Creating a Coordination Layer for AI Agents With Blue Language Labs

[00:00:00] AI agents are only as strong as the data that they're given. When provided with an outdated data set, your agents could end up doing more harm than good. But not with Denodo. With an AI data layer built within your platform, your agents are provided with real-time data changes. So with Denodo, your agents can finally make the right business decisions. Simply visit denodo.com to learn more.

[00:00:27] What happens when AI agents can make payments, negotiate terms and assemble services, but the businesses involved cannot verify what was agreed? Yep, an AI agent might be fast, but speed becomes expensive when nobody knows who approved the action, which conditions apply

[00:00:54] or when money should move. Well, in today's episode, I'm going to be talking with Zor Gorlov of Blue Language Labs. He will talk about the open source Blue Protocol, how that turns agreements, permissions, approvals and settlement conditions into machine executable documents. I will also discuss the difference between authorising an agent to spend and trusting it to form a business relationship.

[00:01:24] And why shared timelines could resolve disputes without one of those heroic searches through emails and spreadsheets. And I suspect we've all been there. So if autonomous agents are going to work across companies, banks, suppliers and customers, intelligence alone is not going to be enough. And today's conversation will be about what infrastructure must come next. But let me introduce you to him

[00:01:50] right now. So thank you for joining me on the podcast today, Zor. Can you tell everyone listening a little about who you are and what you do? Well, first of all, thank you for having me. My name is Zor Gorlov. I am three-time co-founding CEO of enterprise software, cloud and AI companies. My last company, Cosisto, built conversational and generative AI systems for 60 banks around the

[00:02:19] world, including banks like JP Morgan, Standard Chartered, TD Bank, Emirates, Westpac. So I was fortunate enough to have a front row of seats to understand how banks deploy AI to interact with their customers and employees. I started my career in AI back in the Bell Labs days, that was many, many months ago. But today, I'm the CEO of Blue Language Labs. And at Blue Language Labs, we're building the coordination layer that lets AI agents do real business.

[00:02:48] Well, it's a pleasure to have you join me today, especially because you were doing AI before. It was cool, shall we say. And fast forward to present day, every tech conference, every announcement coming out of most businesses now is around AI and more recently, AI agents that can already search, recommend, communicate, and even initiate payments. But I'm curious, from what you're seeing here, and you've seen this evolve from the beginning, where is the dividing line between an agent taking

[00:03:17] action and making a business commitment that other parties can safely rely on? What are you seeing here? Well, I'll respond to the question in a second, but I just want to go back to your observation about AI. Early in my career, I worked at Microsoft. And I remember sitting in 1995 in the audience where Bill Gates declared 1995 the year of speech recognition. So it took a long time for us to

[00:03:43] get to speech systems that actually work. But I'm excited about the moment. I think a lot of things that I did in my career around AI are really beginning to happen now. And AI agents, as he said, are the next evolution of this technology. So to answer your question, in my view, the dividing line is whether an agent can create and execute an enforceable business process. Today, agents can search,

[00:04:08] recommend, communicate, maybe make a payment. They can act. But taking an action is not the same as doing business. Doing business means assembling offers, finding counterparties, agreeing to terms, coordinating fulfillment, managing exceptions, and basically driving the entire process from agreement to settle. And you do that in order to generate revenue. That's what humans do. That's

[00:04:31] what agents will need to do. And that is why we created an open source blue protocol. Each blue contract or blue document includes participants, rules, conditions, state that enable that multi-party coordination that are described. And then ultimately, they drive trusted execution of business agreements between agents and businesses, between agents and other agents. So what blue does, it turns

[00:04:57] an agent intent on agent action to, you know, it's a contract that is human riddable, machine executable, and then let's say an infrastructure that every participant can safely rely on. And for many people listening to our conversation today, they will have human teams that are handling much of the invisible coordination between suppliers, customers, banks, platforms, internal

