Lessons on working with AI in filmmaking
What nearly two years of experimentation taught our team.
At Noora Health, we believe technology exists to extend care, not replace it, whether that’s training families in hospital corridors or supporting caregivers through our Mobile Care Companion after they return home.
This belief shapes not only how we deliver care, but also how we tell stories.
We make health behavior change films which are played in hospital waiting rooms, sent through our mobile service, watched by patients and caregivers across India in their own language, often at the most vulnerable moments of their lives. Every choice we make, whether a medical visual, the pronunciation of a word in a regional language, or an emotional beat, affects whether they trust (and act on) what they’re watching.
About two years ago, I asked the team of filmmakers I lead to start experimenting with generative AI.
The idea was never to replace people or make films more cheaply, but to answer a different set of questions: Could AI streamline our production processes? Could we reach more people in more languages, without compromising medical accuracy, cultural nuance, or ethical storytelling? Could AI carry the time-consuming, repetitive load of post-production correction work such as subtitling and dubbing, so that the team could start spending them on medical review, scripting, and audience research?
This isn’t a manifesto for or against AI in filmmaking. It is my attempt to write honestly about what happened when we brought it into our creative process: the fears, the failures, and the wins.
The risks of using AI in health communication (and the fears no one talks about)
My first fear had nothing to do with technology. It was about the people in the room.
I lead a talented team of scriptwriters, animators, and editors; people who chose this work because they love their craft. Suggesting that we experiment with AI felt like a betrayal of that, almost as if I was telling them that their skills weren’t good enough anymore.
My second fear cut deeper. In health communications, an incorrect medical representation or a mispronounced word isn’t a craft slip. It’s a breach of trust with someone making a real decision about their own or a loved one’s health. And most AI models are trained on generic rather than specific medical data points. The models also don’t know the dialect of a village in Odisha, or what care actually looks like in an Indian household in Karnataka.
If we don’t get the context right, even a well-made film may never reach the people it was meant for.
The third fear was about building on shifting ground. AI platforms change every week, adding and retracting features faster than any team can write a stable process around them. As a leader, you are supposed to give your team certainty. Instead, I was asking them to stand on something that was constantly evolving.
These fears shaped how we moved. Instead of rushing to adopt every new tool, we tested, discarded, and refined — slowly enough that AI earned its place in our workflow rather than being forced into it. Looking back, that measured pace is what gave me confidence we weren’t chasing technology for its own sake.
Some of our biggest lessons came from what didn’t work
Medical accuracy in AI-generated illustrations was hard to achieve. Even highly detailed prompts produced visuals that looked convincing but contained anatomical or clinical errors. We learnt that prompting alone can’t guarantee accuracy — anything a patient sees still needs human review.
Direct translation from one language to another was another challenge. Sentence length, rhythm, and mouth movements differ across languages, so perfect lip-sync often looked unnatural. We chose to prioritize medical clarity and natural language over visual perfection.
AI voice generation also struggled with authentic emotion, regional pronunciation, and conversational rhythm. Getting a patient-facing tone often took multiple iterations.
These failures were a reminder: AI can speed up production, but accuracy, authenticity, and trust still depend on human judgment and medical review.
Out of these learnings, four principles emerged that continue to guide how we use AI in our work.
1. Focus on what the tools enable, not how many you use
We defined success not by the number of AI tools we adopted, but by what they enabled us to do. How was AI integration helping us optimize the filmmaking process? Did it free up our team to focus on higher-value creative work? Are we witnessing efficiency gains?
The results were tangible:
- Faster delivery, without cutting corners: We began with lower-risk, skills-based content, such as training and how-to videos, where an AI misstep was less likely to compromise the trust of a caregiver. AI-assisted scripting, editing, and production reduced timelines from roughly three months to one. Only after learning from these experiments did we gradually introduce AI into medical post-production for voice, visual asset generation, and footage correction. Our video introducing the Mobile Care Companion shows this in action.
- Reducing physical production load: Some medical parts of the videos were created without a physical shoot. AI-generated presenters (avatars), environments, and supporting visuals reduced the need for multi-day shoots, actor coordination, and location scouting — saving considerable production time and effort. For health films specifically, about 50% of production now happens through real shoots, with the remaining 50% achieved through animatics and AI-generated B-roll. Our films on chronic lung disease and stroke are good examples of this approach.
