Nudge AI
Designing a 0→1 clinical product for speed, control and trust
Overview
AI clinical scribe that clinicians could confidently sign off on
For many clinicians, seeing patients is only one part of the job. Every appointment is followed by notes, treatment plans, billing codes and updates to the patient record. A lot of that work happens between sessions or later at night, after the clinician has already gone home. Nudge was built to reduce that burden. It listens during a session and turns the conversation into a structured clinical note in around 30 seconds. The speed was useful, but it was not enough on its own. A clinician is ultimately responsible for every line that enters the medical record. If the output felt generic, difficult to review or even slightly unreliable, they would rewrite it themselves. That made this a particularly interesting design problem. I was not only designing how the product looked or how quickly someone could move through it. I was helping shape how clinicians understood the AI, checked its work and decided whether it deserved a place in their daily workflow.
Challenge
Building trust for non-tech users in the most sensitive sector - healthcare
Clinicians had good reasons to be cautious about AI. They work under constant time pressure, handle highly sensitive information and are personally accountable for the records they sign. A slightly awkward sentence in most software is a minor quality issue. In a clinical note, it can change the meaning of what happened during an appointment. The product therefore had to do more than generate accurate text. It needed to help clinicians review a note quickly, understand where suggestions came from and correct anything that did not sound right. It also had to fit around workflows they already knew instead of introducing another complicated system to learn. Privacy had to be just as clear. It was not enough for the product to be secure behind the scenes; clinicians also needed to understand what happened to the recording, how patient information was handled and what they should tell the patient before starting. The real challenge was not persuading clinicians that AI was impressive. It was making the product predictable enough to use on an ordinary, busy Tuesday.
Approach
Fit into the clinician’s day instead of redesigning it.
The product was designed around the workflow clinicians already followed. During the appointment, capture needed to stay out of the way so the clinician could focus on the patient rather than manage a recording tool. After the session, the generated note needed to be structured, easy to scan and ready for review. Once approved, it needed to move into the clinician’s existing records system without the usual copy-and-paste routine. Nudge also had to work across more than 50 specialties, each with different terminology, note formats and expectations. A behavioural health note does not read like a physiotherapy note, and two clinicians in the same specialty may document the same session very differently. The goal was therefore not to create one perfect template. It was to build a system that could adapt to the specialty, the format and, importantly, the individual clinician’s writing style. If the note did not sound like something they would write, they would not use it. They would edit every paragraph or start again which defeated the point of the product.
Research
Clinicians wanted less typing, not less control.
One of the clearest lessons from speaking with clinicians was that they were not looking for AI to take over the final decision. They wanted it to handle the repetitive first draft while leaving them firmly in control of the record. That distinction shaped much of the experience. The generated note had to feel editable and easy to verify rather than finished and untouchable. It also needed to reflect the clinician’s vocabulary, structure and level of detail instead of defaulting to a generic clinical voice. The same principle applied to billing and coding suggestions. A code presented without context asked the clinician to trust the system blindly. Showing the reasoning behind it gave them something they could assess. In a domain where the user remains accountable for the final output, a useful AI suggestion is not just an answer. It is an answer the user can understand, check and confidently approve.
Design solutions
Helping clinicians review instead of reread
The generated note was organised for scanning rather than presented as one long block of AI-written text. Clear sections and familiar clinical structures helped clinicians move quickly to the parts that required attention. Coding suggestions were shown with supporting reasoning, giving clinicians enough context to approve, change or reject them without retracing the entire session. The aim was not to make the AI appear certain. It was to make its work easy to inspect. That made reviewing the note feel less like proofreading a stranger’s writing and more like checking a capable first draft.
Keeping the workflow moving
The documentation flow followed a simple sequence: capture the session, generate the note, review it and transfer it into the existing health-record system. Each step was designed to require as little management as possible. The clinician could begin capture without interrupting the conversation, receive a note in their preferred format and move it into their records system with a single click. Removing the copy-and-paste step may sound like a small improvement, but it eliminated a repetitive task performed several times every day. Those small interruptions add up quickly, especially between back-to-back appointments. The best workflow improvements were often the least dramatic. They simply removed one more reason for the clinician to stay late.
Making privacy understandable
Clinical privacy could not live only inside legal documents and settings pages. The product needed to make its behaviour clear at the moments when clinicians and patients were most likely to have questions. Sensitive information was redacted during capture, recordings were automatically deleted, and patient data was not used to train the AI. The experience also included a clear moment for the clinician to explain that the tool was being used and ask the patient for consent. That turned a compliance step into a normal, human conversation: what the tool does, why it is being used and what happens to the information afterwards. The goal was not to cover the interface in warnings. It was to make the important safeguards visible without making the product feel frightening to use.
Less time finishing notes. More time left for medical reasoning.
Nudge is now used by more than 5,000 clinicians across a wide range of specialties. In a 10-week study involving an organisation of 250 clinicians, average documentation time fell from 20 minutes to 7 minutes per session. Across the organisation, that added up to nearly 300 hours of clinician time reclaimed. Coding accuracy reached 90%, compared with a national average of approximately 54%. Improved detection of eligible add-on codes also helped practices recover close to $10,000 in additional annual revenue per clinician. But the outcome that stayed with me was simpler. Clinicians spoke about leaving the office with their notes already finished. They no longer had to reopen their laptop after dinner or complete charts once their children had gone to bed. That was the clearest measure of whether the design was working. Not whether someone admired the interface, but whether a clinician could review the output, sign their name to it and get on with the rest of their life.
Duration
2023-2026
I've worked here as a passion project because I've always wanted to do something in healthcare, I did not take home a salary but worked with the handful of team to grow this into a scalable business.
