Tali Talks: From Requests to Reality: What It Takes to Build Tali

Welcome back to Tali Talks, our behind-the-scenes series where we sit down with the brilliant minds behind Tali AI to hear their take on the evolving healthcare AI landscape.

When people think about AI scribes, they often picture a magical tool that listens to a clinical conversation and spits out a perfect note. But ask Reza Takhshid, our Technical Team Lead at Tali AI, and you’ll hear a very different story.

"How hard can it be?" he laughed, when I asked him what makes building an AI scribe so difficult. "Then you get into it."

Here’s the thing: clinicians aren’t just another user group. They’re a deeply specialised, diverse, and demanding audience. And rightly so. They don’t have time to tweak generated notes or adjust phrasing, if an AI isn’t accurate or doesn’t match their existing style, it’s more of a burden than a benefit.

"You’ve got to get it really good and really right, but 'right' means a thousand different things for a thousand different users."

Reza’s team grapples with that daily. Even within the same specialty, clinicians have personal preferences, different workflows, and types of visits ranging from prescription renewals to complex follow-ups. Building one AI that works seamlessly for all of them isn’t just difficult, it’s like building a thousand different products in one.

The Burden of Choice: Templates, Prompts, and Personalisation

That’s why we built the Template Builder, a feature designed to give clinicians more control. But here’s the twist: while many clinicians ask for customisation, they often don’t have the time to do the customising themselves. Talking to an AI or typing out instructions can feel like extra work, not a time-saver.

"We tried to remove the need for prompt writing," Reza explained. "But even then, they prefer to talk to someone who sets it up for them. They don’t have the time to ask the AI to do it."

So we’re constantly walking a tightrope: build tools that are flexible enough to be customised but intuitive enough that no one feels intimidated using them.

Everything Changes, All the Time

Another challenge? The AI landscape itself.

Large language models are improving at lightning speed. While that’s exciting, it also means we need to regularly test new versions, and retrain our systems. Each change ripples through Tali’s backend and the user experience.

"Even if the new model is technically better, clinicians have already adapted to the quirks of the old one," Reza said. "They’ve built it into their flow."

And unlike traditional AI tasks that are easy to measure, like classifying images or extracting key data, AI scribing is subjective. The same transcript might produce five slightly different but technically valid notes. Measuring consistency and quality becomes more about interpretation than strict accuracy.

"It’s not numbers, it’s writing. You can’t just test it against a yes-or-no dataset. You have to interpret the outputs."

That’s why we’re working on new evaluation frameworks, tools that help us assess not just whether a note is factually correct, but whether it feels familiar, usable, and clinically trustworthy.

What Happens When You Request a Feature?

Let’s say a clinician tells us, "I want to see billing codes in my notes." That request might appear simple. But under the hood, it kicks off a complex chain reaction.

First, the team evaluates feasibility. Can we build it? How long would it take? What systems does it touch? Then we prototype, test with real users, refine the experience, and validate the outputs, all before it ever reaches your screen.

It’s not magic. It’s relentless iteration, a deep respect for clinical workflows, and a commitment to getting it right.

So the next time you see a new feature in Tali, or find your notes flowing just the way you like them, know that it took dozens of decisions, hours of testing, and some very clever engineering to get there.

So what’s the takeaway?

AI scribes aren’t hard because of the AI. They’re hard because of humans. Because workflows are complex, preferences are personal, and trust is fragile.

Reza reminded us that success doesn’t just come from great models or good UI — it comes from empathy, iteration, and relentless attention to the edge cases.

Thanks for reading this edition of Tali Talks. We’ll be back soon with more insights from the brilliant people shaping Tali AI, and healthcare, behind the scenes.

Got a topic or person you’d like to hear from next? Let us know. We’d love to hear from you.