Medlog, an AI tool for dental practices, turns treatment documentation into billing. I redesigned it so the AI suggests and the people decide.
Dentists and assistants document every treatment, and the AI turns that documentation into billing positions for the patient's insurance. The first version was barely used. I interviewed the people who document every day, mapped the market, and redesigned the product under a strict brief: frontend only, no new features. Most of the gains came from taking things away.
I interviewed three dentists and two assistants, people who document treatments every day. Five was deliberate: in a first round, five users surface most usability problems, and we wanted to move fast and test again later.
I also reviewed eight AI documentation tools in Germany, the UK and the US, four with a full feature audit. Every one was voice-first or tablet-only. None gave structured reasons for raised billing factors. That told us where to compete instead of copying.
The pain was in templates, the record output and losing work. So we redesigned structure and flow first.
People didn't know they could stack several templates in one session. A real capability, invisible.
Voice recording felt strange to patients. It had to feel calm and in the practitioner's control.
Product set a strict brief: work with the existing backend and add nothing new. Every improvement had to come from structure, flow and removal. For every element we asked: does this earn its place? Ideas that needed backend work went to a later version, and I kept them in the design so the path stayed visible.
The AI drafts billing positions from the documentation. The practitioner accepts or rejects each one; nothing reaches the bill on its own. In the first release the bill appeared on a click, because the backend couldn't update it live. The end result updates as you document.
Catalogue, code, description, teeth, quantity and factor, read-only. Final edits happen in the practice management system, where billing staff already work.
Trade-off Less flexibility here, one source of truth there.
Two recording modes became one, because the audio pipeline was the same. No live transcript, because scrolling text is noise during a treatment. Instead: a reactive waveform, a timer, Pause and Finish.
Trade-off Less visible "AI magic", more calm in the room.
Every field is open at once, no accordions. Practitioners dictate, or type and move to the next field without reaching for the mouse.
Only filled fields are copied into the patient record. If a required field is missing, the copy button takes you straight to it instead of showing a warning banner.
Insurance decides which billing catalogue applies, so it's always visible as a colour on the patient number. Private and statutory patients get equal visual weight. No patient should look "premium".
On a phone, billing stacks under the documentation and can't be collapsed. Hiding it would only hide content people need. We checked the bottom bar against the new iOS Safari layout, because for many dentists Safari is their main browser.
The redesign is live in dental practices, including recording and dictation, and Medlog is now sold to external practices. I ran a task-based usability test on the redesign, recording completion, hesitations, errors and workarounds for each task:
The redesign is too recent for reliable usage data, so I won't quote a number. These are the signals that will show whether it worked:
The tooth chart. Many users asked to select teeth on the documentation page instead of typing codes. I designed it, but it needed backend work the brief ruled out. I kept it in the design with the evidence behind it, and it was adopted for the next version, which is launching soon.
The AI drafts the legal reason when billing staff raise a factor above the standard rate. No competitor offered this.
"For this treatment we'd usually expect five positions. We found four." Trust in the AI without a new billing engine.
Prior treatments and next steps during documentation. Only a product inside the practice system can offer it.
Designing AI for healthcare isn't about showing what the model can do. It's about deciding where it stops, so the people who sign the bill trust what they're signing.
Christy
