Case study / AI-enabled experience leadership
Orchestrating a UX practice.
Keeping judgment human.
Seven advisory specialists connected by evidence, explicit handoffs, and human-owned acceptance.
I designed an AI-assisted operating model for my UX practice: seven specialist roles, explicit handoffs, and a separate quality-review pass. The challenge was coordinating research, design, delivery, and quality while keeping evidence, responsibility, and acceptance visible. I decide where agent support fits, resolve tradeoffs across the work, and retain acceptance and release decisions with the appropriate human owners.
Separate creation from review.
Connect audience needs to design criteria.
Count complete effort and accepted work.
Operating-model design, experience direction, research judgment, and acceptance
Audience models, persona cards, design briefs, agent charters, audit records, and measurement method
Documented practice; working audience hypotheses; tracking accepted work and its complete effort remains pending
A documented design recommendation / Prototype inspection
Give a task one coherent home.
A navigation brief examined duplicate record-management destinations. It proposed a single task-focused home, while keeping account settings distinct. This is a concrete example of turning an inspected experience issue into a reviewable design choice.
- Problem
Competing destinations.
The inspected prototype offered overlapping places to manage records.
- Proposed change
Organize around the task.
Bring record-management activities together and keep account settings in their own context.
- Leadership responsibility
Make the tradeoff explicit.
Judge the task boundaries, decide the direction, and define what implementation and review must preserve.
- Evidence still needed
Test the proposed benefit.
The brief documents a recommendation. User comprehension, accepted implementation, and downstream benefit are not established here.
The operating model below shows how specialist responsibilities can support this kind of decision. It does not claim all seven agents participated in this particular brief.
What each capability helps me decide.
- Research
- Which needs, constraints, and assumptions deserve attention?
- Design
- Which flow, label, state, or navigation choice fits the task?
- Dashboard
- Where does the work need a decision or more review?
- Audits
- What should change, who owns it, and what needs rechecking?
- Value model
- Does capacity gained hold after complete effort and costs?
These are the intended decision benefits of the practice. Measured outcomes require further evidence.
The operating model
Specialists contribute.
Decisions connect the work.
The roles are personified AI advisory perspectives, not employees. I choose which roles a task needs and retain the evidence and acceptance decisions. The sequence below is a typical path, with revision loops as needed.
Meet the personified AI specialists. Their role profiles describe responsibilities, handoffs, and human checkpoints; the audience personas below represent service users.







- 01 / Frame
Riley + Piper
Organize evidence, surface uncertainty, define scope and acceptance criteria.
Riley Research · Piper ProductHandoff: evidence and a reviewable brief - 02 / Shape
Dana + Casey
Turn audience needs into experience alternatives, clear language, and interaction intent.
Dana Design · Casey ContentHandoff: proposed direction for my review - 03 / Build
Ethan + Piper
Translate the approved direction into a reviewable prototype and testable criteria.
Ethan Engineer · Piper ProductHandoff: implementation and traceable scope - 04 / Challenge
Quinn + specialists
Inspect behavior, content, and accessibility. Route findings back for correction and recheck.
Quinn QualityHandoff: findings, evidence, and a review decision - 05 / Learn & share
Morgan + Casey
Frame approved public work for its audience. Carry learning back to the next brief.
Morgan Marketing · Casey ContentHandoff: approved communication and open questions
Review & revision / Before work is shared
Findings come back into the work.
Dana and Casey revisit experience intent and language; Ethan proposes the implementation correction. Piper preserves the requirement and decision trail. Quinn rechecks the result, and Riley helps identify what still needs research.
My checkpoint: I decide what to accept, revise, reject, or investigate before approved work is shared. A review can return the work to an earlier step.
The dashboard is designed to connect evidence, work, review, and measurement across this workflow. Tracking accepted work and complete effort remains pending.
Audience models & design / Research questions to pursue
Give the work a person
and a task to serve.
Working archetypes distinguish people’s relationships to a service. Persona profiles and cards make goals, tasks, constraints, and access boundaries easier to examine together. Their value at this stage is making design assumptions explicit and directing research toward the gaps.
These audience models remain unvalidated working hypotheses. Agent critique and persona-based inspection do not establish findings from research with people.
The working audience model distinguishes record owners, verifiers, data consumers, builders, newcomers, and people acting on someone else’s behalf. Persona profiles and cards add goals, key tasks, constraints, and access boundaries to make those differences usable in design discussions.
A shared pattern of need.
Group people by their relationship to a service and the task they need to complete—not by a fictional biography.
Artifact type: audience model
A concrete working lens.
