Look for Friction Before You Look for AI

Do not look for AI. Look for friction.

Beginning with the technology encourages teams to search for places to insert it. Beginning with friction forces a more useful question: Which part of the workflow is slow, inconsistent, expensive, or difficult to review?

The distinction keeps novelty from becoming the objective.

Identify the friction

Describe the current workflow before designing the intervention.

• What decision or artifact does the workflow produce?
• Who owns the result?
• How long does the work take now?
• Where does rework occur?
• Which errors matter?
• What evidence is routinely missing?
• Which parts require judgment, approval, or sensitive context?

Friction should be observable. “The process feels manual” is not yet a baseline.

Map the decision

Separate preparation from accountability.

AI may help summarize source material, classify inputs, compare drafts, identify gaps, or produce a structured first version. A human owner should retain responsibility for factual accuracy, customer treatment, business trade-offs, and approval.

This is especially important in hiring, performance management, legal, privacy, and other consequential workflows. Assistance can improve preparation. It should not become an undisclosed decision-maker.

Insert AI at the narrowest useful point

Choose one step where the expected benefit and review standard are clear.

For example, an AI-assisted weekly update workflow might:

1. receive supplied project notes;
2. organize them into progress, plans, and problems;
3. flag missing owners, dates, measures, and asks;
4. return a draft for the project owner;
5. require the owner to verify the status and publish it.

The AI does not decide whether the project is green. It exposes the evidence needed for the owner to make that call.

Compare with the baseline

Measure the workflow, not the presence of the technology.

Useful measures may include:

• time to a reviewable draft;
• factual corrections required;
• missing inputs caught before publication;
• downstream questions or rework;
• actions captured with owners and dates;
• customer or business results affected by the workflow.

If the assisted version is faster but creates more review burden, the intervention may have moved the cost rather than reduced it.

Audit the failure modes

Review where the system guessed, omitted, overgeneralized, or followed untrusted instructions. Update the input contract, skill boundary, or human review step.

The next question is not “Where else can we use AI?”

It is “Did this change improve the decision, and what evidence supports expanding it?”