Treat the Agent as an Embedded Worker in the Ecosystem

AI AgentsAI Data
Written By Christina Ellwood

Episode Summary

A contract describes how two companies agreed to work, and Randy Friedman, chief commercial officer of Cognizer, thinks it is time to stop checking on that with a monthly report. His company turns contract documents into structured data, and agentic flows now build the extraction models rather than engineers training them by hand. The interesting part starts where his own product stops. Enforcing an agreement continuously means seeing inside a supplier’s systems as they run, which is where his embedded worker comes in: an agent with access to the data it needs and no more. What stands in the way is the data rather than the models. Today’s infrastructure assumes one organization and one administrator, and this needs decentralized control, work at the edge, and governance from the bottom up. The fabric for it does not exist yet, and his argument for why it will is that supply chain people have had it on their minds for a while.

Key takeaways

The history he runs through is the useful part of the answer. The conventional approach was training models with training sets and aiming them at the problem, which evolved into using large language models and generating prompts for them

  • His definition of contract intelligence is concrete enough to steal. Take contract documents, extract intelligence from them, structure it, graph it and generate insights and analytics, so an organization understands its commercial relationships, obligations and revenue opportunities

  • The reason one model cannot serve every customer is the phrase he keeps returning to. A contract is a data type, the aggregation of a company’s millions of them is a data space, and no two companies have the same one, so each needs its own intelligence to resolve it

  • He is careful about what large language models contribute, which is unusual for a vendor. Large language models provide some of the modeling a contract data space needs, he says, but not all of it

  • What changes with agentic flows is who does the assembling. Instead of prompting the model, you create a team of agents, give them assignments and something close to an anthropomorphic personality, and they do the work of building the AI

  • The line that gives the episode its shape is one sentence. They become the AI that builds the AI, and what a customer supplies is not the words to look for but the legal concepts and the meaning, after which the models get built automatically

  • The constraint he names is the hinge of the whole conversation. You can only really train models, even with agentic flows, on data that you control, which is fine until the work needs data the organization does not possess

  • He sizes the ecosystem by the supply chain, having just said this part is beyond what Cognizer works with today. The companies a business contracts with are in its supply chain and value chain, and he expects the business to end up deeply embedded inside those organizations from a data and AI point of view

  • His worked example turns a contract from a document into a control loop. A supply agreement sets quality, cadence, availability and inspection, and analyzing that contract is one thing while enforcing it is another

  • The status quo he is arguing against is a monthly report, and he is specific about why it fails. A spreadsheet or a periodic report is too slow to optimize against, so the buyer needs a dynamic flow of the data that report was made from

  • He hedges his own claim rather than overselling it. AI makes this possible to do dynamically, and he would not say in real time, but continuously, with the cadence set by how fast the opportunity is moving

  • His one-line history of information exchange lands the point. He calls permeable perimeters the natural continuation of a sequence he has just run through: fax machines, an emailed spreadsheet, a shared table both parties work in, then sensing built into an ERP

  • Asked whether this is an agentic workflow API, he reaches for people instead. At a high level it is an embedded worker, or a team of workers, allowed inside your organization but not free to run amok

  • His access rule is a tight definition of least privilege. The data has to be available to the agent, but not necessarily visible or useful in any way other than its intended purpose

  • He puts the same shape on two industries that have nothing to do with supply chains. One patient seen by multiple doctors in different facilities with different systems needs a continuum of care, and drug discovery runs the same patients through different clinics inside different trials

  • His summary of the arc is the compact version of the whole problem. It goes from producing training data to train a model, to prompting a large language model and doing retrieval on your own content, to agentic analysis across multiple datasets, not all of which he has permission to use or to create

  • The diagnosis is architectural rather than technical. Today’s data infrastructure is organization-centric with an administrator who delegates roles and access, built up from the conventional database, and it was never designed for cross-organizational engagement

  • The first requirement he names comes from blockchain, with its performance left behind. He calls decentralized control through smart contracts a totally different kind of structure, and says it does not scale very well, so what is needed is decentralized entitlement without blockchain’s performance cost

  • The second requirement is physical. The more computationally complex the work, the closer to the data it has to run, so gathering everything in one place to analyze stops working once the work gets complex

  • The third requirement inverts how governance is usually built. Fiduciary, legal and regulatory obligations mean you cannot have a centralized governance structure, and it has to work from the bottom up at the data level rather than top down at the organizational level

