Panel · 4:50 to 5:25 PM · Wednesday, May 13, 2026
The world. Alright. Tokens. Tokenomics. The journey to tokenomics. So FDE is very important. Tokenomics has come up throughout the day as well. The cool thing about tokens is the way we sell a lot of our research. If you have tokens, you can get a Horizon. If you have tokens, you can get Phil to speak at an event. If you need tokens, talk to Neeraj. Neeraj, raise your hand. There's Neeraj. He might even give away free tokens by the end of this. Oh wait, we're talking about different types of tokens, aren't we? We're talking about tokens for AI. Well, Neeraj, like me, is one of the few people in this company who don't have AI in their name. I mean, David, who's speaking, is defined by AI. Can't have David without AI. Ashwin, the same. Ashish, another one of our key analysts, all have AI. Dan has AA in her name, but we won't go into that. This is how I kill time.
One of the best things about tomorrow. So I know most of you, you know, you've been here today. We're going to do it again tomorrow. Tomorrow, Rohan is going to co-MC with me. Rohan leads our industry-focused areas. He's driving the analyst to write playbooks to health care, banking and financial services, telecommunications, media. So he and I are going to banter a lot about industry-specific ways to how-to AI, and hopefully we'll see that cascade through a lot of the discussions. Also tomorrow in this room at lunch will be the AI-First Deal Lab. The deals lab, a lot of you heard about. Some of you expressed interest about seeing what the hell this thing is in action. Be here at lunch. Get your food. Come back in here. There's going to be some tables in here so you can actually not sit and worry about that meatball rolling down your jeans. So that's going on. The data suite is also in 203. If you haven't seen it, talk to Ashish. Talk to Jason. Talk to Hansa. They're all manning the demos. They would love to discuss what's going on. Obviously, see our sponsors, our tech showcase in the hallway. Talk to them during the drinks. See what's going on from JK to Lyzr to Cognitos. Lots of cool stuff going on there. Interesting opportunities to partner and kind of think of new ways to go.
So with that, I'm just waiting to… We're going to skip over a question we didn't do. The flywheel. Wow, there's some slides hidden. From rate cards to tokens. I kind of look at tokens, and I had this recent conversation with some of my friends at a big four, as the tokens are sort of this bridge between the traditional FTE model and outcome-based models. Right now, they're kind of in the middle as we move to how we're consuming AI to do stuff, and we're measuring it on tokens. As you saw from Phil's view here for the keynote, a little example of where that can go awry when you only game on it. So what does token economics mean? Where is David and his panelists? See, that works. There's a little TV out there that works on cue. But please join me in welcoming David, Surojit, Cliff, Ashok, Bijit, Umang. Where did you come from? You broke protocol. Shook it all up. All right. With that, David, you are the panel holding them from drinks. Oh, thank you. So make it good.
So I'll just go really slow now, yeah? Everybody just be patient with us. We're going to draw this out as long as possible. Honestly, no. But look, thank you guys for joining us today. I wanted you just to do a little intro, and then we'll get into some of the harder stuff. So starting with Surojit at the far end.
I'm Surojit Chatterjee. I'm CEO and founder of Ema.ai, Enterprise Machine Assistant. We're an agentic orchestrator, build agents for HR, IT, and customer support, primarily based in Silicon Valley. Previously, I led mobile advertising for Google, and I was chief product officer at Coinbase and helped Coinbase go public.
Cliff Justice, recently retired from KPMG at the end of the last year. Currently, I'm involved with a number of AI startups, sit on some boards, working through a number of advisory engagements helping companies get started. At KPMG, I led enterprise innovation. We had a ventures group, made investments in startups. We had an internal studio and incubator, built startups internally, and I led our AI disruption for a number of years within KPMG. So kind of looking at transforming the firm around AI business models across the multiple functions at KPMG.
Hi, Bijit Ghosh. I manage the portfolio for one of the private equity companies. Underneath, we have more than seven different companies, some of them in the growth stage, some of them are in the seed to the early stage. I provide the practice of the engineering, the product, and the GTM, how does it really look like. And prior to that, I recently moved from Wells Fargo. I used to manage the consumer banking and digital banking AI practices, applied AI team and the complete enablements of the AI, including the agent orchestration platform.
