The user
Expert, time-poor, accountable
Clinical assessors writing lengthy reports after each assessment. The bottleneck wasn't judgement — it was turning judgement into fifteen structured sections of formal prose.
Case study ProblemShared · Clinical AI
A draft.
Never a decision.
Clinicians spent hours writing diagnostic reports. AI could draft them in seconds — but a clinician's name goes on the outcome. I designed a co-pilot that saves the hours without ever letting anyone sign something they didn't check.
One field, three states — AI generated draft
Summary of criteria A1 (DHI)
Differences in social-emotional reciprocity are evident across contexts, with reduced initiation of interaction and atypical response to social overtures.
AI generated draft
Review for accuracy
Regenerate
Mark reviewed
01 Context
The report is the product.
A ProblemShared assessment ends in a written diagnostic report — a legal and clinical document a named clinician stands behind. Writing one took hours of synthesising notes, questionnaires and observations into structured, defensible prose. AI was the obvious accelerant, and the obvious risk.
The temptation
AI that's good enough to trust blindly
A capable model produces fluent, plausible clinical text. That fluency is the danger: the smoother the draft, the easier it is to sign without truly reading — and a wrong diagnosis is not a typo.
The bar
Every word must be owned
Regulation and clinical safety require a human to be accountable for the final report. The design couldn't just speed writing up — it had to guarantee a human genuinely reviewed each part.
02 Research
I watched clinicians work, and worried about automation bias.
The literature on human-AI teams has one loud finding: people over-trust fluent automation. I designed the whole experience to fight that, and validated it against how clinicians actually behaved.
M1
Contextual sessions with assessors
Watching clinicians write reports the old way — where the hours went, which sections were formulaic, and where their expertise genuinely couldn't be shortcut.
M2
Automation-bias review
Studying how people rubber-stamp confident AI output, then designing deliberate friction so review couldn't be skipped without a conscious act.
M3
Prototype testing on real structure
Testing the draft-review-sign flow against genuine report sections with clinicians, watching for where they'd accept text without reading it.
M4
Compliance & clinical governance
Defining, with governance, what "a human reviewed this" must mean and prove — the requirement that shaped the entire interface.
01
A good draft got trusted too easily
The more fluent the AI text, the less clinicians scrutinised it. Fluency was being read as correctness — the classic automation-bias trap, live in a clinical setting.
So what
The interface had to make "AI wrote this" impossible to forget, and reviewing a deliberate, visible act.
02
"Reviewed" meant nothing without proof
Governance couldn't accept a report as clinician-authored unless the system could show, section by section, that a human had actually engaged with it.
So what
Review state had to be tracked per field and written to an audit trail — not a single checkbox at the end.
03
Editing needed to feel like ownership
The moment a clinician changed a word, they were taking authorship. The system had to recognise that instantly and reflect it, so the state matched reality.
So what
Typing into an AI field flips it out of "AI draft" the moment a human touches it. State follows behaviour.
03 The design
Three states, and a name on every one.
The whole system is one honest state machine, applied to every section of the report. It can't be gamed into signing unreviewed AI text.
State 01 · AI draft
Labelled, never hidden
Every AI-written field carries a visible "AI generated draft" tag and a "review for accuracy" prompt. The clinician always knows what they're looking at before they read a line.
No AI text is ever mistaken for a human's
State 02 · In progress
Touch it, and you own it
The instant a clinician edits an AI field, it flips to "in progress". Authorship transfers on the first keystroke — the state can never lag behind what actually happened.
Editing is authorship, and the UI knows it
State 03 · Reviewed
A deliberate, logged sign-off
"Reviewed" is a conscious action per section, written to an audit trail with who and when. Preview and export stay gated until every section has been owned.
"A human checked this" becomes provable
Product The report editor
Draft fast. Review honestly. Sign with proof.
One section-by-section editor. AI drafts on demand; the clinician reviews each field; the sidebar and audit trail track exactly what's been owned.
Richard Hendrix — Assessment report×
Preview & export
Developmental history assessment
Summary of criteria A1 (DHI)
Differences in social-emotional reciprocity.
Reduced social-emotional reciprocity is evident across home and school. Richard initiates interaction infrequently and responds atypically to social overtures, consistent with reports from both parents and teaching staff.
Reviewed
Owned by Dr A. Okafor · 14:22
Regenerate
✓ Reviewed
Summary of criteria A2 (DHI)
Differences in nonverbal communicative behaviours.
Eye contact is inconsistent and gesture use is reduced. Clinician note: cross-reference against the school observation before finalising —
In progress
You're editing — authorship transferred
Regenerate
Mark reviewed
Summary of criteria A3 (DHI)
Differences in developing and maintaining relationships.
Difficulties developing and maintaining peer relationships are reported, with a preference for solitary activities and limited imaginative play observed during assessment.