  • He is describing something that does not exist and says so plainly. Those kinds of things do not exist right now, and what is needed is a new fabric connecting organizations in a secure way, federated and decentralized in its entitlements, so each organization controls what data is used, how, by whom, and by which agent

  • His own company already runs a version of the trust problem. Cognizer is asked to build models and AI capabilities that understand a customer’s contract data space without ever being allowed to see it, which he calls an interesting trick

  • He names the tension without pretending it is solved. People in his and the host’s shared network have put their models behind the firewall to protect their customers’ data, which he says sets an upper limit on what they can do

  • On whether change comes from the business or the technology, he says both, and the cloud is his evidence. The cloud solved a technical problem of scalability, redundancy and economy, virtualization is what made it possible, and only once people had it did they push it up to the application level and get the collaboration they had always wanted

  • His analogy for repurposed technology is a drug developed for one thing that turns out to do another. He expects reuse of technology people did not think was good for a particular purpose, the way a blood pressure drug turned out to grow hair

  • He accepts the host’s framing that new kinds of companies force the issue, and extends it. The evolution will drive new enterprises and new providers, and he points at large language models being productized as virtual workers you hire and pay by the month

  • He draws a distinction between two kinds of agent negotiation, and notes in the same breath that Cognizer is not doing the negotiation case. Two agents representing two parties sit outside each other’s security perimeters with no overlap; doing the business activity the contract describes is what forces the perimeters to overlap

  • His answer on what executives should do starts with the demand already inside their own teams. If you run procurement or manage a supply chain, he says, making it more performant, more observable and more controllable is already on your mind, and the data your work depends on already belongs to somebody else

  • His closing idea is that most of what is coming is old ideas coming back, and he has a name for it. Bell bottoms: concepts from earlier technology implementations reimagined and repurposed with more power, and he tells enterprise leaders to ask the big questions that follow from it

  • The example he ends on is a company of people, not software, which is the point. Deloitte’s few hundred thousand people work inside their customers’ perimeters and effectively become embedded workers, and his question is how you make that happen with AI

About Randy Friedman

Randy Friedman is chief commercial officer of Cognizer, which applies AI to legal business processes and specializes in contract intelligence: taking contract documents, extracting the intelligence in them, then structuring and graphing it into insights about commercial relationships, obligations and revenue opportunities. He describes the company as working in what he calls somewhat of a double-blind, building models and AI capabilities that understand a customer’s contract data space without ever being allowed to see it. His argument on this episode reaches past that work: once agents are treated as embedded workers, the constraint stops being the model and becomes the data, because the data an enterprise needs often belongs to its ecosystem partners rather than to itself.

 

In this episode

00:58 Welcome, and who Randy Friedman is
01:35 The opening question: using AI to build AI
01:46 A continuum of technology development
01:51 What contract intelligence extracts, and what it is for
02:32 The contract data space, and why no two companies share one
02:57 What large language models provide, and what they do not
03:26 Agentic flows: a team of agents with assignments
03:57 They become the AI that builds the AI
04:36 The constraint: you can only train on data you control
04:58 Data you need but do not possess, and the ecosystem beyond Cognizer
05:27 The supply chain, the value chain, and the next horizon
05:46 Checking the claim: data from another organization
06:35 A supply contract as quality, cadence and availability
07:27 Why the monthly report is not good enough
08:40 Not real time, but continuous, at the cadence the opportunity needs
09:19 A natural continuation, and permeable perimeters
09:40 Is this an agentic workflow API?
09:45 The embedded worker, let inside and not free to run amok
10:07 Available data, not visible beyond its purpose, and the gray area growing
10:27 Healthcare, and one patient across many systems
10:48 Drug discovery, and the same shape of problem
11:20 The whole arc: training data, prompting, agentic analysis
12:00 A different infrastructure for the data
12:34 Why today’s infrastructure is organization-centric
12:58 Where it came from, and what blockchain did differently
13:21 Why it does not scale, and control held by the owner
13:47 Utilization across organizations, and tolerance for latency
14:12 Computational complexity, the edge, and the third issue of trust
14:37 Governance from the bottom up at the data level
15:01 A new fabric that does not exist yet
15:43 Who controls what, how, by whom, and by which agent
15:57 A fundamental change in how data can be used
16:05 Somewhat of a double-blind, in his own words
16:13 Building models for contract data they are never allowed to see
16:35 Why nobody has done it, and what agentic architectures change
17:01 Models behind the firewall, and the limits that sets
17:53 New technology as a chance to re-architect what you already have
18:47 Or does it come from the technology outward?
19:09 Chicken and egg, and what the cloud actually solved
19:42 Virtualization, then collaboration nobody had a method for
20:32 The forward-thinking operator, and the goal that pulls technology along
20:56 COVID as the lesson, and the reuse nobody expected
21:35 People invent new kinds of companies with new technology
21:49 The one-person company where agents do everything else
22:28 New enterprises, and models productized as virtual workers
23:08 One entity is one level of complexity, more than one is another
23:29 Two agents negotiating, with no overlap between them
24:16 Doing the business the contract describes, and overlapping perimeters
24:37 Guidance for executives getting ready
24:56 The people who have been thinking about this all along
25:22 Data that belongs to somebody else, and the vendors who will come
26:10 Resources
26:19 Medium, and Steve Jones at Capgemini
27:17 What listeners should take away
27:34 Bell bottoms, and old ideas coming back with more power
28:24 Deloitte’s people as embedded workers, and the same pattern for AI
29:32 Close

