NDTV’s AI strategy is less about models and more about what publishers build around them
Most AI products get dumped after the novelty wears off. The challenge for publishers is building products that continue to create value, according to Rohan Tyagi, Chief Product Officer at NDTV.
Building AI products is becoming easier but developing products that continue to create value is getting harder.
“This is because the entry barrier to building product is getting lower. So there is not much challenge in building new product,” Tyagi said during our Digital Media India conference in New Delhi.
The AI models may be the same, but publishers can still build different products. Rather than treating AI models as the product, publishers should focus on the layers they can own – software, workflows, data and content that are difficult to replicate, he said.
To explain how NDTV approaches these decisions, Tyagi shared a framework the company uses internally to evaluate AI products and features.
The framework divides products into two categories:
- Thin AI layers: These are built largely on existing AI models, often as “wrappers” around chatbots or existing LLMs. Their value lies in helping teams move quickly, test ideas, improve workflows and manage organisational access and credits. However, these products can contribute to what Tyagi described as “vibe-coded slop”, making it easy to build AI products that do not solve a real problem.
- Thick AI layers: These require more time, investment and judgement. They combine AI with something proprietary, whether software, data, workflows, processes or content that cannot be replicated easily. While they take longer to build and may not always make sense from a business or cost perspective, they have greater potential to create long-term value.
Building AI for newsroom workflows
NDTV’s internal tools show how the publisher is using AI to solve workflow problems while keeping editorial teams involved.
One example is Echion, an infographic tool that converts articles into visual graphics and is optimised for news and multiple languages.
“It’s mostly off-the-shelf image model underneath, with a UI on top,” Tyagi said.
The company invested in templates, default prompts and branding guidelines so newsroom teams do not have to begin every prompt from scratch. Users can also reuse designs created by other teams and build templates for different brands or sponsored campaigns.
For instance, AI models can repeatedly make the same mistakes, such as generating an incorrect India map. Templates can help solve these recurring issues by embedding instructions that teams do not have to repeat every time.
“You have to repeatedly prompt it every time. You can set these things into your template, and then they can be implemented,” he said.
Another tool is Liza, which acts as an analytics agent for editorial, product, social and revenue teams.
Rather than creating another analytics dashboard, NDTV gave the AI agent access to Search Console, Google Analytics, Chartbeat, YouTube and other data sources. Teams can ask questions, generate reports and schedule recurring updates through Discord.
“This is actually one of the only real value agentic AI products I have seen internally,” Tyagi said.
Building new consumer experiences
NDTV is also using AI to rethink how news is presented and discovered in its consumer products.
Inside the NDTV app, the company is testing multiple versions of AI-generated feeds along with more editorially curated experiences.
One feature creates AI cards from articles by combining summaries, short videos and GIFs. Editorial teams can review, correct and grade AI-generated summaries inside the CMS after early versions produced inconsistent results.
“We created sort of a playground for editorial teams where they can rate the summaries, because we were going wrong with the summaries a lot,” Tyagi said.
The company is also testing recommendation systems based on LLM embeddings instead of traditional machine learning models. These recommendations better understand the context of articles, while another experiment automatically clusters stories around entities to create more fluid ways of navigating news, Tyagi said.
“We don’t have considerable result on the A-B test yet, whether this works better than editorially aggressive feeds or not,” he added.
However, the approach has shown positive results for acquiring new users, he said.
Another product is askNDTV, an answer engine built on the publisher’s archive.
Rather than returning articles that match keywords, the system is designed to answer questions using NDTV’s journalism and structured datasets.
“The value is that it is grounded in NDTV data and trusted sources,” Tyagi said.
The company is also experimenting with AI-generated audio experiences. During development, the newsroom found that audio scripts required different editorial standards from written summaries.
“What works in text in terms of summaries and text doesn’t work in audio,” he said.
Editorial teams graded scripts to improve quality before the company built an audio feed using cloned anchor voices and AI recommendations.
Investing beyond the AI interface
Building consumer AI products also exposed the limits of relying entirely on existing AI services, Tyagi said.
Running vector search and LLM embeddings across decades of published articles created significant cost and performance challenges.
To address this, NDTV invested in research to improve retrieval across millions of articles.
The company developed a new way of structuring its database for LLM retrieval and submitted the work as a research paper, Learned Rotation-Aware Binary Projections for Efficient News Retrieval, which was accepted at an information retrieval conference.
Tyagi described this as an example of a “thick AI layer” – an investment that takes longer to build but becomes more valuable over time.
Lessons from building AI products
One lesson for publishers is that not every AI workflow needs the same level of editorial review.
Analytics products working with structured data have a lower risk of hallucination because they answer logical questions using structured inputs. AI-generated summaries, however, require closer editorial involvement, according to Tyagi.
Rather than asking editors to approve every AI output, NDTV is building grading interfaces where journalists can review, rate and provide feedback on summaries and infographics. The feedback is then used to improve future outputs rather than simply approving or rejecting individual pieces of content.
“I think it’s not about approval and rejection. You’ll need grading models,” Tyagi said.
While publishers may have access to the same AI models, the opportunity to differentiate lies in what they build around them, he said.
