People team using an agent to prepare learning-reimbursement requests. An employee submits a course, cost, date, and receipt. The agent reads the approved policy and drafts a recommendation. Five decisions still belong to the organization, even if the agent handles most of the preparation.
Here are five decisions that should not disappear inside an agent.
1. What outcome are we optimizing?
“Be helpful” is not an operating objective. Speed, cost, consistency, employee experience, risk, learning, and quality can pull in different directions.
The system can optimize a stated outcome. It should not quietly choose the tradeoff.
The product brief should name the primary outcome, guardrails, baseline, and the measure that would cause the team to stop or revise the work.
2. Which evidence is authoritative?
Retrieval is not governance. A system may be able to access twenty documents and still not know which policy supersedes another, whether a local exception applies, or whether a source is current.
The organization must define approved sources, source owners, effective dates, conflict rules, and what happens when evidence is missing.
The agent can assemble the evidence. Authority remains an organizational decision.
3. When does a recommendation become an action?
There is a material difference between drafting, recommending, deciding, and executing. Products often cross those boundaries gradually because each incremental automation looks convenient.
Name the approval gate. Name the person or role with authority. Show them the evidence and uncertainty. Record their disposition. Define whether any low-risk cases can proceed automatically—and who approved that policy.
“Human in the loop” is too vague unless the human has time, information, authority, and a meaningful way to intervene.
4. What uncertainty is acceptable?
Every useful system encounters missing, conflicting, or unfamiliar cases. The design decision is not whether uncertainty exists. It is how the product behaves when it appears.
The organization must set thresholds for clarification, escalation, refusal, and manual fallback. A confidence score alone is not a policy. It needs calibration, a consequence, and an owner.
The agent can express uncertainty. The organization decides what uncertainty it will accept for a given consequence.
5. Who owns the system after launch?
An agentic workflow has at least four kinds of ownership: service performance, source information, model behavior, and business risk. Giving all four to “the AI team” usually means the actual operating function has not accepted responsibility.
Name who can change the workflow, approve new sources, review failures, respond to incidents, and shut the system down. Define the evidence reviewed at each cadence.
Ownership is what turns a build into a capability.
A simple design review
Before approving an agentic workflow, ask the team to put these five decisions on one page:
- outcome and guardrails;
- evidence authority;
- action boundary;
- uncertainty policy;
- operating ownership.
For each decision, require an owner, the current choice, the evidence supporting it, and the condition that would trigger reconsideration.
For that reimbursement workflow, a one-page decision record could look like this:
- Outcome: Reduce back-and-forth on complete requests without increasing mistaken approvals.
- Evidence: Use the current policy owned by People Operations; show its title and effective date.
- Action boundary: The agent may prepare and cite a recommendation. A designated person approves or rejects payment.
- Uncertainty: Missing receipts or conflicting policy passages go to a person; the agent does not guess.
- Ownership: People Operations owns the policy; the service owner reviews failures and changes the workflow.
Those are illustrative choices, not universal reimbursement rules. Their value is that a builder, reviewer, and employee can see what the agent is—and is not—authorized to do.
If a box contains “the agent decides,” ask whether the organization has consciously delegated that authority or merely failed to notice it.
The best agentic systems do not remove organizational judgment. They make judgment more focused, better informed, and easier to examine.
The BUILT approach helps leadership teams decide which work to change, who owns each decision, and what evidence will show that the new way works better.