Updated October 14, 2026
When AI enters a workflow, define who may act on its output and under what conditions. A recommendation becomes consequential when someone uses it to allocate work, approve a request or change a service. The decision process should make that responsibility visible.
Start with one recurring decision. Write down the outcome it is meant to support, the people affected and the consequences of getting it wrong. Then design the following elements around that specific choice.
Give the decision a clear owner
Name the person or role authorised to make the decision. Identify who supplies evidence, who must be consulted and who implements the result. Where an independent approval is required, make that authority explicit.
Define the limits of delegation. A team may be able to change an internal queue while needing additional approval to change a customer commitment. Agree on an escalation route before a disagreement or unusual case occurs.
Set the evidence standard in advance
List the information needed to make the choice and distinguish observed facts from estimates and assumptions. Check the source, date, relevance and gaps in important evidence.
Compare realistic options, including keeping the current approach or running a limited test. Ask what evidence would change the preferred option. Scale the review effort to the potential harm and difficulty of reversing the decision, with a deadline appropriate to the situation.
For an AI-supported analysis, require the reviewer to inspect the underlying material for material claims. Record where the system’s output could not be verified. A confident explanation should never substitute for an adequate basis to act.
Make human review a defined task
Specify what the reviewer must check, which exceptions require referral and whether the reviewer can reject or pause the proposed action. Provide the evidence and time needed to exercise that authority.
In a hypothetical service-request workflow, AI might suggest an urgency category. A meaningful review would let staff inspect the original request, correct the category and escalate ambiguous cases. The operating plan should also address missing information and requests that the tool cannot process.
The NIST AI Risk Management Framework offers a voluntary reference for considering trustworthiness during the design, use and evaluation of AI systems. The practical arrangements still need to fit the actual workflow and its consequences.
Keep a concise decision record
Capture enough context for someone else to understand the choice later:
The decision, date and accountable owner.
The options considered and evidence used.
Important assumptions, uncertainty and unresolved concerns.
The selected action and the reason for choosing it.
Implementation responsibilities, review date and conditions for reconsideration.
If AI materially informed the decision, record its role and the relevant system version or configuration where feasible. Link to supporting evidence with appropriate access controls. Keep sensitive information only where it is needed and properly protected.
Check what happened after the decision
Set the review before implementation. Compare the intended outcome with actual results, including exceptions, rework and effects on the people using the service. Separate a weak decision process from an unexpected outcome; either may require a different response.
Decide whether to continue, adjust, pause or reverse the change. Update the decision record so the next team can learn from the result. This creates a practical connection between authority, evidence and accountability throughout the work.
Discuss decision rights and AI governance with Daly Ventures.
