← back to work case study · service design

Five tangled boards, read as one lifecycle.

A seven-person art studio ran on five workflow boards that barely connected. I mapped the real flow, rebuilt it one department at a time, and found where a self-hosted AI could carry the busywork without touching the teaching.

Client A small, China-facing art studio that prepares students' portfolios for university. Identity withheld; operational details abstracted.
Sector Creative education, student services
My role Service designer. Research, service blueprinting, and the AI systems architecture.
Team Solo, working with the studio's director and its department leads.
Duration 2025. An as-is read, a to-be redesign, and a phased build plan.
Methods Workflow-board analysis, service blueprinting, stakeholder mapping, an adversarial design-critique loop, and AI systems design.
Outputs An as-is map, a to-be service blueprint, a self-hosted AI architecture, a dashboard per role, and a phased build roadmap.
Status A real engagement, mid-build. Studio anonymized pending its sign-off; sensitive operational specifics abstracted.

The studio had drawn how it works. Almost nothing connected.

The studio had already done the honest part. Five large workflow boards, one for each side of the business: how a new student enrolls, how a student gets coached, how money moves, how the studio finds the next student. Each board was detailed and real. Side by side, the boards barely touched.

A student's name was typed fresh into a new board at every stage. A handoff between two teams was a message someone remembered to send. The same approval waited on the same role in five different places. None of the boards were wrong. They were drawn as five separate machines, and the studio is not five machines. It is one studio that a student moves through, and the gap between those two facts was where the work kept getting lost.

Five boards were one lifecycle. That was the whole unlock.

I stopped treating the boards as five processes and traced the one thing that runs through all of them: a single student. They arrive as a lead, sit a consultation, get assessed, decide, enroll, get onboarded, get coached for months, apply to schools, and graduate. If they do well, their story becomes the marketing that brings in the next student. One arc, not five.

So I fixed that arc as a spine and held it still, then rebuilt each department against it, one at a time. The order mattered as much as the rule. Cut the steps that should not exist, simplify what is left, then automate, and only then connect the systems. Automating a broken step only makes a broken step faster.

Before and after, in one picture.

The point of the redesign is the difference between these two shapes. On the left, the old one: five boards, work crossing between them by hand, a student re-entered at every boundary. On the right, the new one: a single record that moves through states, where each step hands to the next on its own.

As-is · five machines

board board board board board

Manual handoffs. Dashed lines are work moved by hand, and the gaps are where it fell through.

To-be · one record, moving

attract onboard develop result advocate loops back

Automatic handoffs. One record through states; the last stage feeds the first.

Figure The same studio, redrawn. The redesign's whole value is the move from the left shape to the right one.

I designed each department, then tried to break it.

A redesign that only its author has read is a guess. So every department went through the same loop. I drafted the new flow, then handed it to an independent reviewer whose only job was to find what was wrong with it. The reviews were uncomfortable and worth it.

One pattern kept surfacing. I would call a bottleneck solved when I had only moved it. Approvals for nearly everything funneled to one role, so I routed the routine ones away, then noticed I had rebuilt the same queue on a different screen. The fix was a rule that defines which decisions actually need that person, and a number on their own dashboard that shows when they are the thing holding work up. I kept the critiques in the work instead of hiding them. The second draft is the real one.

The to-be service, with AI drawn in as an actor.

The redesign is one blueprint. Across the top, the four phases a student moves through. Down the side, the people and systems that serve them. Every backstage step carries one tag, so the shape reads at a glance: what stays human, what runs on its own, and what an AI drafts for a person to approve. The dashed rules are the blueprint's own lines, the borders between what a family sees and what they never do.

Unless tagged, a backstage step is AI-assisted: the model drafts, a person approves. human judgment, kept human auto runs on its own, reversible
Attract & convertinquire → enroll
Onboardset up
Develop & delivercoach → apply
Advocateloops back
Student
Inquire, decide, enroll
Join the platforms
Coaching, upload work, apply
Share results, refer a friend
line of interaction
Frontstage
Consultation & assessment human
Orientation human
Coaching & portfolio critique human
Results follow-up
line of visibility
Sales
Triage, proposal, contract
·
·
Referral intake human
Operations
CRM record & scheduling auto
Onboarding pack
Record state machine auto
·
Education
Assessment write-up
Development plan
Notes, reports, essay feedback
Case-study draft
Marketing
·
·
·
Content & publishing
line of internal interaction
AI layer
recordsdocumentsmessagesscheduling Self-hosted models behind one gate: anything that sends, pays, or publishes waits for a person.