Add context, goals, tasks, constraints, and authority. Keep assumptions and source evidence visible.
Artifact types: profile and front/back reference card
A proposal we can question.
Translate a task or constraint into a flow, navigation choice, label, state, or acceptance criterion.
Artifact types: design brief and prototype instructions
A model that can change.
Check the hypothesis with relevant people and task evidence. Revise the model when observations conflict.
Validation remains distinct from agent critique
Sanitized illustrative adaptation / Not a research finding
One service. Different responsibilities.
The record owner
Task: Review and maintain information they are responsible for.
Design hypothesis: Make status, changes, and confirmation easy to understand.
A person acting for someone else
Constraint: Their authority may cover only certain people or actions.
Design hypothesis: Show who they are acting for and the limits of their access.
The artifact trail: Audience models inform design questions; a brief makes the proposed choice inspectable; criteria describe what the prototype must preserve. Research with relevant people is the next step for testing the assumptions.
Adapted to explain the method. Original project persona decks, internal research details, and product screens are not reproduced here. The personified agents are contributors to the practice; personas represent the people a service is intended to support.
Engineering handoff / From direction to a reviewable build
Make the next step clear.
I use the handoff to preserve the task, the reasoning, and the behavior the implementation must support. The aim is shorter clarification and revision loops between design and engineering; a measured delivery-speed benefit is not yet established.
A bounded brief.
Dana and Casey clarify interaction intent and language. Piper connects scope, dependencies, and acceptance criteria.
Artifacts: design brief, task flow, criteria
A reviewable proposal.
Ethan translates approved direction into a prototype, including responsive states, keyboard behavior, errors, and recovery.
Artifacts: prototype, implementation proposal, check results
A traceable correction.
Quinn records the finding, inspection method, owner, and recheck needed. The appropriate specialist proposes a correction.
Artifacts: audit finding, proposed change, verification record
A human checkpoint.
I review the experience with the appropriate product and engineering owners. Work can be accepted, revised, or rejected; recommendations do not approve their own release.
Artifacts: review decision and release boundary
The agent roles support the work; they do not replace accountable engineering or product owners. Acceptance of a prototype is distinct from authorization to merge or release production changes.
Program dashboard / Public demonstration
Put the next decision
in view.
I designed the measurement framework to help decide which workflows to continue, change, or stop. It considers complete effort, quality, sustained use, and service outcomes. This public dashboard demonstrates the decision views with illustrative records and synthetic calculator inputs.
Which assumptions need research? What direction is ready for review?
Inspect the audience-to-design method →What needs correction? What can proceed? Where is review effort accumulating?
Explore the illustrative audit records →Which workflow still offers value after production, review, rework, and tool costs?
Change the ROI assumptions →3 illustrative audit records · All await a review decision · No released correction is claimed
- Documented practice
- Role charters, audience models, design briefs, governance analysis, and audit records.
- Public demonstration
- Generalized examples, three representative audit records, and a synthetic value calculator.
- Measurement next
- Track accepted deliverables and complete effort before reporting measured ROI.
Governance audit / A separate control layer
Distinguish advice from automation.
Advisory roles
Human-invoked specialist perspectives propose work and challenge assumptions. Their recommendations remain subject to review.
Controlled processes
Monitoring, local refresh, and publication workflows have different access and action boundaries. A control matrix makes those distinctions inspectable.
Human decisions
Acceptance, consequential changes, and release remain explicit decisions. Proposed controls and evaluation sets are labeled separately from controls tested in operation.
The documented control review is a governance analysis, not an independent security certification.
Quality audit explorer / Representative examples
Make a finding actionable.
I use audits to compare interface behavior, design intent, and declared standards. A finding can call for a code correction, a design-artifact update, or no change when the two are aligned. The audit makes that decision and its owner explicit.
| Audit lens | Questions and artifacts |
|---|---|
| Design and implementation | Do task flows, navigation, states, and recovery match the intended experience? Record differences and the proposed design or code correction. |
| Accessibility and responsive behavior | Can people follow the task with a keyboard and across viewport sizes? Inspect semantics and rendered behavior; record the specific recheck needed. |
| Content and task clarity | Do labels, instructions, and feedback preserve the action’s meaning? Create a language proposal and identify comprehension questions. |
| Agent governance | Who can advise, change, approve, or release? Use a control matrix to distinguish permissions, human checkpoints, and controls still to evaluate. |
The framework connects each finding to its task risk, inspection method, correction owner, review decision, and verification needed. Code inspection, rendered checks, and research with people remain distinct.
