Earley AI Podcast - Episode 92: Supply Chain Intelligence, Knowledge Graphs, and the Limits of the Easy Button

Why Supply Chain Visibility Is One of the Most Consequential and Underestimated Applications of AI in the Enterprise

Guest: Ilya Levtov, CEO|Founder at Craft

Host: Seth Earley, CEO at Earley Information Science

Published on: June 1, 2026

Podcast Transcript: Supply Chain Intelligence, Knowledge Graphs, and the Limits of the Easy Button

Transcript introduction

This transcript captures a conversation between Seth Earley and Ilya Levtov about why supply chain visibility remains one of the most consequential unsolved problems in enterprise operations - and why AI is finally making it solvable. They cover the gap between relationship-based supplier knowledge and real intelligence, how knowledge graphs ground AI in proprietary data that LLMs do not have, why a single risk score always misleads, how cross-correlating data streams surfaces hidden systemic risks, and how the give-to-get model turns supplier intelligence into a shared competitive asset.

Transcript

Seth Earley: Welcome to today's Earley AI Podcast. I am your host, Seth Earley, and in each episode, we explore how artificial intelligence and data are reshaping business strategy and operations.

Today, we are talking about one of the most consequential blind spots in enterprise operations, and that is supply chain visibility. Most organizations know their top suppliers well. Below that, the picture gets murky fast, and as recent years have shown, the risks hiding in the middle and long tail of a supply network can be catastrophic.

Joining me today is Ilya Levtov, Founder and CEO at Craft.co, a supplier intelligence platform that uses AI and knowledge graphs to give enterprises and government agencies visibility into their full supply networks. Ilya founded Craft after careers spanning Goldman Sachs, venture capital at Venrock, and enterprise partnerships at Deutsche Telekom, and has spent nearly a decade building what he describes as the source of truth on companies - which has evolved into a platform for mission-critical supplier intelligence. Ilya, welcome to the show.

Ilya Levtov: Thank you, Seth. It is great to be here, and I am looking forward to our conversation.

Seth Earley: When you talk to procurement leaders and executives about supply chain visibility, what do they most consistently get wrong?

Ilya Levtov: The biggest misconception is that having good knowledge of the top 100 or 200 suppliers - through a strong relationship between a category manager, a buyer, and that supplier - is good enough. That really is kind of where most organizations are. And as you alluded to, there is that whole middle and long tail. It can be up to 20,000 or 30,000 suppliers easily for a Fortune 500-style enterprise, and many more when you go deeper into the multi-tier, the suppliers of those suppliers. So one of the biggest misconceptions is that we are fine because we have visibility into our top few hundred.

What compounds that misconception is the feeling that getting visibility into the rest is so hard it is insurmountable - that we may as well not even bother trying because it is too much data. That is really the big one. And it is exciting to be able to overcome it.

Seth Earley: It is a constantly evolving and dynamic environment. People think it is impossible to really understand it, too many data points, too many moving parts, too many interrelationships. And that is where Craft has made such significant strides. So where exactly does that visibility start breaking down, and what are organizations actually missing?

Ilya Levtov: The visibility that exists on those top critical suppliers is really done mainly by relationship - between the buyer and a relationship person at the supplier. And right there you can see the blind spot emerging. Because what is the information that relationship gives you access to? It is the information the supplier is providing. And of course the supplier is saying everything is great, we have capacity, everything is fine.

So you are immediately seeing the end-tier problem, because that supplier is also representing that they have got great visibility of their own supply chain - which you know they do not, because you do not. So the gaps come in immediately.

One of the first ways to mitigate this is to establish a 360-degree view of the supplier using data aggregated from very many different sources. For example, crawling the web for every public signal - a news article, or the number and type of job openings a supplier is publishing today and how that has changed over the last few weeks. That is giving you clues about what is going on inside that company. You probably are not visiting every supplier's website every single day. Then connecting with specialty data providers that evaluate a company's cyber footprint - monitoring whether that company has had any data leaks, any appearance of confidential data on the dark web that would indicate a vulnerability. These are the kinds of data sources that are actually available to bring together and integrate to enrich this picture.

Seth Earley: And the challenge is separating signal from noise, because the data is enormous and the permutations are endless. One of the things you have built is a progression from structured supplier data through knowledge graphs. Can you walk us through that evolution and why each layer matters?