In Randy’s words

“They become the AI that builds the AI.”

Randy Friedman   (03:57)

“you can only really train models, even with agentic flows, on data that you control”

Randy Friedman   (04:36)

“the perimeters are becoming permeable to these organizations”

Randy Friedman   (09:19)

“At a high level, it’s like an embedded worker.”

Randy Friedman   (09:45)

“you have to make sure that data’s available, but not necessarily visible or useful in any way other than its intended purpose”

Randy Friedman   (10:07)

“You almost have to have a governance structure that works from the bottom up at the data level, not at the top-down organizational level”

Randy Friedman   (14:37)

“We’re being asked to help companies develop models and AI capabilities to, to understand their contract data space, but not ever see it.”

Randy Friedman   (16:13)

“Old concepts from old kinds of technology implementations are being reimagined and repurposed in new ways.”

Randy Friedman   (27:34)


 

Resources

Ideas and terms discussed

  • The contract data space: His term for the whole of an organization’s contracts treated as one body of data rather than as documents. It is the reason one model cannot serve every company: the space is different at every company, so each needs its own intelligence to resolve it

  • AI that builds the AI: His compression of what changes with agentic flows. A team of agents with assignments builds the extraction models, from a description of the legal concepts a customer wants captured rather than from a list of words

  • Agentic flows: His name for the generation after prompting. Rather than asking a large language model for answers, you give a team of agents assignments and something close to an anthropomorphic personality, and let them do the work

  • The embedded worker: The metaphor that carries the episode. An agent working inside another organization’s systems, allowed in deliberately, with access to the data it needs and no ability to run amok

  • Permeable perimeters: His description of where enterprise boundaries are heading. Information exchange went from fax to emailed spreadsheet to a shared table to sensing inside an ERP, and agents crossing organizational lines is the next step in the same sequence

  • Decentralized entitlement: What he takes from blockchain without taking its performance. Control over how data is used stays with its owner, while the use of that data happens across organizations, which is the opposite of how databases delegate access today

  • Governance from the bottom up: His answer to the trust problem. Because fiduciary, legal and regulatory obligations sit with the data owner, governance has to be applied at the data level rather than at the top-down organizational level

  • The edge, and latency: The physical constraint on all of it. Richer agentic work tolerates less latency, so the more computationally complex it is, the closer to the data it has to run, which is why gathering everything in one place to analyze stops working once the work is complex

  • Double-blind delivery: How his own company already works. Cognizer builds models that understand a customer’s contract data space without ever being allowed to see the contracts

  • Bell bottoms: His closing metaphor for what is coming. Old concepts from earlier technology implementations, reimagined and repurposed with more power, which is what he tells enterprise leaders to ask big questions about

Named on air

Related AI Realized episodes and events

 

Frequently Asked Questions

  • An embedded AI worker is an agent allowed to work inside an organization’s systems, with access to the data it needs and no more. Randy Friedman of Cognizer reaches for the phrase when asked whether an agent working across companies is simply an agentic workflow API, and the distinction he draws is about control rather than plumbing: the agent is let in deliberately and is not free to run amok. His access rule is that the data has to be available to it, but not necessarily visible or useful in any way other than its intended purpose.