AI generated draft
Review for accuracy
Regenerating
Mark reviewed
Section audit trail
14:22Dr A. Okafor reviewed A1Reviewed
14:19Dr A. Okafor edited A2In progress
14:17AI drafted A1–A3AI draft
01
The label never leaves
AI text wears its origin until a human changes or reviews it. Fluency can't quietly pass as authorship.
02
State follows the cursor
Editing flips a field to "in progress" on the first keystroke — the record always matches what actually happened.
03
Sign-off is evidence
Every review writes who and when to an audit trail, and export stays locked until each section is owned.
Product Review anywhere
The same accountability, in a pocket.
Clinicians review between appointments. On mobile the report collapses to a checklist of sections, each carrying its state, so no one can export a report with an unreviewed field — regardless of device.
Report readiness — Richard Hendrix
Sections reviewed
11 / 15
ReviewedDevelopmental history · A1
In progressDevelopmental history · A2
AI draftDevelopmental history · A3
AI draftFormulation
9:41
Report reviewr-2291
Sections reviewed
11 / 15
In progressDev. history · A2
AI draftDev. history · A3
AI draftFormulation
AI draftRecommendations
Export locked · 4 to review
04 Trade-offs
Three calls where speed met safety.
Friction, on purpose
Per-section review is slower than one "accept all" button — and that was the point.
Fewer clicks⇄Genuine review
Decision
Keep the friction where accountability lives. Speed comes from drafting, not from skipping the read.
How loud to label the AI
Heavy AI branding builds distrust; hiding it invites automation bias. Neither extreme was safe.
Reassuring⇄Honest
Decision
Label clearly but calmly — a quiet, persistent tag that informs without alarming.
Regenerate vs. protect edits
Clinicians wanted to re-roll drafts, but a careless regenerate could wipe human edits.
Freedom to redraft⇄Protecting work
Decision
Regenerate only from AI-draft state; once edited, it warns before overwriting a clinician's words.
Process AI in the loop
I used AI to design the safe way to use AI.
There's a nice symmetry here: the tool I was designing was an AI co-pilot, and I prototyped it with one. That let me test drafting quality, review friction and edge cases against real report structure before engineering committed to anything.
01 · Prototype
A working state machine, fast
I built the draft → in-progress → reviewed field as a live prototype with Claude, wired to real generation, so the team could feel the automation-bias risk instead of debating it. Seeing a fluent draft you're tempted to sign is worth a hundred slides.
My judgementAI built the mechanism. Deciding review had to be un-skippable was the design.
02 · Pressure-test
Trying to break my own safeguard
I had Claude attack the flow: how would a rushed clinician sign without reading? Every loophole — bulk-accept, blind regenerate, export with gaps — got found in prototype and closed in design, not in production.
My judgementI chose which loopholes were unacceptable. AI made finding them cheap.
03 · Validate
Governance on something real
Because the audit trail existed in the prototype, compliance and clinical governance could react to a working artifact — confirming that per-section, logged review was enough to call a report clinician-authored.
My judgementThe prototype turned a policy debate into a design they could approve.
Me
Build the criteria field with three states: AI draft, in progress, reviewed. Typing must flip AI draft → in progress on the first keystroke. Export must stay locked until every section is reviewed.
Claude
Done. I gated Preview & export behind an all-reviewed check and log each transition with a timestamp. One risk: a clinician could hit "mark reviewed" without reading. Want me to require the field to be focused, or scrolled, before that button enables?
Me
Good catch — enable it only after the field's been opened. Now try to defeat it as a clinician who's running late.
Clinical
Per-section logging is exactly what governance needs to call it clinician-authored. The "AI draft" tag has to stay visible until it's edited or reviewed — that's the line.
Eng
State-per-field and the audit log are straightforward. Regenerate needs a guard so it can't overwrite edited text without a confirm. Buildable.
Reconstructed from the working sessions. Prototyping the safeguard let me stress-test accountability with clinicians and governance before it became code — which is the only responsible way to ship clinical AI.
05 Outcome
Hours saved, accountability intact.
Report drafting got dramatically faster
Clinicians started from a structured draft instead of a blank document, reclaiming hours per assessment
Every report stayed clinician-owned
Per-section review and audit trail satisfied clinical governance — nothing shipped unreviewed
Trust in the tool grew
Clinicians adopted the co-pilot because it made them faster without asking them to cede judgement
What I'd do next
- Surface the model's own uncertainty per section, so review effort goes where it's needed most.
- Let AI cite the source note behind each drafted sentence, making verification a click.
- Learn from clinician edits to improve drafting quality over time, with governance oversight.
- Bring the same draft-review-sign pattern to other clinical documents.
Rahman Malik — Product designer
ProblemShared · Clinical AI co-pilot