top of page

Before Building an AI Feature, Define the Decision It Should Improve

7 days ago
5 min read

Before approving an AI feature, define whose decision it should improve, what makes that decision difficult today, and how you will know the feature helped.

Summarizing documents, generating responses, and recommending actions describe capabilities. Their business value depends on what happens when someone uses them.

A convincing demonstration can show that AI produces an output. The investment decision requires a clearer connection between that output and the work.

Why I start with the decision

I have spent the past 20 years working in and around PMOs, beginning in a Product PMO. Across that experience, I have seen teams spend hours preparing a decision, wait for approval, and then return for more information.

That is one reason I focus on the AI Operating System, or AIOS, and improving how leaders make decisions. In Building an AI Native Company, I explain how that thinking is shaping NBD’s own operating model.

Consider a project change request.

When scope, schedule, or budget changes materially, someone must capture the proposed change, explain why it is needed, and assess its impact on all three. That can mean hours working with individual contributors to understand the implications, obtaining agreement from leads on the assessment, and writing the request for approval.

The request may then wait for the steering committee to meet or for specific stakeholders to become available. After all that preparation and waiting, the people responsible for approval may still lack enough information to decide whether to proceed.

The effort is real. So is the delay. The team may need another round of analysis and another opportunity to secure a decision.

This is the decision latency problem I discuss in Where Can AI Give Delivery Leaders Time Back This Week?: preparing information and reaching a decision are connected parts of the workflow.

Write the use case around the approval decision

“Generate a change request” describes an output.

The decision is more specific: should the proposed change be approved, declined, or returned for additional information?

The product brief should identify who prepares the request, who assesses the impacts, who has approval authority, and what each person needs to perform that responsibility.

It should also identify what remains outside the feature’s authority.

An AI feature for this workflow could help assemble an impact assessment, connect statements to supporting evidence, and flag missing information before the request reaches the approver. Those are proposed uses to evaluate.

The feature should distinguish documented impacts from assumptions and unresolved questions. A polished request that conceals uncertainty could leave the approver with a misleading picture.

A system that helps someone prepare a decision has a different acceptance standard from a system authorized to make that decision.

For this use case, approval remains with the designated stakeholders.

Establish the baseline before the demonstration

A persuasive demonstration shows what the capability can do. A baseline helps establish whether it improves the work.

For the change request process, examine:

1. The effort required to gather information and prepare the assessment. 2. The number of requests returned for missing or unclear information. 3. The time between a request being ready for review and an authorized decision. 4. The effort required to check and correct the completed request.

Use available operational evidence and record its limits. Separate active preparation time from time waiting for review or approval.

That distinction matters. If preparation takes hours but approval waits for the next steering committee meeting, faster drafting may have a limited effect on elapsed decision time.

Define the intended improvement before choosing a success threshold. The business should understand why the threshold matters and how it will be measured.

Include the complete workflow in the assessment. Time saved generating a request may be offset by time spent verifying its claims, correcting the assessment, or locating missing evidence.

Define acceptance around evidence and exceptions

Agree on what an acceptable change request must contain.

For this example, that includes the proposed change, its reason, the assessed impact on scope, schedule, and budget, and the supporting evidence. Assumptions and unresolved impacts should be visible.

Build an evaluation set that includes ordinary requests and meaningful exceptions. Examples could include a missing cost estimate, conflicting schedule assessments, an unconfirmed dependency, or an impact that the responsible lead has not validated.

Decide how the feature should behave in each situation.

If the available information does not support a conclusion, it should identify the gap and route the question to the appropriate person. It should preserve that uncertainty in the assessment.

Specify who validates each impact, who can access the underlying information, and when escalation is required. These requirements should shape the product from the beginning.

The acceptance question is whether the request gives the authorized stakeholder a dependable basis for deciding what happens next.

Choose a bounded first release

Begin with one project group and one defined category of change request.

Keep the existing approval authority explicit. Establish who checks the generated assessment, who resolves missing information, and who decides whether the request is ready for review.

Observe preparation effort, corrections, returned requests, and elapsed decision time. Record why requests wait so that an improvement in drafting is distinguishable from a change in stakeholder availability or approval cadence.

Agree in advance on the decision after the trial: expand, revise, continue learning, or stop.

The trial may show that AI helps prepare more complete requests while the main delay remains elsewhere in the approval process. That is useful evidence. It tells leadership which part of the operating model needs attention next.

Give AI product work the same ownership as other product work

The product leader needs to connect business value, user needs, delivery decisions, and operational acceptance. Technical feasibility is one input to that responsibility.

In the change request example, the product challenge includes understanding what approvers need, how evidence is validated, and where the process loses time.

Product Intelligence Foundation provides a structured way to clarify that problem, sequence the work, and establish acceptance conditions before a larger delivery commitment.

Before approving the next AI feature, ask the team to complete this sentence:

“This helps this person make this decision or complete this task, and we will know it helps when this measure changes.”

For the change request example:

“This helps the designated approver decide whether a proposed project change should proceed by presenting validated impacts and unresolved questions. We will assess its contribution through preparation effort, requests returned for missing information, and time to an informed decision.”

That gives the team a use case to test and leadership a basis for evaluating the investment.

About the author


Duane L. Nicholson is Founder and Chief Transformation Officer of NBD Consulting Services and a Fractional Executive & Strategic Advisor. He works with technology, product, and delivery leaders on modernization, operating model change, and strategic execution. His experience includes DTCC, Experian, Amgen, Toyota Financial Services, and Pacific Life. He is a SAFe Practice Consultant (SPC).


Comments


bottom of page