Saara
Saara is an AI documentation platform for occupational health teams. A fitness-for-duty visit becomes a signed form, a work disposition report, and an employer letter.
Overview
I joined Saara when the product was an early prototype. I started with its design language, then became the primary builder as we focused on occupational health.
Our first design partner, an occupational-health provider group, needed ten capabilities for a July alpha. I mapped their workflow end to end, from the exam room to the letter an employer receives.
Ask Saara records the visit, turns it into a structured note and fills the form, showing where every value came from. The provider confirms it’s accurate and signs. The employer letter is drafted from that signed form, never the raw recording.
A fitness-for-duty exam ends in three documents that share most of their facts, so one verified form could produce the rest. If the visit itself filled that form, providers would review instead of write.
On my whiteboard, I mapped the letter flow from the clinic’s template to a letter that names itself.
A fitness-for-duty (FFD) exam decides an employee’s work status and any restrictions. Each one ends in three documents: a detailed form, a work disposition report with that outcome, and a letter for the employer. Providers typed them all up long after the exam.
Any AI filling them in had to be trustworthy: one wrong restriction on an employer letter is a liability.
Each gap in the table names the part of Saara that closes it.
Providers wrote all three documents by hand after the exam.
Each document retyped facts the others already held.
Trust broke fastest when the AI said it did something it hadn’t.
Clinic staff started without instruction. One “what next?” choice kept them moving.
A separate clinician review went unused, so reviewing and signing had to be one step.
The work splits into six features. Pick one to see its final design, how each flow got there, the prototype and the research.
01 / 06
I designed and built Saara's core loop for fitness-for-duty visits: record from chat, get a drafted note, fill the work disposition report with every value sourced, then attest, sign and lock it.
Review and sign, at iPad width
Each fitness-for-duty exam left the provider three documents to type by hand, long after it ended: the detailed form, the work disposition report and the employer letter. Any AI filling them had to earn trust, because the provider signs for every value.
01Record the visit, pick the form, sign it
02A note headed by what was discussed
03A correction as a signed addendum
04Ready to sign, in dark mode
Each piece numbered beside the component, then where else it is used.
Where it’s used
Earlier rounds, oldest first, then what shipped and any later pass. Open any screen full size, or step through the flow.
Flow 1 of 6
Moving to any other page in Saara ended a recording in progress.
How does the provider start recording in the exam room, tied to the right employee, and keep it running while they check the chart?
What landedAsking Saara to record opens a recorder card in the chat, with the employee already picked and an occupational visit reason. The transcript builds live. Leave for the chart or schedule and a floating pill keeps recording on every page; click to open it, or stop from anywhere.
Flow 2 of 6
The first version drafted a SOAP note (Subjective, Objective, Assessment, Plan) into four fixed boxes.
How does a long, unstructured transcript become a note shaped like the visit, one the provider can correct without Saara overwriting their edits?
What changedThe note drafts itself from the transcript when the tab first opens, headed by what was discussed, here job demands and the exam. Regenerate asks before replacing anything, nothing saves until Save notes, and form fills and letters read this note.
Flow 3 of 6
Our bug list flagged a dead “0 fields” row in the chat's form picker for forms nobody had mapped yet.
Once the visit is saved, how does the provider reach the right form in one move, and know up front which forms Saara can fill?
What landedEvery saved recording offers one next step: Fill out a form with this recording. The picker already holds the recording, so it asks only which form. Each form shows how many fields Saara can fill; a form with none stays selectable as a manual fill, completed by hand in review.
Flow 4 of 6
Trust broke fastest when the AI said it did something it hadn't.
When Saara fills a work disposition report, how does the provider see which values came from the profile, which from the visit, and what's left?
What landedEvery value names its source: Profile · Date of birth, Recording · job duties. Violet marks profile values, blue the recording. Anything Saara couldn't map stays empty and amber, and the header counts it: 20 of 22 filled, 2 to complete before saving. Dates and vitals always show as written.
Flow 5 of 6
Nurse practitioners sign these reports, and one wrong restriction is a liability.
How do they complete the blanks, own every value, and lock the report?
What landedAmber fields are edited right on the form. Sign & lock asks for the signer's name, credentials and an explicit attestation that it's accurate; Sign & apply stays off until it's ticked. The signature is stamped with the time, and the signed form keeps its values when reopened or downloaded.
The parts
Flow 6 of 6
Once signed, the report joins the employee's record.
How does the provider know it landed, and fix a mistake later without rewriting what they signed?
What changedSigning ends on Form complete, confirming the report is signed and saved to the patient chart, with a thumbnail of the signed PDF, the signer, and View form in patient profile. A later correction is an addendum, a new signed, dated page, never an edit to the signed content.
The parts
Ask Saara to record the visit. The transcript builds live, and a pill keeps it running on any page. Stop, and Saara drafts the visit note. Pick a form and it fills from the recording and the employee's profile. Check where each value came from and complete the amber fields. Attest and sign. The locked report lands on the employee's chart.
I tested with our design partner and a clinic tester, worked through their QA punchlists and our bug list, and reviewed builds on my whiteboard. The flow only clicked once we broke it into smaller steps.
Clinic staff started without instruction. A long step-by-step flow lost them; one “what next?” choice kept them moving.
Trust broke fastest when the AI said it did something it hadn't.
In partner QA, date and number inputs silently hid values that didn't match their strict format, so filled vitals and dates looked blank.
Moving to any other page in Saara ended a recording in progress.
Nurse practitioners signed the reports, so a separate clinician review went unused. QA found the signature box silently disabled when they lacked a signer role.
Our bug list flagged a dead “0 fields” row in the chat's form picker for forms nobody had mapped yet.
Looking back
Chat became the front door: in the last 90 days, people started nearly three Ask Saara conversations for every form. And Saara did the work, not just the talking: none of 638 tool calls returned an error.
Every number here comes straight from the product database and our sign-in provider.
| Period | conversations |
|---|---|
| Jul 13 | 30 |
| Jul 20 | 57 |
| Jul 27 | 93 |
| Aug 3 | 134 |
| Aug 10 | 160 |
| Aug 17 | 177 |
| Aug 24 | 200 |
| Aug 31 | 225 |
| Sep 7 | 236 |
| Sep 14 | 241 |
| Sep 21 | 243 |
| Sep 28 | 252 |
| Oct 5 | 255 |
| Oct 12 | 257 |
Let users fall in love with one thing that works properly.
What I learned
Our partner’s own users came back to the record, fill and sign loop in 7 of 11 weeks. Next: make a recording impossible to lose, bring trial users back for a second week, and measure time saved per exam.
Teams used Ask Saara to look things up and the forms page for form work. So chat should hand off to structured screens, not replace them.
In testing, trust broke fastest when the AI said it did something it hadn’t. So Saara says “I can’t do that yet” instead of guessing, and names the source of every value.