    Transcript 09:45 to 10:27


  • Models can only really be trained on data a company controls, because a company cannot train against data it does not hold. Randy Friedman of Cognizer names this as the limit that everything else in the conversation runs into, and he applies it to agentic flows too: a company can build models against its own contract data space, and through mergers and things of that nature, once another organization’s data is involved, it becomes, in his word, really tricky. His conclusion is that the next step is working with data an organization needs but does not own, which is a permissions problem rather than a modeling problem.

    Transcript 04:36 to 05:46


  • Cross-organizational agents need infrastructure where control over data stays with its owner while the use of that data happens across organizations. Randy Friedman of Cognizer contrasts that with what exists today, which is organization-centric: an administrator sets up the environment and delegates roles and access inside one company. He takes the decentralized structure from blockchain and leaves behind its performance, which he calls really bad, saying it does not scale very well, and he is explicit that the fabric this would need does not exist yet.

    Transcript 12:34 to 15:43


  • AI governance for agents working across companies cannot be centralized. Fiduciary, legal and regulatory obligations mean the organization that owns data has to govern it, so Randy Friedman of Cognizer says the structure almost has to work from the bottom up at the data level rather than top down at the organizational level. He names latency as a separate requirement, before that one: the more computationally complex the agentic work, the closer to the data it has to run, which is why gathering everything in one place stops working.

    Transcript 14:12 to 15:43


  • AI building AI means agents assemble the models rather than engineers training them by hand. Randy Friedman of Cognizer describes the generation after prompting: instead of writing prompts for a large language model, you create a team of agents, give them assignments and something close to an anthropomorphic personality, and they do the work. In Cognizer’s product the customer describes the legal concepts and the meaning they want captured rather than the words, and the models needed to extract it get built automatically.

    Transcript 03:26 to 04:36


  • Executives should start where the demand already sits, in procurement and supply chain, and treat data ownership rather than technology as the constraint. Randy Friedman of Cognizer observes that people running those functions already have making the process more performant, more observable and more controllable on their minds, and that the data their work depends on belongs to somebody else. Getting at it safely, with a certain level of opacity preserved, is the problem he expects innovative vendors to arrive with a rethought process for, and he expects early innovators to try it first.

    Transcript 24:56 to 26:10


 

  • [00:58] Christina Ellwood: Welcome to AI Realized, the podcast for enterprise executives deploying AI. From addressing security, data, and operations challenges to managing the organization on management changes, AI deployment presents the opportunity to redesign our organizations from the inside out. I’m Christina Ellwood, your host for today’s episode, and we’re talking today with Randy Friedman, the chief commercial officer of Cognizer AI. Welcome, Randy.

    [01:27] Randy Friedman: How are you?

    [01:28] Christina Ellwood: I’m good, thank you. How are you doing on this- very early 2025 date?

    [01:33] Randy Friedman: Yeah, very busy, but good. Thank you.

    [01:35] Christina Ellwood: Good. We were talking earlier about using AI to build AI. How are you using AI to build AI for your customers at Cognizer?

    [01:46] Randy Friedman: That’s an interesting question. There’s been a continuum of technology development. There’s a lot of interesting things that have come up over the last, uh, couple of years. I- in-- at the beginning, what we would do is we would, uh, create models. In our case, we apply AI to legal business processes. Uh, our specialty at Cognizer is contract intelligence, taking contract documents, extracting intelligence from that, data extraction, structuring, graphing, insights, and analytics to help an organization better understand their commercial relationships, obligations, revenue opportunities, things like that. Every company has somewhat different data space. If you think of, of a set of millions of contracts a large organization may have, you can think of that as a data type, the contract, and the, the aggregation of all that as a data space. And no two companies are the same. That means that they’re going to need some kind of intelligence to resolve their own contract data space. It’s gonna be unique, and they need some kind of modeling to, to do that. Large language models provide some of that, but not all of that And so the conventional approach had been that you would, uh, train models, uh, with training sets and then aim those models at this problem. And, and that’s evolved to using some large language models and then generating prompts and prompting those large models, uh, to give answers. The, the, the next generation of that, call it gen two of large language models, is using agentic flows. So ins- instead of prompting the large language model, you create a team of agents and give them assignments. And those assignments, really almost an anthropomorphic personality to those particular agents, they go about doing their work and they create the AI that is doing this work for you. They become the AI that builds the AI. We’ve created some application technology that allows a customer to describe the legal context and legal concepts that they’re interested in capturing. Not the words, but the meaning, what they’re looking for and what they’re trying to understand from their giant contract data space And then the AI builds the models necessary to do that kind of extraction on their behalf, and all of that is automated. But that’s using just their own contract data to begin with, so they’re focused on their own needs. Now, through mergers and things of this nature, now they start to deal with other organizational data, and it becomes really tricky because you can only really train models, even with agentic flows, on data that you control. And the next step is in working with data that you don’t possess as an organization, but need. You can think of that as like an ecosystem, where you’re working in an ecosystem, and to optimize your process, you need data from your ecosystem partners as well. That’s a little bit beyond what we’re working with right now at Cognizer, but I can absolutely see how that would work. When you think about the companies you do contracting with, those are in your supply chain, your value chain, and eventually you’re gonna get deeply embedded inside of those organizations from a data and AI point of view. And, and that’s what I think is the next horizon in all of this work.