Awesome. Hi, everyone. Ashok Panduranga. I lead AI strategy for US Bank. I primarily focus on operations business. We have about 19,000 folks in operations, and my job is to translate strategy into execution and how do we drive value. I am more focused and biased towards execution and value than strategy. Look forward to the discussion.
Umang Nagpal, Executive Director, Office of the CIO Chief of Staff within the CIO Organization of Consumer and Community Banking at JPMorgan Chase.
Thanks very much. So we'll get straight into this because I think we need to define some terms. We've been talking about tokenomics. It's come up a few times today. So what the hell does that mean? What does it mean for you? Let's just pick on Cliff first.
So to me, it's a continuation of the unit of cost that's relevant for the world we're in now. You know, when we went to outsourcing and offshoring, you know, it became FTE. When cloud came along, we had a new unit of cost, which was compute second. And now that required a whole new discipline to be invented, FinOps, to calculate that and manage those costs. I don't know if you remember, but in the early days of cloud, the expectation was we're just going to get a whole lot of computing for free. And then the economics of that really got upside down really quick. And there was a lot of turmoil in the IT finance groups about how do we measure this cost. And that's kind of how FinOps came around. And then now we're dealing with tokens, which isn't one for one as a cost of compute, but it's a cost of intelligence as you look at how AI is measured in the form of a token. And so token, it's a pretty good proxy for the cost and for the unit cost for economics. It's not exact. And there's going to have to be some discipline. The discipline of FinOps will likely have to evolve to account for the cost of intelligence. It's moving so fast that I don't think anybody really has a full grasp of how much intelligence is going to be required. Jevons paradox is going to be playing out in full force because as this becomes, as this materializes in the business, the demand for this is going to skyrocket. So it's very hard to predict what the intelligence usage is going to be over the next two, three, five years. Think about what an outsourcing contract, how unpredictable that was and how difficult
That was to price. And yet we're seeing increasing compute capability at the same time as cost. Those tokens are falling through the floor. Something like a hundred times a year, the differential. So how do you write a contract in those kind of conditions becomes really difficult, I think. The Jevons thing, I think, is going to be massive because if we think that this has only really touched software so far. It's really only touched writing code. There's an awful lot more for this to go at. So personally, I'm a believer that there is not a ceiling in sight to the kind of demand we're looking for. So the token thing, then, is really a proxy for work. And, of course, it's never been a good thing just to pay for work. So what else do we need to add to this mix to make tokenomics work? Anyone leap in?
I can take a stab. I think, look, the proxy for work is kind of a dangerous thing. So you can easily do all kinds of nonproductive work. I have a, you know, I was talking to Uber's CTO recently, actually, in an event. And he was saying their token cost is 15% of the salary cost of engineers. And they don't know if it is giving any ROI or not. Because they have just said, go use as much, experiment, and so on. I think the industry has to move to an outcome-based pricing, which is, I mean, we do this all the time, which is you pay for actual useful work, not just any work.
Any other thoughts? Yeah, I fully agree with Cliff and Surojit. I'm just trying to reinforce that we have seen the consumption model in the SaaS and also how does it really works with your FTE, right? So I think it's a new enterprise resource model where we see how it's going to maneuver with the outcome-based model. We want the token to be consumed in the enterprise so that they can do vibe coding, they can play with the product, the prototypes first play that game so that we can cop out what is really needed. But I think it's mostly maneuvering to the outcome-based pricing model where you have the reasoning, you have the chaining, but with the outcome-based.
So it seems to me there's an initial step, which is get the thing working.
Yeah.
At which you're quite at liberty to spend what you like in terms of tokens. But then it becomes something you have to operationalize in order to scale. And then you have to get a little bit smarter with it. So where are you seeing the application of that kind of smarter with it element working in any of the companies you've been working with?