[00:05:23] teams, and so much more. There's a lot of room or a big opportunity here for improvement. But I think many, even in their personal lives, they might be wary of connecting their inbox and calendar to an AI agent, for example. But in the corporate world, stakes are much, much higher. So what process has become especially fragile when agents attempt to run them at machine speed, where the risks of creating

[00:05:49] mistakes at machine speed might worry some people listening? Yeah, I agree with that. That's a great question. And you know, the order to answer the question, I mean, we need to look at today's business infrastructure, right? Yeah. That infrastructure is scattered across emails, PDFs, spreadsheets, APIs, handshakes, phone calls, institutional knowledge. This infrastructure is designed for human speed, right? So the processes that become most fragile are the ones that cross organizational boundaries,

[00:06:20] like agreeing to terms, confirming authority, coordinating dependencies, approving inspections, delivering, right? And you know, the humans today fill the gap, right? You call your colleague in another business saying, well, I'm missing this, can you send me an email on that? They resolve ambiguity, they chase missing approvals, they reconcile whatever conflicts, and ultimately decide what to do. At machine state, it all breaks,

[00:06:45] right? Those informal, you call your counterparty on behalf of your agent is just not tenable. So in one understanding, especially in the agentic world, it can trigger a really, really bad reaction when those errors that agents or humans can make and become compounded. So human infrastructure is designed for human state. And the answer to agentic is not just better or smarter agents. People in Silicon Valley

[00:07:13] build better harnesses and create better sandboxes, all with the goal of making sure that agents do what they're supposed to do. But ultimately, these agents are probabilistic systems. So what is needed in our world is a shared enforceable business process and contracts in which commitments, authority, evidence, approval, exceptions, everything is explicit and visible to all parties participating in that contract. And that's what blue documents or blue contracts provide.

[00:07:43] And payment protocols can confirm whether an agent is authorized to spend money. But what additional information is needed to establish what was agreed, which conditions apply, and when payment should be captured or released? And one of the reasons I wanted to ask that is, I don't know if you saw this, but Target recently hit the headlines for showcasing their terms that said that AI purchases or AI agent-made

[00:08:09] purchases, they're ultimately considered authorized by the user and there will be no refund. So there is, there's a need to be careful here, isn't there? Yes, absolutely. And I think this whole idea of delegated buyer, right? That's what we're discussing. And delegated buyers focused on payments and maybe checkouts, right? And payment authorization answers

[00:08:35] only one question, right? Is the agent allowed to spend? And if you listen to Stripe and maybe Target, I mean, people are creating guardrails around payments, right? The amounts, you know, or you can only spend and, you know, so much at Target and, et cetera, et cetera. This is essential. This is important. But the payment itself does not answer whether, whether it is appropriate, right? An authorized agent can still pay the wrong amount at the wrong time to the wrong party. So what

[00:09:04] what we at Blue believe that payments are really important, but they're one block in a chain of events that lead to execution of business process. What is important are context and the operating model. The context establishes what was agreed by who under whose authority and operating model defines who must do what which approvals are required, what evidence process was fulfilled and deal with all the

[00:09:30] exceptions, right? What is the return? What the user is not satisfied? I think sort of the Blue connects the payment directly to that business process and makes it more robust. And, you know, the other thought here is the Blue agents are designed for businesses. Businesses can generate new revenue, assemble offers,

[00:09:55] collaborate and cooperate with each other and enable new revenue streams, right? What we think, what we believe in Blue and why it is important, it takes the concept of agents from delegated buyers to delegated business formations. For the agendic economy, I think in one of your podcasts, you said that agentic economy or agentic commerce is projected to grow to 200 billion dollars by 2034. And for this economy

[00:10:23] to reach its full potential, delegated buyer model where I tell agent to go buy something for me is not good enough. I think the real opportunity is delegated business formations where business agents can do more than buy existing products. They must be able to identify, you know, participants, assemble new multi-offerings, generate those offerings, negotiate and coordinate the terms, secure approvals. So I think that