    [05:46] Christina Ellwood: So let me just make sure I understand what you’re saying here. You’re saying that if I need to be able to use data from another organization, like I’ve got hundreds, thousands of ecosystem partners in my supply chain, for example, and I, in the case of Cognizer, have contracts with all of them, but there’s other data associated with it that’s not contract related, I don’t know, like forecasting data or stuff like that. If I need to be able to use that data, or my-- more accurately, from what I gather, if I need my agent to be able to use that data, then I have to have some way for those agents to use that data and comply with all of the privacy and so forth requirements that are- in place? Is that what you’re saying?

    [06:35] Randy Friedman: Yeah, you’re, you’re, you’re tapping into the right theme. So here’s an example. Let’s start at the top. I have a commercial relationship with a supplier. I’m a manufacturer, and, uh, and my commercial relationship is codified in a contract. And those-- the contract has provisions in it that says that, that you as my supplier have to provide me a certain level of quality, a certain level of a cadence of delivery, availability. We have possibly a return or an inspection process, all of these things that the contract describes So analyzing that contract is one thing, and then enforcing that contract is the second thing. If you think about the natural extension of what a contract is, it’s a, an agreement to do things in a certain way. The extension is you want to do those things in that way. How will you do that? Right now it’s some periodic report. Maybe every month I get a spreadsheet or some kind of a report from you telling me how things are going. Dig deeper. Now we wanna make this dynamic. We wanna optimize this process. We wanna do it in real time. I don’t wanna rely on a piece of paper and a spreadsheet or a report to do this work. It’s too slow. I can’t optimize So now I have to get inside your perimeter, your security perimeter, and gain access to some data that’s describing... That’s what that report was made up of, but not in a fixed static report form, but in a dynamic flow. I wanna be able to actually optimize my demand forecasting and my own production and sales forecasting around what you’re doing, and now multiply that by 10, 'cause I have 10 suppliers, and I have to do that with everybody. So if you think about a contract very dynamically, a contract is just a way in which you’ve agreed to do something, and then what you wanna be able to do is monitor that. You monitor right now, it’s a very slow static approach. AI makes this possible to do it really dynamically, and I wouldn’t say in real time, but continuously. So the faster, the more accelerated the opportunity you’re seeking is, then the faster you want that cadence to go. AI makes this possible. Think about fax machines was how we delivered information, and then it became a spreadsheet that we’d email, and then it’s an air table that we’re both working on, and then it’s a ERP that we have, uh, sensing. This is just a natural continuation of all that. Uh, and that’s what’s i- interesting, is that the perimeters are becoming permeable to these organizations. And, and you, you can imagine the AI has to flow across these organizations in order to really deliver the value people are seeking.

    [09:40] Christina Ellwood: So is that like an agentic workflow API?

    [09:45] Randy Friedman: Yeah, it’s a good question. At a high level, it’s like an embedded worker. If you think about this agent as a worker, or there’s a team of workers, what you really want to do is allow them to be embedded inside your organization. You, you don’t want them to be able to run amok and do anything, right? You need to control them. And they need data to do their job, so you have to make sure that data’s available, but not necessarily visible or useful in any way other than its intended purpose. So if we’re doing business together as ecosystem partners, if you visualize a Venn diagram, that gray area is getting bigger and bigger, right? That’s what this technology really can do. And if you think about, for example, not supply chain, but healthcare. There’s a patient. They’re seen by multiple doctors in different facilities with different systems. But you need a continuum of care and a care plan- That patient is at the center of all of that. How’s that gonna happen? Or in pharma with drug discovery, same set of patients being examined for many different kinds of potential cures and therapies in different clinics, but those clinics are a part of many different clinical trials. The efficacy of all of that work would be so much higher if there was a way for this agentic re- embedded workers, uh, to actually gain access to the information they need to do their work. So there’s this big continuum. It’s gone from, “I’m gonna produce a lot of training data to train a model,” to then, “Now I have a large language model that I’m prompting. I can do RAG-based type of questions with my own content.” And now I realize to do my job and really to accomplish my goal that’s in that original contract, I actually have to do agentic analysis across multiple datasets, not all of which I have permission to... or create. They may not be part of my business. They may be a part of my partner’s business, and that’s the direction things are headed.