So, you know, at JPMorgan Chase, you know, we talked about scale. In my simple mental model, I mean tokenomics is essentially just like kilowatt power. It's a currency of sorts where you pay for the piece of work, but not every token may cost the same. What I mean by that is you could use a token with quote-unquote a cheaper model to actually pay less. And that's very critical because in my mind, I break every token request sent out, processed in three tiers. Tier one would be a simple command which may say summarize this email for me. And this email may be two pages long. There you don't have to use the most advanced sophisticated model because you're paying extra for nothing. Tier two may be what you may want to call as a single agent. Think about you call a call center and a wrap invokes an agent to actually pull up your information and then uses that to actually give you whatever information you need as a customer. Here itself, there is a good possibility the simple, cheapest model can do the work for you. It's gonna cost a little higher because more than quote unquote one token will be used, but it's very defined. Tier three is when you want to actually call about orchestration, where you have an agent which is orchestrating multiple agents. And this is where, quote unquote, the real problem may get amplified because if you don't define what you want to do and there's no clear end point, it can keep consuming tokens to whatever model it thinks it needs to use for that certain work. Think about if someone says, I have a contract, do a risk-based analysis on the contract. And I've seen contracts anywhere from 100 pages to more than 1,000 pages. Since it's so ill-defined, you'll not only look at that contract for analysis of risk of this contract versus several others, and there's no end point, it can go ballistic and use all your token that may be allocated for the entire week or month. So the point I'm making here is you've got to, as an organization level, you've got to first estimate how many tier one, tier two, tier three you may have. And then, also start defining what your tier three work should be. What should be the end point? You shouldn't ask a general question, give me insights. Just say, give me top three insights, period. So it's very defined, it's boxed, and now it can deploy whatever model it needs to deploy.
Okay. Anyone else got any examples where?
I'll argue, I think Umang's points are great, but I'll argue it's almost impossible to implement that at scale in any enterprise, because when you are giving a task, you don't know the boundaries of the task. You don't know how to specify exactly how the execution will happen. The approach we have taken, so the problem is, I think, the theoretical level, that's exactly the approach that needs to be taken. But I think you need some kind of a solution, not like everybody cannot probably go and discipline every user. The approach we have taken is we have a model called Ema Fusion, which is exactly what you are saying. It's a small model that sits on top of 100-plus models, takes an enterprise task, breaks it up, figures out how to execute optimally using multiple models or single model and iteratively and asks these questions back to the user to optimize cost. So you need something like that where the system actually helps you optimize. Today, I'm very concerned the way, say, Claude is designed. It's almost on the other side, which is like, come spend as much token. Oh, you want me to do more? It's like getting addicted to token and make you spend.
Yeah, 100% agree, right? I mean, the way we are approaching it is how do we bring the operational governance into it and discipline into it through telemetry? Because you can set all the thresholds you want. It's very difficult. It's a dopamine, right? I mean, once you start building agents and then it starts consuming, it is very difficult to measure. So one through operational discipline, second through telemetry, and third through architecture. There are tons of things you can do in architecture, architecting an agent, and have a decent idea about how much are you going to consume, how much are you going to give the context window, what model I'm going to select. There are so many things that you can control there and not look downstream, right? So try to control it upstream. That way you can have some semblance. Again, all of us are trying to figure out, I think if anybody says in the room I have figured out how to control agentic token consumption. I would like to talk to them. We're all trying to figure out and evolve as we try out things.
Okay, well, just to add some more complexity to the problem, let's talk about how anyone should think about comparing human AI or hybrid delivery. And ask ourselves, what are the real trade-offs between an FTE-based model and a token-based execution. And how should we think about the cost per task, productivity, speed, quality, and ROI in one model? This seems to me where we're headed very rapidly. We've heard talk today about people expecting to manage teams of agents. My expectation is that we all become managers. Anyone in any organization becomes a manager of teams of agents, but they are also managed by agents. There's a three-way view of this. So we're very much moving towards a hybrid model. I don't know how quickly we get there, but we are going to have to think about the trade-offs, aren't we? How are you guys approaching that? Has anyone tackled that challenge?
