AI Human-Centered Design Product Strategy

What Should AI Own, Assist, or Leave Alone?

A practical framework for separating mechanical work from the decisions that still require context, accountability, and human judgment.

Human-Centered AI
Automation Strategy
Decision Design
AI Governance
Systems Thinking

The Problem

AI conversations often start with capability instead of responsibility.

The first question is often, “Can AI do this?” That is useful technically, but incomplete from a product and experience perspective.

A model may be capable of summarizing a case, recommending an action, assigning a risk score, drafting a decision, or routing work automatically. That does not mean each of those responsibilities should be transferred to the system.

The better question is what kind of work is happening underneath the task. Is it mechanical? Is it interpretive? Does it require context? Is there meaningful consequence if it is wrong? And most importantly, who needs to remain accountable for the outcome?

Questions to ask before adding AI

Is the task predictable or judgment-heavy?
What happens if the system is wrong?
Can the output be independently verified?
How much context does the decision require?
Who remains accountable?
Does automation remove friction or hide it?

The Framework

Own. Assist. Leave Alone.

I use three levels of responsibility to help teams decide how deeply AI should participate in a workflow. The goal is not maximum automation. It is the right distribution of work between people and technology.

01 / Own

Let AI handle the mechanics.

AI and automation can own tasks when the work is repetitive, predictable, easy to verify, and low-risk enough that failure can be detected and recovered from.

02 / Assist

Let AI improve the decision.

AI should assist when it can reduce preparation work, surface evidence, identify patterns, or suggest options while leaving interpretation and final ownership with a person.

03 / Leave Alone

Keep human responsibility explicit.

Some decisions carry too much ambiguity, consequence, ethical weight, or contextual nuance to delegate. Technology may support the surrounding work, but it should not own the judgment itself.

01 / AI Owns

Automate work that behaves like machinery.

The strongest candidates for ownership are tasks with clear inputs, repeatable rules, observable outputs, and limited consequence when something goes wrong.

Routing and Classification

Categorizing requests, identifying document types, routing cases, organizing content, and assigning standard workflow paths can remove significant administrative overhead.

Data Preparation

Extracting structured information, identifying missing fields, normalizing data, detecting duplicates, and preparing material for review are strong automation opportunities.

Routine Coordination

Status updates, reminders, handoff notifications, queue management, and other predictable coordination work often create friction without requiring meaningful human judgment.

Mechanical Verification

Checking completeness, known policy conditions, formatting requirements, or obvious inconsistencies can reduce review burden before a person engages with the work.

The Test

If the work is predictable and the failure is recoverable, ownership may be appropriate.

AI ownership should still include monitoring, exception handling, and a path for human intervention. “Automated” should never mean “invisible.”

02 / AI Assists

Use AI to make people faster and better informed.

This is where I see some of the most valuable AI opportunities. The system reduces cognitive and preparation load while keeping the person firmly inside the decision.

Summarize Context

AI can synthesize long histories, documents, interactions, or evidence into a useful starting point.

The person spends less time assembling context and more time evaluating what that context means.

Surface Relevant Evidence

AI can identify supporting documents, previous cases, policy language, trends, or anomalies that may matter to the decision.

Evidence becomes easier to find without allowing the model to become the evidence itself.

Suggest Next Actions

The system can recommend possible steps based on patterns, rules, or similar situations.

Recommendations become decision support rather than automatic decisions.

Identify What Deserves Attention

AI can flag uncertainty, unusual patterns, missing evidence, or higher-risk situations that warrant deeper review.

Human attention moves toward the places where judgment creates the most value.

03 / Leave Alone

Some decisions should remain unmistakably human.

There are places where the right product decision is not deeper automation. It is preserving human judgment while using technology to improve everything around it.

This is especially important when decisions involve meaningful consequences for another person, incomplete or contradictory evidence, ethical judgment, exceptions to policy, or contextual factors that cannot be safely reduced to a score.

AI may still prepare the case, summarize evidence, identify inconsistencies, or organize the workflow. But the consequential judgment should remain attributable to a person who understands the decision and can defend it.

Signals to preserve human ownership

The outcome materially affects another person
The evidence is ambiguous or incomplete
Exceptions and context matter
The decision requires ethical judgment
Accountability must remain attributable
A wrong answer is difficult to reverse

Designing for Risk

The higher the consequence, the stronger the human controls should become.

AI experiences need different levels of review and transparency depending on the consequence of failure. A typo in a generated summary is not the same risk as an incorrect eligibility recommendation or compliance decision.

Low Consequence

Optimize for speed.

Routine assistance can often move quickly when errors are easy to detect, easy to reverse, and have limited downstream impact.

Moderate Consequence

Require verification.

AI can prepare or recommend, but users should be able to inspect sources, understand confidence, correct mistakes, and explicitly confirm important actions.

High Consequence

Preserve accountable judgment.

The system should organize evidence and reduce administrative burden while keeping the consequential decision clearly owned and documented by a person.

Human-Centered AI Patterns

If AI participates in the decision, design the relationship.

Human-in-the-loop cannot simply mean putting an approval button after an AI output. The experience needs to help people understand, challenge, correct, and meaningfully participate.

Show the Source

Generated summaries and recommendations should connect users back to the underlying material so verification is fast and practical.

Expose Uncertainty

Confidence and ambiguity should shape the workflow instead of presenting every generated answer with the same level of authority.

Make Correction Easy

People should be able to correct extracted information, reject recommendations, add missing context, and recover without fighting the system.

Keep Accountability Visible

The workflow should make it clear what the AI produced, what the person reviewed, and who ultimately made the consequential decision.

Applying the Framework

Break the workflow into decisions before deciding where AI belongs.

Instead of labeling an entire workflow “AI-enabled,” I separate the work into smaller responsibilities and evaluate each one independently.

Map the Work

Identify the actual tasks, handoffs, decisions, and sources of friction.

Classify the Judgment

Separate mechanical work from interpretation and consequential decisions.

Evaluate Risk

Consider impact, reversibility, confidence, accountability, and failure states.

Assign Responsibility

Decide whether AI should own, assist, or remain outside each part of the workflow.

A Practical Example

Document review should not become one giant AI button.

Imagine a reviewer receives a lengthy document and needs to determine whether a case meets a complex set of requirements.

AI could own extraction, field matching, completeness checks, and document classification. It could assist by summarizing relevant evidence, identifying conflicts, surfacing policy references, and suggesting areas that deserve review. The final determination may still need to remain with the reviewer because context, exceptions, and accountability matter.

The technology can remove hours of preparation without pretending the preparation and the judgment are the same thing.

Example responsibility split

Own: Extract and organize information
Own: Flag missing or conflicting data
Assist: Summarize supporting evidence
Assist: Surface relevant policy
Assist: Suggest questions for review
Human: Make the final determination

The Principle I Come Back To

Automate the mechanics. Preserve the judgment.

The goal of AI should not be removing people from every workflow. It should be removing the work that keeps people from applying the context, expertise, and judgment we actually need them for.

AI & Complex Systems

Designing AI around accountability, trust, and real work.

This framework reflects how I approach AI inside complex products: start with the work, understand the decisions, evaluate the consequence of failure, and then decide where technology creates genuine leverage.

That keeps the conversation focused on product and human outcomes rather than simply maximizing the amount of automation inside the experience.

Explore AI & Complex Systems