    [12:00] Christina Ellwood: So that sounds like a different s- infrastructure for the data, that there has to be a way that this data that is, resides in multiple companies in v- in the system that you’re describing, whether that’s clinical trial data or whatever, that those multiple parties need to be able to allow these agents to do their work in order to provide the value of the insight. Is that correct?

    [12:34] Randy Friedman: Yeah. It’s a great question. There, the, the data infrastructure that’s in use today is principally organization-centric. There is an administrator. That administrator sets up the environment. It can, you can delegate different roles to different users and different access, but they’re not really designed around cross-organizational engagement. They’re around single organizational engagement. They sprang up from the conventional database from however many decades ago. And if you look at the concept of the blockchain, which was a totally different kind of structure, they had this decentralized idea where control was managed in a decentralized way through smart contracts. But problem is it just doesn’t scale very well. The performance is really bad. But if you take those two concepts at a high level and put them together, you need something decentralized where the control of how data is used is in- exclusively managed by its owner, the organization, in a cross-organizational framework. But the utilization of that data is not within that organization, it’s across organizations or by partner organizations. So you definitely need a different kind of infrastructure to be able to do that And the other thing is that the more rich these agentic activities are, the less tolerant they’ll have, the less tolerance they’ll have for latency. So the, the more, um, computationally complex, the, the closer you need to perform that work at the edge. And so this idea that you can bring all the data together and analyze it or put it in a room and analyze it, so to speak, that doesn’t work when these things are really complex. And there’s the third issue, of course, of trust, which is I have to secure this. I have fiduciary obligations, I have legal obligations, regulatory obligations to make sure this data’s managed in the right way and governed in the right way, which is interesting because that means you can’t have a centralized governance structure. You almost have to have a governance structure that works from the bottom up at the data level, not at the top-down organizational level. Those kinds of things don’t exist right now, but we can all see them coming and why that would be needed to facilitate this kind of optimization where there’s a lot of trapped value There needs to be some new kind of fabric that connects organizations together in a secure, trustless way that allows these sort of embedded workers to do their job, and do that in a, in a manner that is controllable and federated and decentralized in its terms of entitlement to the data, so that one organization doesn’t have to abide by the same rules as another organization. They get to control what, what and how and by whom or by what agent that data can be used at any point in time. And, and that’s new. That’s different

    [15:57] Christina Ellwood: It’s a, it’s a really fundamental change in how data can be used.

    [16:05] Randy Friedman: Yeah. If you think about it right now, we operate in a somewhat of a double-blind mechanism at Cognizer. We’re being asked to help companies develop models and AI capabilities to, to understand their contract data space, but not ever see it. We’re never allowed to see it. It’s a, it’s an interesting trick. So now if you multiply that across multiple organizations now, that’s... We don’t have a way to do that, and that’s what we’re talking about. There’s, there-- You can tell why you would wanna do it. People have been talking about this for a long time. There’s just been no way to do it, and now with agentic architectures, that need is growing, right? And there’s a lot of misuse of data. You can see the litigation involved in companies that have scraped data to train their models. We have people we know, both of us in our network, who’ve decided to put their models behind the firewall to really protect the data of their customers, but that provides some upper limits on what they can do. But, and so there’s this tension between what you can imagine the technology could do versus what you want to govern it to satisfy any one, one of your customer set. So that tension has to give rise to some new way of doing this that preserves privacy, preserves that differential privacy and that ability to do this decentralized control, and yet accomplish this lofty objective of really optimizing and improving the outcomes. Those two th- those two things don’t go together with current technology.