I think we, as an enterprise, we need to see from the portfolio perspective, where the human element is differentiating. We are great in strategic thinking, we are great in decisioning, we did a great work in terms of the relationship. I think that's a given, right? And we need to see where the agent can scale phenomenally, right? With the speed and also with the predictive understanding which agent already have. The mechanics should be hybrid where we can fuse both together and get the better reasoning and chaining to do work in a proficient way. And I think most importantly, the token economics, or we call it token maxing, is giving an opportunity where we have the biggest business value and the outcome we can translate into and create a hybrid mechanics where we need to operate the most.
I can add to that. You know, I would go back to how many of you have heard radiologist paradox here? Okay. That's good. And so 10 years back, when AI and ML picked up this team, one of the first jobs AI was going to eliminate completely was of that radiologist. It was so easy for AI to look at scans and then infer data, extract data, and then give it so that next five years we don't need radiologists. Guess what? Today, average salary of radiologists in the U.S. is $400,000 to $600,000, and demand is increasing. Add to that, next six to eight years, we're going to see shortage of 40,000 to 50,000 radiologists in the U.S. alone. Okay, what happened here? I think the way we need to look at some of these processes, jobs, is we need to revisit and then see how it is going to evolve in the era of AI. It's not like gloom and doom. It's not like we're going to eliminate a certain set of jobs. It's like how we augment AI with human capabilities. Yes, token-based AI can deliver task-based, repeatable, and things like summarization, doc extraction, and so on, coding assistant, and so on and so forth. But you need human expertise for the judgmental, high-risk, edge cases, and so on and so forth. I think it's time we take a step back and then certainly look at how are some of the jobs going to evolve in next four to five years, then talking about, okay, how many jobs are we going to eliminate?
Cliff, I wanted to get your view on this because…
Yeah, I think it's easy to try and kind of group all of this into one model or one answer. And you have to look at the task or the job to be done and the complexity of the job, where AI fits, where humans fit, and how the customer consumes, the end user consumes that output. You're going to have a change in demand, to your point, you're going to have a change in demand because the economics of this are going to change. Costs are going to go down. Demand for certain tasks are going to go up. Some will stay the same. Some are just kind of back office commodity tasks. But when you get into something like radiology, the cost comes down because machines are reducing the cost. It's causing MRI centers to pop up everywhere, and you can go in for $300 or $400 and get an MRI of your shoulder or your knee. You don't even need to deal with insurance. And so demand is spiking. That's causing a demand for radiologists to sign off on the machine read. It's an exploding business. So you have to look at each one of these jobs to be done before you can, you know, develop the operating model around that task. What does AI do? What does a human do? How does a user consume? Because it's going to be very different. Some commodity tasks are, you know, you're just going to be driving the cost down. It's going to be dependent on the business that that commodity task is supporting. In some of these more core business operations, that's going to be different. That could change everything. You know, when we see robots come on the scene, physical robots, it's going to have a very disruptive effect on the cost of everything. And the entire economic value chain is going to have to be rewritten. And whether that's tokens or other units of cost, human cost, the skill sets, that's all going to have to be redefined. We're going to be in a very disruptive period over the next three, five years.
I'm kind of interested in what order we need to approach this in. Do we think about how the organization works first? Or do we think about it task by task and learn from that and scale that into the organization and make the organization different as a result?
It's likely all of the above dependent on what job there is to be done. I would take it as the job to be done first and look at the overall job and then break it down into those tasks because the operating model is going to be very different because the work that you used to do is no longer going to be considered work. It's just going to be absorbed into technology.
Okay, so this is going to require a different set of skills to understand how to make this work in the first place. What's the kind of expertise we're going to need to develop to make this work? I'll open that up to anyone who's got a thought.
I'll pay for that answer. All right, go ahead.