Ilya Levtov: It starts with collecting structured data, semi-structured data, and unstructured data. That in itself is an enormous amount of work when you consider all the different lenses: financial, operational, geographic, cyber, ESG, reputational, regulatory, and compliance. We call it the data fabric. It is the foundation of the entire solution, and it is hard in itself. It is necessary, but not sufficient to get to value. Anyone who has just collected a lot of data finds themselves stuck exactly where you said - tons of data by itself just looks like noise.

The next layer is converting that data into intelligence. And that is where the concept of the knowledge graph became clear to us. At the beginning, structuring all of that data is typically done in a relational database - columns and rows, static attributes and data points. But the reality of the world around suppliers is that it is genuinely more a network than a chain. The whole supply chain concept is a misnomer. It is a supply network.

Company A is a supplier to Company B. Quite often, Company B is also a supplier back to Company A, or to a different division of it. And then there are people - a really important type of node. The graph construct is very helpful because it gives you nodes and lines. Nodes are corporate entities - but also the parent company and its subsidiaries, manufacturing locations, and the people involved. And then all of the connections among those entities. In totality, that gives you an accurate and dynamic picture of what is going on.

Seth Earley: And the knowledge graph is what allows you to ground your LLMs in something that is actually real for your organization. Because only about one percent of enterprise knowledge lives inside LLMs today.

Ilya Levtov: Exactly right. Only approximately one percent of enterprise knowledge exists inside the LLMs today - and of course that makes sense. Companies do not want to give all of their data to the LLMs. Data providers who have built that cyber data or that ESG data do not want to give it away for free either because they would lose their business. And so the LLMs actually do not have this data fabric that we describe. That is why you need a specialized approach - leverage the power of the models on your own data set and on your knowledge graph.

Seth Earley: And organizations compete on their knowledge. Knowledge of customers, knowledge of solutions, knowledge of supply chains, knowledge of routes to market - those are competitive advantages. You do not want those inside an LLM, because you do not want to lose them. So tell me - when you start treating supply chain as a complex network rather than a linear chain, what kinds of insights become possible that were not there before?

Ilya Levtov: The first example is how insight comes from cross-correlating different data streams. We found that by looking at the digital footprint of companies across social media platforms, we were able to build a picture of the culture inside the company - including the quality and level of employee engagement. You get almost a window into the morale of a company whose products you depend on. What surprised us was a very high correlation between that level of morale and the ultimate cybersecurity posture of that company. Companies with lower employee engagement, lower morale, a slightly disaffected workforce - were much more prone to spear phishing and leaks. One example of taking multiple data streams, integrating them with strong entity resolution, and finding a very significant risk insight.

Here is another example. Everyone thinks about financial risk in a supplier - will they be financially solvent six months from now? What people were not really thinking about is that when the answer to that becomes shaky, a financially vulnerable supplier actually becomes vulnerable to adversarial capital. The influx of investment from entities that might have a hidden agenda. In the world we operate in - mission-critical supplier intelligence for national infrastructure, aerospace and defense - these are real concerns. But they are concerns for any enterprise, because there is always a concern about IP theft and exfiltration. A financially vulnerable supplier becomes a target for outside capital because they are looking to survive. And if one of their options is taking money from an entity coming in from an unfriendly nation, this is happening across industry in very concerning ways. You are connecting two different data sets, connecting entities, and getting to a very significant risk insight that you need to act on before it becomes a problem for your enterprise.

Seth Earley: And who would have thought that a hidden morale issue could put an organization at systemic risk. That is what makes this so significant - finding these hidden signals, these latent attributes that are immensely impactful and that you just do not know exist. Talk about the limits of the simple risk score, because everyone wants an easy button.

Ilya Levtov: Everyone wants a single score. We want that for them as well and we are always pushing towards distillation. But in practical reality, early efforts to distill supplier risk to a simple score or a high, medium, low - it was convenient but it was not accurate and not actionable. It might lead you down a course of action that could be very wrong.

My top example is the deep interconnectedness of the global supply network with upstream materials from China. It is very easy for an aerospace defense company to say, I do not want any China in my supply chain. It is just not possible. Decades of evolution of this supplier network means some material is coming from China, possibly going through other countries, and the manufacturing capacity is simply not available elsewhere. Rare earth minerals are the obvious example.

A risk analyst looking at this might want to say, that is high risk because it is coming from China, therefore I want to take it out. But that is not possible - the whole product line goes down. What the analyst really has to be doing is saying, is this level of risk acceptable? Can I continue to produce with it, or do I really have to stop? Because if stopping costs $20 million to go get supply from a more expensive source, what I need to know is whether that is greater than or less than the value of the risk. That nuance cannot be expressed by a single score that says supplier risk equals high because domicile equals China.