    [17:53] Christina Ellwood: But every time we have an, a big, new technology available, in this case AI, we also have the opportunity to re-architect how our existing systems are being used, not just in service of the new technology, but being used, period, right? Because there’s, that tension breeds innovation, but it also breeds the opportunity for us to make changes in our infrastructure. Infrastructure has a tendency to get pretty s- rigid, um, over time, and these are the kinds of things that unlock that rigidity and give us a chance to have some new flexibility. And do you think that is going to be driven from the business need down into the organization to say, “Hey, I need to be able to do these things with this, with our data that I can’t do today”? Or do you think it’s gonna be something that happens from the technology outward that, hey- The technology is being used for this purpose, and its value is constrained by not having the right data access. Which, which way do you think that’s gonna work? Or maybe it’s both.

    [19:05] Randy Friedman: Yeah, it’s, it’s a great question, uh, and it’s nicely phrased. Where’s the chicken? What’s the egg? H- how do those two things relate to each other? And speaking just colloquially, I think the answer is both. So on the, on the one hand, you can think of the cloud as being a technical solution to a technical problem of scalability and redundancy and economy. People had databases that were on-prem, and having it in the cloud means that they don’t have to maintain those on-prem instances and, and figure out how to double their capacities and things of this nature. So virtualization technology really helped in making this possible, and then people started to rethink everything, right? Once they understood that was something they could do, they took a step back and said, “Hey, I’d like to collaborate. I’d like to have more collaboration. I have been able to do that, the same virtualization technology, if I push it up to the application level now, I can m- my people can collaborate regardless of where they are. That’s pretty cool.” So collaboration was always a desired outcome, but there wasn’t a method to do it, and the method to do things don’t always match to the intention. Like we, we all know about drugs that were developed for one thing, and turns out they’re hugely beneficial for something else, and that’s true with technology too. So the really forward-thinking operators, the enterprises that are thinking about their desired goal, let’s say desired goal back to the supply chain example is, I don’t want to disrupt my business. I want to be able to have a smooth flow of goods, and I know there’s lots of changing events in the world, and I wanna be able to adapt to those. COVID taught a big lesson in that. They know they want to do that And they’re gonna pull on that chain, and that’s gonna pull technology along, uh, in a direction that allows them to do that. And some of the things we’re talking about are my imaginings of where that heads. And I think, yeah, you’re gonna see some reuse of technology that people didn’t think was good for a particular purpose, and then all of a sudden, like that drug that all of a sudden helps your hair grow, it was for your blood pressure. There you go, now people are gonna use it in a new way.

    [21:35] Christina Ellwood: The other way that it happens too, Randy, that we didn’t mention yet is people invent new kinds of companies with these new technologies. And when you invent a new type of company, you represent the ability to do things in a way that weren’t thought, weren’t considered before. So if we were starting a company that we wanted to have, I’m gonna be extreme and use Jeremiah Owyang’ prediction that you’re gonna build a company that has only one person, the founder, and everything else is done by an agent. Right. If you really did build something like that, you would have to have a way for those agents to be able to do what you’re talking about. They’d have to be able to work with outside entities and use data outside the organization as well as inside the organization without ever having a human touch something. So if someone actually built a brand new type of company, then the need for some of these capabilities would not only become more clear, but it would be a competitive threat if you didn’t have them.

    [22:28] Randy Friedman: Absolutely. I think that’s what the evolution of the technology will drive is new enterprises, new providers of technology. We can think of some of the work OpenAI is doing. They’re taking their large language models and productizing them as workers, virtual workers that you can hire, pay them by the month, and they do a certain job. Yeah, you need a PhD in biochemistry. That PhD in biochemistry is gonna have to work inside Of some organization, right?

    [23:01] Christina Ellwood: Possibly more than one, right?

    [23:02] Randy Friedman: Yes, massively more than one. And how will they do that, and how will they collaborate? The-

    [23:08] Christina Ellwood: Which is a big part of what you’re talking about, that doing the job that you described with one other entity has one level of complexity, which you’re dealing with every day. But when you need to do it with more than one entity, and there’s differential need for the use of the data and, and rules about, uh, accessibility and so forth, now you’re talking about a whole 'nother level of complexity.

    [23:29] Randy Friedman: It’s fantastic to think about. It’s very inspiring. So use-- going back to the legal use cases at, at Cognizer that we’re focused on, although we are not doing it, there’s a lot of, um, apply- uh, application of agentic, uh, capabilities to negotiating contracts and understanding each company’s, each party’s best interest, and having that agent represent that interest. You can think of those two agents are being on the outside of each of their security parameters, but there’s no gray in the Venn diagram. There are two lawyers and two companies doing this work. You just have agents facilitating that job. But when you take the next step and say, “It’s not the contract itself, but the business activity described in the contract. Now I wanna do that with you,” now we have to put the perimeters overlapping, and you get the gray area, and then you get into this space that we’re talking about that requires new infrastructure, new technology, new data structures, new security paradigms, um, and new collaboration paradigms. That’s what’s coming, for sure.