The records below are representative examples of documented audit patterns. They do not report live inspection, accepted fixes, or released outcomes.
3 illustrative records
Illustrative audit record / accessibility
A question group has no accessible name.
A Yes/No choice can be announced without the question that gives it meaning.
- Inspection method illustrated
- Rendered-interface and semantic-grouping inspection, illustrated here.
- Proposed correction
- Group the choices semantically and expose the question as the group name.
- Responsibility
- Quinn reviews; Ethan proposes the correction.
- Review decision
- Pending human review.
- Verification needed
- Recheck the corrected experience against the stated task and inspection method before closing the finding.
Illustrative audit record / responsive and accessibility
Keyboard order leaves the task.
On a small screen, the next focus target can jump away from the visible task before the action controls.
- Inspection method illustrated
- Small-viewport keyboard walkthrough, illustrated here.
- Proposed correction
- Align focus order with the visual task sequence and verify it in the rendered interface.
- Responsibility
- Quinn inspects; Rob decides the intended experience.
- Review decision
- Pending human review.
- Verification needed
- Recheck the corrected experience against the stated task and inspection method before closing the finding.
Illustrative audit record / content
The same action has competing labels.
Different terms can make people question whether they are changing data, reviewing it, or confirming it.
- Inspection method illustrated
- Terminology review against declared task meaning, illustrated here.
- Proposed correction
- Choose terms based on the actual action and test comprehension.
- Responsibility
- Casey proposes language; Riley identifies research questions.
- Review decision
- Pending human review.
- Verification needed
- Recheck the corrected experience against the stated task and inspection method before closing the finding.
A separate AI review does not establish independent model validation. These neutral examples illustrate the structure of documented audit patterns. They contain no current product screenshots, tickets, participant information, or internal routes.
Dashboard and UX program / Transparent value model
Measure accepted work and complete effort.
The adopted method measures a deliverable accepted against stated criteria, rather than a prompt or ledger row. Compare like-for-like work and include all contributors’ production, setup, integration, coordination, review, corrections, and attributable downstream rework. Tracking this complete effort and piloting the method on accepted work remain pending.
| View | What matters |
|---|---|
| Delivery economics | Comparable baseline, complete human effort, incremental cost, and accepted output. |
| Quality | Review burden, corrections, rejected work, and defects after acceptance. |
| Sustained use | Repeated use relative to eligible opportunities. |
| Service outcomes | Task success, errors, support needs, and the experience people receive. |
A speed ratio cannot stand in for all four. Older dashboard totals use estimated baselines and are not presented here as verified financial ROI.
Explore the ROI assumptions.
Illustrative scenario only—not Rob’s measured results. Inputs describe one comparable deliverable accepted against stated criteria. Include unsuccessful attempts and all related human effort.
Include costs allocated to this work package, excluding labor already counted above. These are illustrative assumptions.
Count all contributors’ setup, integration, coordination, review, corrections, and attributable downstream rework here. Keep those hours separate from production time.
- Estimated capacity difference
- 5 hours
- Net estimated capacity value
- $725
- Estimated capacity-based return
- 93.5%
Labor-capacity model, not cash savings. Negative results are retained. No public performance claim is inferred from this scenario.
Illustrative values; change an input to compare assumptions.
Calculation and limits
Capacity hours = manual baseline − production − setup/review/rework.
Net capacity value = capacity hours × labor rate − incremental tool cost − other allocated setup/maintenance cost.
Capacity-based return = net capacity value ÷ (AI-assisted human labor cost + entered nonlabor costs).
The new components replace a complete legacy labor total; they must not be added to it again. Missing costs, unmatched scopes, or unaccepted outputs make a measured return incomplete. Zero cost denominators are unavailable, not infinite returns.
Documented contribution / Remaining proof
A practice people can inspect.
A result still to measure.
The documented contribution is an operating model connecting specialist responsibilities, audience and design artifacts, review decisions, and a complete-effort measurement method. It makes the work and its decision boundaries inspectable.
The next proof point is one accepted deliverable traced from evidence through drafts, rejected or revised recommendations, criteria, audit, my review decision, complete effort, and downstream use. That record is needed before claiming realized ROI or improved service outcomes.
Sources and evidence limits
This account draws on documented role charters, audience-model notes, design briefs, governance analysis, audit patterns, and the adopted measurement method. Audience models remain unvalidated; example audit records and calculator values are illustrative. Proposed evaluations and controls are distinct from checks completed in operation.
The saved interview narrative is a reconstruction, not a recovered verbatim transcript. No direct interview quotation is used here.