- Better value for every rupee invested: AI-powered dubbing, voice cloning, and lip-sync significantly reduced costs, bringing our dubbing expenditure down to zero.
- Faster adaptation: The same tools also made AI video localization for Indian regional languages operationally feasible, allowing us to produce state-specific content in Bengali, Hindi, and Kannada. Our early cancer screening videos were contextualized for an Assamese audience in Bengali, using direct language-to-language AI translation.
But ultimately, what actually changed my mind was watching my team stop losing entire days to repetitive post-production fixes, and start spending that time on work they actually enjoyed. If anything, work started feeling more human, not less.
How AI helped us make our diabetes medication adherence film: HeyGen for graphics and B-roll generation and ElevenLabs for a post-production voice fix and lipsync that kept the message medically accurate.
2. Test broadly, then commit narrowly
We learned early that too many tools create more complexity than they solve. So we tested more than 13 AI platforms, including Kling AI, PixVerse, Grok, Flow Adobe Firefly, Gemini, and Perplexity. From there, we settled on the following four that consistently met our standards for quality, speed, multilingual delivery, and medical accuracy:
- HeyGen — AI avatars, lip-sync, and video translation to regional languages
- ElevenLabs — Voice cloning and dubbing across learning modules, medical videos, and reminder messages for patients and caregivers
- Google’s generative video and image tools — AI B-roll, backgrounds, and medical environment visuals blended with real footage
- Higgsfield — semi-realistic video, B-roll, and medical environment generation
3. Flexibility is just as important as capability
No AI tool is perfect, and pretending otherwise only creates frustration. We learned to design our workflows around AI’s strengths while planning for its weaknesses.
- Avatar quality still has limits: AI avatars and photo-to-video outputs depend heavily on high-quality inputs, and body-gesture control remains limited. They can look convincing, but they don’t always feel natural. Therefore, we use avatars only where they genuinely fit — for instructional segments with medical or trainer personas laying out directives, but not for the human choices, attitudes, and approaches that AI still struggles to capture.
- Translation still needs human review: Local-language video translation continues to throw up pronunciation and contextual errors in Kannada, Telugu, Hindi, and other regional languages. Medical accuracy leaves little room for approximation, so every output is reviewed and refined by our team.
- Voice cloning isn’t plug-and-play: Text-to-speech and speech-to-speech models often require multiple iterations before the emotion, pacing, and intonation feel authentic enough for patient-facing content. The team keeps refining until the output feels right, and falls back on traditional dubbing when it doesn’t.
- Platform volatility is real: AI platforms don’t just evolve quickly, they also change how they charge you. Credits fluctuate, pricing shifts, and features move behind new paywalls, making it difficult to forecast budgets or build stable production workflows. So we keep testing newer platforms against what we already use, and being ready to switch has become part of how we work.
These constraints taught us to budget in ranges instead of fixed figures, keep backup tools for critical workflows, and avoid becoming overly dependent on any single platform.
4. Lead with people, not platforms
For all its strengths, AI can’t replicate the judgment, empathy, and creativity that health communication demands. It can carry the load of asset creation, but it can’t carry the discretion that filmmaking still requires from a human.
That’s the principle I’d offer anyone bringing AI into producing health communication videos: let AI scale what’s repeatable, and protect human judgment where it counts. Medical accuracy, cultural nuance, the storytelling instinct that makes a patient trust us — that still lives with humans, and always will.
Where we go next
Our focus now is on turning experiments into discipline: building an internal AI workflow library, standardizing the processes we’ve developed over two years, expanding our audio-visual library for future usage, and testing emerging formats, including interactive and immersive tools, with the same care we’ve applied from the start.
AI is now a core part of how we work, but it earns that place only as long as it strengthens the human connection at the heart of the care we’re trying to support.
This is an edited version of a post originally published on Ankita’s LinkedIn.
A big thank you to Preetham Tegginamane and Sajeev G Nair for all the creativity and care they brought to the AI work, to Seema Murthy and Parul Singh for backing us, and to Tanaya Jagtiani for her editorial support.
Ankita Das
Film Lead, India
Subscribe to The Companion