Could employee digital twins streamline collaboration? Viven co-founder Ashutosh Garg thinks so

Emerging from stealth mode this week, Viven has pioneered a new approach to employee digital twins that enables colleagues to find answers to questions based on a dynamically generated knowledge graph. Rather than interrupting those colleagues with simple questions, it pulls together appropriate tidbits from their interactions with various enterprise apps and SaaS applications. This thoughtful approach also limits the propensity of Large Language Models (LLMs) to hallucinate.

Viven CEO and Cofounder Ashutosh Garg gives me an example of asking his CFO for a summary of a certain part of the business. If he had asked her directly, she might have spent an afternoon making a comprehensive report, since he is the CEO, even when he was only looking for an approximation to guide a new business analysis. Garg explains:

For the same question, she will end up spending half a day or four hours answering that for me. She's like, ‘Okay, my CEO is asking for the answers to be very well thought out,’ but what I'm really looking for is a simple checkpointing system. So, it's a very interesting psychological barrier. She's like, ‘Okay, fine. If my digital twin is answering that lightweight answer, I'm totally fine with that.’

His digital twin knows what he is working on and can adjust the context for his analysis. She gets a report of the chat and can correct any details that it may have gotten wrong. In the meantime, her afternoon is freed up for other work. Garg can also ask other digital twins questions about the status of various client engagements during a meeting with an exec, rather than needing to follow up later.

At first blush, there is something slightly creepy about the possibility of people getting answers to questions they would not ask in person. But Garg assures me there are privacy and security safeguards, and you get to see everything people ask your twin. You can’t ask about someone’s spouse, medical condition, or income, even though they might have some emails about those things. Also, it won’t share details like credit card numbers, personally identifiable information or other confidential data that someone has not been authorized to see.

Digital twin of people

The idea of a digital twin of people deserves some contextualization. There are numerous efforts to “upload” people into digital or virtual twins in various shapes and sizes. Venture Capitalist Tim Draper has been popularizing the idea of uploading old talks and videos to generate AI content in your likeness. Draper now has four fine-tuned for PR messages, content creation, advising entrepreneurs, and creating a hologram of himself. Other researchers are trying to create digital twins of lots of people to improve market research. This extra complexity and compute did enhance some kinds of analysis, but more work is required.

There are also efforts to build digital twins of a wide variety of other things, like products, factories, physical infrastructure and supply chains. Dr. Michael Grieves, widely credited with coining the digital twin concept, once told me he likes to start here because it's easy to understand the value. It's much cheaper to prototype virtually before making physical changes. But increasingly, digital twins are being built for supply chains and business processes as well.

These various kinds of digital twins are good at surfacing different views of things across multiple contexts. But the most difficult concept to understand is the essential role of the knowledge graph for linking this context together. This is precisely what’s special about Viven’s approach to digital twins of employees.

Building a better knowledge graph

Garg has spent the last several decades innovating better knowledge graphs to support various use cases. In the early 2000s, he pioneered techniques to disambiguate searches for people with the same name before joining Google. Later, he applied the same approach to make it easier to find details about Apple Computer rather than the fruit.

After he left Google, he co-founded Bloomreach to apply knowledge graphs to improve e-commerce personalization. Then he cofounded Eightfold to develop a technique to dynamically determine a person’s talents even if they were not directly listed on a resume. For example, someone working at IBM would probably have some familiarity with the Watson tech stack, even if it were not directly mentioned on their resume.

HR Guru Josh Bersin wrote about it in 2020, noting that he initially didn't quite understand what Eightfold was doing when they first explained it to him, as many other companies with bigger pockets were doing something similar in the talent space. It took Bersin some time to recognize the real innovation, which he called “skills inference,” as a weak signal about what someone really knows. This lead helped Eightfold to expand across the saturated talent market, achieving a $2 billion valuation, all before LLMs and AI became what they are today.

Flash forward to today, and now Garg and his team are applying similar principles to distill weak signals about an employee’s knowledge, experience, and expertise from their interactions across various enterprise and SaaS apps into this employee's digital twin. Knowledge is the information the employee has collected over the years, and experience is how they have interacted on that information. Expertise represents the deep knowledge someone might have gleaned from working in a particular domain over the years.

It also works for employees who have left. For example, Garg walks me through a recent chat he had with the twin of a recently departed executive involved in engaging with the public sector. The twin helped Garg improve his understanding of the different folks involved in a major certification for Viven and explained the role of various partners in supporting the process. This helped him craft a more engaging thank-you letter for their support.

There is an LLM on top that simplifies the process of asking questions and summarizing data. But the context fed into the LLM is limited to what can be pulled out of the knowledge graph. Garg explains:

As of right now, our bigger focus is on knowledge, experiences, expertise, less on tone tonality, for example. For each person, we try to model their context and memory, and we are conducting various experiments right now with reinforcement learning and fine-tuning as well. The second thing is that when I ask someone's twin a question, the first thing it does is try to figure out why I'm asking and what other context I might have received. What was my previous conversation with this person in whichever channel prompted me to ask this question in the first place? Capturing that makes everything very, very different and very powerful. Then it looks across those conversations I have had with this person in the past, and in what level of depth I've shared the information.

With all the talk about LLMs these days, knowledge graphs are emerging as a powerful complementary technology for grounding their work in a more focused context. Knowledge graphs have been around for a while, but they seem to be only now coming into prominence. Garg weighs in on seminal improvements over the last couple of decades he has been working in the field:

One is the quality of the knowledge graph that you can generate in a scalable fashion, which is much, much better today than back in the days when we would have a list of attributes for each person. Now we don't have to worry about that. We can very easily scale. So, for example, this product scales across every role in the company. As a CEO, I can use it as a support person and as a marketer.

Second, because of that, we can also capture things across time, which is a very important concept we are trying to bring over here: how your knowledge has evolved over the years. And the insight for that came from Eightfold, where we said it's not important what skills you have. What is more important is, what is your learnability of those skills? How quickly can you learn those things? So the way you might have responded to let's say, someone like me five years back is very different from how you might respond to me today.

My take

LLMs and agentic AI seem to be attracting all the attention these days. But LLMs on their own are prone to hallucinating and provide ungrounded answers. Getting these to work well at scale will require equal effort in building better knowledge graphs. The pioneers behind Viven have a strong record in applying knowledge graphs to solve a variety of enterprise problems. It will be interesting to see how adding LLMs into the mix for employee digital twins re-shapes the collaboration experience.