    [24:37] Christina Ellwood: Mm. Wow. That’s quite a, a future that you’re painting, and I really appreciate that. What guidance do you have for the executives listening to our podcast today about getting ready for this kind of future or for learning more about the possibilities of this type of tech?

    [24:56] Randy Friedman: I, I actually think that people are thinking about this now. They’ve always been thinking about it. If you work in procurement, if you manage the supply chain, on the top of your head are things that relate to making that better, more performant, more efficacious, um, more observable and more controllable. It’s been on your mind for a while as a head of procurement or supply chain. And you work in an ecosystem that the data that’s part of your work belongs to somebody else, and you try to figure out how in the heck in the world can you get that safely and protected, or do so in a way that preserves a certain level of opacity, and you start to multiply these ideas. I think people are wondering about this, and then I think there will be innovative vendors who come along and go, “We have rethought this whole process that you’re in the middle of doing with a completely different paradigm that delivers this outcome you like. Would you like to give this a try?” And I think you’ll find some of those companies, early innovators, embracing this and giving it a try, and that’s where the change will come from.

    [26:10] Christina Ellwood: Great. Any resources that you have to suggest for the, to the listeners? Resources you might point them to?

    [26:18] Randy Friedman: Yeah, you know, that... It’s a good question. I do a lot of reading on Medium. There’s a lot of interesting thought leaders on Medium where they’re talking about agentic flows. Certainly, I follow a bunch of people who write blogs and make observations. One of them is a mutual friend of ours, Steve Jones at Capgemini, who’s talks a lot about these things. I think that in terms of what we’re talking about, you can see the handwriting on the wall with what, uh, OpenAI is doing and what Microsoft is doing and what Google’s doing. And I think everybody understands these things in a somewhat different way. It takes some thought and some imagination to figure out how you could apply that in a productive way. And- Overall, I think that’s what I’ve been doing in terms of expanding my thinking and trying to create new ideas and figuring out what’s next.

    [27:17] Christina Ellwood: Great. As we wrap up, what would you like our listeners to take away from our conversation today?

    [27:26] Randy Friedman: Bell bottoms. Yeah.

    [27:28] Christina Ellwood: Please no.

    [27:29] Randy Friedman: So I, I mean that in a funny way. It used to be bell bottoms were very hip, and then they went away. Nobody thought they were cool, and then they came back, and all of a sudden they were interesting again. I think this is what’s gonna happen, that’s happening now. Old concepts from old kinds of technology implementations are being reimagined and repurposed in new ways. A lot of what you’re gonna see next has been a repackaging of what was before, and with an extension, more power, more capability. Remarkably, there’s a lot of older ideas that can be re- con- reused and repurposed in new ways, and I think that we’re seeing that. And the other would be maybe anthropomorphizing a little bit more this technology. If you think about a company like Deloitte, we’re, we’re very proud at Cognizer to have a good relationship with Deloitte. Deloitte has a few hundred thousand people who work all over the world, and they work inside their enterprise customers’ perimeters. They become part of their organizations. They become essentially embedded workers. Now, if you can anthropomorphize AI, and you’re gonna see the same pattern. H- how do you make that happen? How is all that going to happen? Once you have the question, you can begin to ask the critical, critical de- uh, deconstruction task of, okay, how would I make that happen? Once I understand that’s a possibility, I think everybody starts to say, “How do I unpack that into something that’s actionable?” And I think that’s fascinating. I think that’s what enter- le- you know, enterprise leaders should be doing. They should be asking these big questions. They have the template. They just need to apply it. It’s like bell-bottoms. Like, okay, we’ve had people doing embedded work for decades. Now, how would that be done with AI? It’s an old idea, just rethinking it.

    [29:32] Christina Ellwood: That’s lots of good food for thought, thank, thank you so much, Randy Friedman, commercial, chief commercial officer of Cognizer AI. Thank you so much for talking with me today on AI Realized.

    [29:43] Randy Friedman: Pleasure. Thank you.


 
Listen to other episodes
 
 
 
Join AI Realized
 
 
 
Subscribe to the newsletter
 
 
 
Christina Ellwood
Previous

Bridging Data Science and Generative AI
Next

Write the AI Policy Before You Write the AI Feature