Seth Earley: Right - and it is the components that matter. Some areas are very mature, some are not, and the divergence of those scores is what drives the real insight. In our AI maturity diagnostic, you can get an aggregate score, but the lowest area is actually the bottleneck, and that is what you have to act on first. The same principle applies here. So what is your approach to building confidence in data when you can never get to 100 percent accuracy?

Ilya Levtov: Multi-sourcing. We get different sources all to opine on a single question - like what is that supplier's cybersecurity posture? We have partnerships with not one, not two, but three independent providers of cyber intelligence, so we can see each of their opinions. When all three align, I still would not say that is 100 percent, but you are at high confidence. Three independent sources all came to the same conclusion - that is one you do not need to spend more time on.

But when they diverge, that is an extremely useful flag. It signals that this is the one to go and do extra work on. The misconception was that accuracy has to be 100 percent. It does not. If you can get into the 90s on accuracy and absence of false positives and false negatives, you are doing well - and it is vastly better than doing nothing, which is often the alternative.

Seth Earley: And we are finding material risk findings in a significant percentage of supplier portfolios. What does AI actually change for the analyst doing this work day to day?

Ilya Levtov: The acceleration is enormous. An analysis that used to take eight hours per supplier - building a deep and comprehensive view to determine whether this supplier is acceptable in your supply chain - we can now get a first draft ready in about a minute. With another 15 to 30 minutes of working with that draft, you end up with an outstanding and comprehensive supplier report. That is an 80 to 90 percent acceleration, and we are still only just getting started.

But the decision of what to do with those findings still very much lives with the human, and will for the foreseeable future. In our domain, the cost of error is very high - do I buy more from that supplier or less, do I cut them out entirely, how do I renegotiate my renewal? All of those decisions we see still living with the human. Who, however, has just been given superpowers from the data fabric, the knowledge graph, and the LLM synthesis of that.

Seth Earley: Let us talk about the strategic opportunity side of this - not just risk. What can organizations see about their competitive position with this level of fidelity that they could not see before?

Ilya Levtov: Many of our procurement teams use the intelligence they get on their supplier network to also understand how they are positioned against competitors for particular elements of supply. When you illuminate the end tier and go down to Tier 2, Tier 3, Tier 4, you can look at large data sets that give visibility into who is supplying what to whom across global trade routes. You can hone in on choke points in the supply network - rare earth minerals are a great example - and see the volume of shipments coming from certain upstream manufacturers to downstream processors including yourself and your competitive set. From that you can get visibility into how well you are positioned competitively. That is tremendously important in areas where supply is constrained.

Seth Earley: And one of the things we talked about in our prep call is the network effect at work here. The more suppliers you have information about, the richer the picture becomes. Tell me about the give-to-get model and how you balance that with protecting each customer's competitive position.

Ilya Levtov: The most fundamental principle is the sanctity of confidentiality and privacy of data from each enterprise customer - including who their suppliers are. As a multi-tenant architecture, data from customers is extremely tightly walled off, with absolutely zero leakage between tenancies. That is the zero-term foundation of every engagement.

With that said, there is real business value that customers want to create for the benefit of all - by observing certain relationships across the knowledge graph. The way we have squared that circle is with an opt-in, give-to-get model. An enterprise can, with full agency and control, say that under specific circumstances I am willing for some of the insight that really belongs to me to be shared - explicitly, with my permission - because I am generating value from someone else doing the same.

In the aerospace and defense industry, there is approximately 70 percent overlap of suppliers, Tier 1 through Tier N. When one large enterprise starts working with us and we start analyzing one of their Tier 1 suppliers, that enterprise is effectively getting visibility into their Tier 2. In many cases, because two enterprises are very interdependent, one will say - I have developed and invested in a key insight about a particular supplier that both of us depend on, and I am happy for you to have access to that. Because it makes me more secure as well. We call that supplier passporting - taking a report, once it is derived, and sharing it with others such that the entire network becomes more secure.

Seth Earley: Ilya, thank you so much for joining me and helping our audience really understand why supply chain visibility at this level of fidelity is one of the most important and underestimated applications of AI in the enterprise. It is going to make a material difference to risk, to profitability, to reliability, and to the ability to deliver. Really appreciate your time today.

Ilya Levtov: Seth, I really enjoyed the conversation. Thank you very much for having me.

Seth Earley: And to our listeners, thank you for tuning in to the Earley AI Podcast. We were recently named one of the top 80 AI podcasts globally, so please share with your colleagues, subscribe, and we will see you next time.