No, we are all figuring out. But the way we have looked at it is I think we spoke a lot about process debt, tech debt, data debt, everything, all debt today morning, right? And we keep hearing them in most of the conferences too. I think we need to take a step back. Again, look at some of your complex processes. How are you servicing your customer? Five years from now, is your customer going to consume the service the same way as he or she consuming today? That's going to change significantly. In our banking, we can tell you so many things have changed. The way the Gen Zs, millennials consume some of these versus the baby boomers are completely different, and we need to cater to everybody. So breaking that customer journey, breaking the process into consumable and executable tasks, and then see where AI can definitely help, right? I mean, AI can surely increase productivity in some of the repeatable tasks, some of the tasks where you can actually measure and then augment human capability. There's no one fit size all answer here, but it depends on the business models. It depends on what are your top 10 complex processes that you're trying to solve for and how you're gonna put tech on top of it. I think tech should be the last one that you should, you know, thinking about solutioning.
If you think about the expertise layer and what we've considered as you know skilled skilled work it's going to be very hard for AI not to move into that space so that that expertise level will be performed by agents as they get smarter and smarter but the judgment is not the judgment is still going to write that's the radiology example you know we're not just going to turn it all over to AI. Somebody has to be held accountable. It's our legal system. It's our culture. It's everything. It's going to require some level of human interaction and judgment. There's so many undefined characteristics that go into a decision that that judgment layer has to be there. There's going to be a higher demand for that judgment layer because of the volume of demand for the lower cost expertise, which will be provided by agents.
Yeah, a slightly contradictory point of view. I think if you look at an example like self-driving cars, say, oh, humans should have the final judgment and so on. Turns out the error rate of self-driving cars is far, far lower than humans. And you actually don't even want the human in the seat at all. So the way more cars are better than the Tesla's where the human takes control because human gets worried and takes control and makes a mistake. We work with very large enterprises. We see this all the time. Customer support consistently, consistently. AI employ. We call our agency. I employ their agentic systems, multiple agents. It's a slightly provocative term, but our view of the future is where AI and human are working together. AI is augmenting human workforce. You have to treat them as kind of other employees. Consistently, higher CSAT, a better judgment, higher response. Very consistently. Of course, there are sensational news like Air Canada and so on. Just like there is a sensational news like, oh, one Waymo car is stopped in San Francisco. Yes, every day there are so many humans make mistakes that doesn't make the news. We have seen in healthcare similarly, if you look at diagnosis, actually AI is performing a lot better than doctors today, this is just published data. So I do think humans will have a role, that role will evolve, maybe higher level kind of judgment, but a lot of judgment that humans are making are not that great. I mean, radiology is probably different, it's much more experts, but certain roles will forever be changed like customer support. A lot of back office will be very dramatically evolved. Even coding software, we have seen if we train our AI on the code that humans have written, it learns bad decision-making. It's actually better for AI to write code on its own, and it will do…
So I think this becomes probably our last question. I just want to add one point.
I'll take a middle seat between Cliff and Surojit. I think as you're scaling into the enterprise, if you know the mechanics of LLM as a judge, right, and if you have a multi-agent running simultaneously, this is the only way you can scale it. And again, the human taste and the personalization bring to the picture where we are not there yet with the baseline or the frontier model, but we will get there very, very much, very soon as we're optimizing, as AI is optimizing AI, right? So I think that's the theme of how we're seeing the ecosystem is changing and maneuvering to lift that. And I think very soon we're going to see that if you have played with your open Claude and doing some home automations and all, I think it understands your mechanics. It understands your taste very well. And you don't have to repeatedly say the same thing, right?
Yeah. So I just want to go back to the point about driverless cars because I think there's a question around trust there. How do we get to a place where Waymos were allowed to run around San Francisco and then get extended all along the Bay Area and down the 101 there? That took an awful lot of proof points. And I'd just like to ask this kind of final question, more focused on the guys who are buying from the guys who are providing. What's it going to take for you to trust in the whole tokenomic model?