[00:10:51] business-to-business infrastructure is worth the real opportunities in the world. A hundred percent with you. And it can feel right now that so much is changing while so much remains the same. So I'm curious, how does this coordination problem that we're talking about here today, how does that differ from challenges already addressed by workflow software, API, smart contracts and enterprise

[00:11:17] resource planning systems? And also where do these technologies fall short with what we're seeing now? Yeah. So that's a three part answer really. Yeah. One we just discussed, which is like, okay, well, the CRP systems have human-powered infrastructure and humans to operate until we touch on that. But there are two more. One, as I found quite interesting, I mentioned that I listened to your podcast and there's a gentleman from IBM,

[00:11:41] talk about IBM orchestration. And that's a framework for, you know, creating guardrails and controlling agent behavior and all that. Blue is very different. Blue is a protocol that can operate between enterprises than requiring every party to join the same agent control of a workflow plan. Now, by definition that blue enables shared multi-party state rather than again, centrally

[00:12:10] managed workloads. Blue is focused on commercial commitments as a first class protocol objects in blue. Blue is a blue contract that offers and orders and mandates and pay notes. There's an infrastructure that is designed to enable commerce between different businesses. Blue supports independent

[00:12:34] verification of every action that agents takes, but in this multi-party framework and not relying on orchestrator logs within single enterprise. So it's very different from that perspective. And then the third component of it, so we talked about human, we talked about enterprise orchestration frameworks, and third component of a smart contract running on blockchain. You can think of, there are some similarities, right? The similarities are that blue, just like smart contracts or blockchain,

[00:13:02] are deterministic and executable. But blue is different because blue works on the existing payment rates. Blue is not dependent on a blockchain. And we feel that if two businesses decide to work together and deploy their agents, and I don't know, a florist in Manhattan decides to take their, to send their arrangements to 12 restaurants in a three block radius, I don't think global consensus of putting

[00:13:28] events on the blockchain is required. At blue, we have a concept called the timeline and timelines of local processes. Timeline is an append-only hash link record of actions. And those actions can be taken by person, company, agent, bank, bank, system. And the hash chain proves that the record has not been altered. And the timeline provider, which is a key concept in blue protocol, establishes the operational facts,

[00:13:57] who was authenticated, who performed the action, whether authority was valid, when the event occurs, and whether record is complete. This allows independent processors to create composable trust and combine multiple participant timelines and reach the same state deterministically. So that's how blue infrastructure is different from human orchestration, from Asian orchestration, like the ones IBM has, which is amazing product and smart contracts on blockchain.

[00:14:26] And before you join me on the podcast today, I was reading about your travel package example of a boutique hotel, which involved several independent businesses contributing to one customer experience. But what happens when a hotel accepts a restaurant decline or a customer disputes whether part of the package was delivered? Can you expand on that example for anyone that's not heard it?

[00:14:52] Yeah, so that's an interesting question. First of all, blue is a young company that technology has been under development for a couple of years now, but we launched the business at money 2020 in Amsterdam six weeks ago. And this is a travel package is the agentic execution that we prepare for the company launch. You know, there's the, I think the use case was the travel agent

[00:15:18] creating a romantic weekend package by sourcing off season, a room availability at a boutique hotel, coordinating it with dinners at a romantic dinners at a local restaurant as well. And as you said, these businesses did not need to know about each other. We actually took before we went to money 2020, we went, took this process A to Z and we had people who stayed at hotels and ate dinners, which was quite

[00:15:44] exciting. We wanted to make sure that blue worked the way we envisioned it to work before we launched the company. But this whole process, um, that what blue does, it turns this travel package into one shared verifiable process. Each business, you know, travel agency hotel accepts only its own roles and obligations. So the hotel accepts, but the restaurant declines. There are predefined rules within blue

[00:16:09] documents that can trigger replacement repricing customer approval, or even consolation delivery evidence when somebody showed up and checked in a hotel or had dinner at a restaurant is reported against in a timeline and it's supported by each documents. And we manage all the refunds, partial refunds, no show fees, uh, et cetera, et cetera. So disputes in the blue documents are resolved against degree terms approvals and proofs of performance. And you don't have to deal with