→ scroll the blueprint sideways on a narrow screen

Figure The to-be service at phase altitude. The moments that earn the fee stay human. The busywork beneath them moves to AI that a person still signs off. Finance runs underneath every phase from enrollment on, and the director sits above it, handling the exceptions.

AI carries the busywork. It never carries the judgment.

The studio already used AI in the corners, informally. The redesign turns that into a system with one rule at its center: a person approves anything that sends, pays, or publishes. The model drafts the proposal, the progress report, the invoice, the reminder. A human reads it and commits it. Nothing irreversible fires on its own.

The work that touches a student's records runs on a model the studio hosts itself, on its own infrastructure, by design. That keeps a young person's data where it belongs and where the law expects it to be. The teaching, the critique of a portfolio, the read on whether an idea has landed: those were never on the table to automate. They are the product.

Role dashboards

One screen per role. The few numbers that drive an action, plus the one-click drafts that role needs.

The gate

Every send, payment, or publish waits here as a draft with a preview. A person approves, edits, or rejects. A log records who did what.

Connectors

Narrow, least-privilege links to the systems the studio already runs. The model reads freely and writes only through them.

recordsdocumentsmessagesschedulingsocial stats
Models

Self-hosted for anything touching a student's data, by design. A hosted model only for the rare hard problem on non-sensitive work.

Figure The system, top to bottom. The gate is the whole safety model: the machine proposes, a person commits.

Every role gets one screen that says what to do next.

Each team gets a dashboard scoped to the decisions that role actually makes. Every figure has to answer one question: what do I do about this? The ones that cannot, I cut. A few carry a quiet signal the studio needed without having asked for it, like whether the work is creeping toward the limit of how many students the mentors can take.

Figure The sales view. Urgency reads through the accent and weight, not a traffic-light palette; the starred signals tell the studio whether it is short of leads or short of capacity. Every other role gets the same kit, tuned to its own decisions.

The mistake I almost shipped.

The loudest signal in the studio was a single note on the intake board: too many leads. I read it as the problem and started designing a faster funnel. The reviewer stopped me. A faster funnel only helps if the studio can teach the extra students it converts. If the mentors were already full, converting harder would fill a longer line and make the experience worse, which is the opposite of what this kind of studio sells.

So I held the redesign and asked the one question only the director could answer. Are the mentors full? They were not. The studio had room to grow, which made the intake work the right call after all. But I had been about to bet the engagement on a guess. The lesson stuck. The loud symptom is not the constraint until you have checked.

What I handed over.

This is a real engagement, and it is mid-build, so I will not claim numbers I have not measured. What exists is the design. An as-is map the studio recognized as itself. A to-be blueprint that holds together across every department. A self-hosted AI architecture with a human gate on every risky action. A dashboard for each role. A build plan sequenced so the cheapest, safest wins come first and earn the trust for the rest.

The reframe held: the studio is one lifecycle, and seeing it that way changed what every department was for. The principles held too, and they are the part I would carry anywhere. Do not call a bottleneck solved when you have only moved it. Check the loud symptom before you build for it. Let a person keep the judgment and let the machine keep the busywork.

What I would do differently.

I trusted the loud signal too long before I tested it, the same habit I am still unlearning from years of teaching. I have gotten faster at treating the first clear sign as something to test, and the reviewer caught this one before I did. I would also sequence the build the other way around. I designed the visible parts first and left the unglamorous prerequisites for last, when a real build has to run in reverse. The studio needs a clean place to keep its money before any of the finance work is worth drawing.

My first real service-design project taught me that the craft is mostly restraint. The interesting move is rarely the new automation. It is deciding what to leave exactly as it is.