So I'll take a stab and I think I'll kind of go on a slight tangent. It'll sound like a tangent, but it's not. So if you think about driverless cars, there was a lot of data available which essentially became a baseline. Miles driven, per person, per day. And you can actually see if the accident rate has gone down with driverless cars. In a lot of the work we currently do, imagine you, if you're an auditor, you're being told to actually see if there's an issue or not. I don't think there's a foundational baseline available. And we also need to understand and be very clear, AI is a probabilistic model. A lot of work, especially in the financial industry, when we are reporting numbers, is deterministic. And there's a level of accuracy that has not been achieved for reporting those deterministic values using probabilistic models. So, the reason I bring that up is this is purely from a use case point of view. You really need to be sure this is my use case and if this model, be multi-model or multi-agent, whatever approach you use, is able to deliver that and you can compare to a baseline of sorts. Because fundamentally, financial industries are trust-based industries. You go to your app today, tonight, tomorrow morning, you need to see the same balance. Even if penny's off, you'll be freaking out. You can't actually open an app and say, today your balance is X dollars, but we are not fully sure. Go to a banker and ask. You can't do that. So the point I'm making here is, again, when I share about boxing, what you want, it has to be very clear, and you need to approach it from that point of view, and then essentially see if AI as a solution is a good fit. And the back end, when you're using the tokens, you need to be very clear for this token, this is the cost I'm paying. Because there is a lot of opaqueness right now. And especially if you go to the SaaS contract, it may be buried along with 10 different line items, this is the cost, this is the maintenance cost, token. And you need to be very clear. What am I paying for? How much are being allocated, if I exceed, if there's a penalty, if I actually don't use that, is there a penalty, and what kind of additional benefits. In many ways, this is cloud 2.0, because we went through the journey, because organizations, the reason hyperscalers, I hyperscalers, because you actually can go to a GCP website, tells you, for this queue, this is how much you're going to pay. Now, at an enterprise level, you can obviously create a contract, get the discounts, and at the back end, this is what the FinOps organization is doing. It essentially has created the dashboard at AWS site, and you have your own dashboard where you have multi-tenancy, you can actually compare cost, and you add that contracting discounted there, and you actually see, and a similar system will, in my mind, evolve in this tokenomics as well, where you have different providers, be it Bedrock or others. They have a dashboard, you have a dashboard, and there's a reconciliation of sorts. But that essentially is the accounting piece. But fundamentally, you have to start — this is what I want, can AI deliver or not, and what model do I use to actually make that happen?
And the models can shift their performance on a daily basis.
Potentially, actually, if you now use Claude, even a free version, it actually automatically allocates the model. So it's not rocket science. See, we talked about architecture. You can actually now actually not just have, you can have multi-model approach, you can have open AI and Anthropic models right there. And if you now think about integrating the cost element through your contract at the back end, based on the request itself, it's gonna actually pick the automatic model. You can build that as part of your architecture. Sure.
I'll just add to that, right? I think it's, yes, absolutely the transparency, the telemetry underneath, and reasonable explanation about the consumption is important. But again, point to note is, tokens are just activities, right? I think it's important for us to move towards the outcome and then break that outcome into unit-based pricing. How much am I paying for per-cost conversation with client? How much is going to cost me to extract information from complex lawyer contract? How much it is going to cost me to write a complex code with agent assist, right? I think once we start moving towards that, economics will start making sense rather than burning tokens and then expecting the transparency from providers. I would love to get that, but I think, you know, I'll just give one example, right? Today, 40 to 50% of the contact center calls in utility companies, cable companies, and then your telephone companies, are related to consumption and billing. And what are we consuming here? Gallons of water, wattage of electricity, and then number of data, and then calls. It is so deterministic, right? Even there, we struggle to provide the transparency and then translate that to billing. And then here, we are talking about probabilistic output that token is going to generate for you. And then how am I going to measure if I am so focused on input and activity and not focused on outcome? I think we need to pivot that equation a bit.
That's a great question. We're gonna have to wrap up, unfortunately.
If you don't have an account, Register here |
With the exception of our Horizons reports, most of our research is available for free on our website. Sign up for a free account and start realizing the power of insights now.
Our premium subscription gives enterprise clients access to our complete library of proprietary research, direct access to our industry analysts, and other benefits.
Contact us at [email protected] for more information on premium access.
If you are looking for help getting in touch with someone from HFS, please click the chat button to the bottom right of your screen to start a conversation with a member of our team.