[00:16:38] con conflicting emails, you know, reservation systems, booking systems, et cetera, et cetera. And I'm curious, you mentioned the money 2020 there, and I've been incredibly busy with announcements, et cetera. How was that, uh, uh, greeted by the attendees? Uh, what kind of questions were you getting? Was it a mixture of excitement there? What, what were people saying to you afterwards? Yeah, we had a, an amazing show. The, you know, the company came out of

[00:17:04] stealth. We announced the presence of the show. We had strong interest both from marketplaces, but also through financial institutions because blue protocol, as I mentioned, eight is an open source protocol. We have a product called my OS business that is designed for small business and merchants to create their own agentic presence. And I think it's a real opportunity for marketplaces

[00:17:28] to enable agentic revenue generating opportunities between their businesses, but also it allows financial institutions to not only participate in agentic economy and agentic commerce, but actually become a guarantors and drivers, uh, business opportunities. So we have a strong interest and, um, very successful show. One thing that is worth noting and, you know, sort of vision of, uh, blue and why we exist. We believe

[00:17:54] that the reason real opportunity for small businesses and small merchants to deploy agents that will help them be a very, very, very, very, very, very, very, very, very, very, very, very, very, very, very, very, very, very important. And again, for people listening inside, uh, their, their, their companies and organizations, and they're hearing about this here, what competitors are doing, what, how should their company decide which actions an agent

[00:18:19] should take independently? What would require human approval and, and which should remain outside an agent's authority altogether? I suspect this is a question you've been asked a lot this year, but which way would you direct them? That's, that's in everybody's mind. I mean, our key blue principle is that humans remain in control, not in the loop. There's different. If you put, uh, humans in the loop of every agentic commerce

[00:18:44] interaction that, that you were back to human speed, but humans need to remain in control. We do that by introducing a special type of blue construct called the mandate. Mandate is an explicit revocable grant of authority defining what an agent may do on whose behalf, within what time limits for how long, and when human approval is required. So human approval itself and the rules are baked in a

[00:19:12] document. You know, routine low risk actions can proceed automatically. Fire value unusual or irreversible actions may require, uh, human approvals. And then the, as the agents get better and, you know, these, the AI learns, you know, the mandates can be updated, um, accordingly. So the guardrails will evolve with evolution of agentic commerce, but mandate is a key concept that controls

[00:19:38] when and how humans get involved and what conditions need to be met. For example, if I'm a restaurant operator and I deploy an agent that says, well, go to, uh, buildings near me, find startups and finance companies in New York and offer catered lunch to them. But then you can say, well, I want to know that I want to approve every catered lunch, uh, that, that is then, uh, 20 people, because if it's more

[00:20:07] than 20 people, I need to deal with this kitchen capacity. So that's what mandates do for you. And when several agents and indeed businesses participate all inside one process, another question I'm sure you get a lot is around accountability. So who is accountable when something goes wrong, what evidence audit records and dispute mechanisms will needed to establish who is responsible before we end up with one of those Spider-Man memes where everyone's just pointing at

[00:20:35] each other? Yeah. Look, I think accountability should follow obligation and authority behind it. Yeah. You know, not simply looking at how the last agent fail in blue, blue document, each participant accepts the defined roles and every agent operates under a, this construct that I call blue mandate. It identifies who authorized it, what it could do within what limits, et cetera.

[00:21:01] Blue timelines come into play then because blue timelines create a sign, attend on the record of actions, approval, state changes, communication, everything is in a timeline. So if something goes wrong, the blue document shows what was agreed, who was responsible for each step, whether agent acted within the story and what evidence was required. Now, the dispute, remediations,

[00:21:25] refunds, escalations, PAC can all be triggered within blue documents, but the evidence itself is stored in the timeline. So in other words, blue does not eliminate disputes, but it replaces these fragmented records that exist today and finger pointing with verifiable responsibility and shared basis for conflict resolution. And for any company that's interested in experimenting with agent driven businesses today, I would

[00:21:52] imagine there is a lot of people are going to be hanging on your words here, not knowing which way to go. What, what foundations should they be putting in place now around identity permissions, process rules, human oversight, exception handling and payment controls before even thinking about granting agents greater autonomy. Any foundations or advice you'd offer there? Well, you know, first of all, the starting point is that not to whether to rest an agent.

[00:22:19] Yeah. Right. Uh, I think that, and also it is not going to work to try to program every agents to always do the right thing. You just simply were not work because this agents are probabilistic systems. Business conditions around them change and unexpected situations are almost inevitable. They're, they're going to happen. So I think it starts with every agent having a verifiable identity,

[00:22:43] like a clearly identified person organization whose, whose behalf this agent acts. Right. And then it also needs to have rebuttable permissions defining what it may do within what limits and for how long. And then you layer on top a business process that humans are comfortable with that should establish approval thresholds, prohibited actions, escalation path when humans need, needs to get involved and then high

[00:23:10] risk or irreversible action must remain subject to human review. You know, uh, payment controls we talked about, it should include spending limits, uh, limits, approved counterparties, conditional authorization, capture all of that. So you need, we like to say at a blue language labs is you don't trust your agent. You trust the system. So you need to build a system that will ensure that agents

[00:23:36] behave within the rules and conditions defined by humans. So, uh, blue language labs, you're building this backbone of the AI economy with a secure trust layer that lets people enterprises and autonomous agents transact, negotiate and collaborate at machine speed. But ultimately are you just, you're turning agent actions into verifiable business processes. And most importantly, with shared rules, auditability, settlement on

[00:24:05] existing rails, et cetera. And I think this will tick the box for so many organizations listening and anyone that just wants to start a conversation or find out more information about what you're doing, how you're evolving, how you might be able to help. Where would you like me to point everyone listening? Well, the best place is their website, uh, blue language labs.com. My email is door and blue language labs.com. You know, I'm a LinkedIn

[00:24:29] on the ax, uh, blue language labs.com. You can read about coordination layer and blue protocol, uh, that we open source as well as my OS, which as I mentioned is an operating surface, uh, for blue, uh, documents for, for businesses and look and your podcast. Of course, I'm sure a lot of people will listen to it and reach out to us. Well, thank you so much. We covered a lot in a short amount of time today. And I think

[00:24:55] we've set off a few light bulb moments around the world as people want to dig a little bit deeper on this stuff. So I urge anyone listening, if this has struck a chord with you, please go check out the website. There's a lot more to come on that as well. So there'll be a lot of big announcements I would imagine in the future, but more than anything, thank you all for sitting down with me today, bringing this to life in a language that everyone can understand. Really appreciate you, Tom. Thank you for having me. I think Zor's travel package example brought this coordination problem we discussed today to life.

[00:25:24] A hotel, restaurant, local guide, travel agency, and customer can all participate in one offer while keeping control over their inventory, pricing, approvals, and obligations. And the practical lesson here is to define authority before giving an agent the freedom to act. Decide what it might approve,

[00:25:49] which decisions require a person, what evidence confirms completion, and what happens when one participant cannot deliver in that process. But I think the opportunity extends far beyond just making current processes faster. Agents could help smaller businesses combine their capabilities and create offers that none could provide in their own right. And the challenge here is going to be persuading

[00:26:13] banks, platforms, suppliers, and customers to all trust one common coordination method without creating another closed system. So a big thank you to Zor for bringing this to life today. And over to you, what would your AI agents need before another business could safely rely on their commitments? techtalksnetwork.com. That's the place where you'll find me. 4,000 interviews, a lot to keep you

[00:26:41] quiet over there. And if you don't want to stay quiet, send me a message, send me a recorded message, and I can get back to you. But that's it for today. Thanks for listening as always. Bye for now.