Gam Dias: Agents Unleashed, understanding the Agentic AI stack
An interview with the author (May 8, 2025)
The following is a transcript of our recent interview with Gam Dias on Masters of Privacy. The original recording can be found on the podcast’s website, as well as on your favorite podcasting feed.
Gam Dias is a seasoned technologist and entrepreneur with a rich background in software engineering, AI, and product innovation. As a consultant, he has helped write the data strategy for Fortune Global 500 companies, innovative startups, and ambitious non-profits. He has a degree in Computer Science from the University of Liverpool and an MBA from Warwick Business School. Gam has lived in London, Leeds, Salt Lake City, Santa Cruz, San Francisco, and he currently lives in and works from Madrid, Spain.
Gam’s latest work, Agents Unleashed, distills years of experience into a compelling look at the rise of autonomous AI agents and their growing role in marketing, sales, and beyond.
INTERVIEW
Sergio Maldonado
Gam, thanks for joining me. Again.
Gam Dias
Hey, good to see you, Sergio.
Sergio Maldonado
I understand that Hubble process analytics had a lot to do with your path, with the path to writing this book and getting into AI agents. So what’s the connection between Hubble and agents?
Gam Dias
So as you know, I come from the whole Oracle ecosystem. Salesforce is very, very new to me. And as I got deeper and deeper into it, I found it more and more interesting. mean, Hyperforce and the whole built on Oracle thing and the way they’ve kind of built this massive configuration engine. Fantastic, fantastic idea. And so we were building process analytics. It is a non-platform application to actually understand the business processes that are running through Salesforce. It’s certainly essentially a process mining application. I handed off the reins of that product to the Hubble technologies team and they are selling and marketing it alongside Hubble Diagnostics. And so one of the things we discovered was if you are automating processes using Agent Force, the first thing you need to understand is the process you’re automating. And Hubble does exactly that. You also need to understand whether the org that you’re implementing on is up to running Agent Force. So you need to relieve it of tech debt. And so the two products together, really serve Agent Force very well. And so as I was coming off of that, I wanted very much to document my knowledge and understanding of how you apply both products if you’re thinking about building agents and Agent Force in particular. So it really started off as a guide, a deep guide to understanding your org and the processes around it before you apply agents. So that was the genesis.
Sergio Maldonado
What was exciting about the Salesforce ecosystem?
Gam Dias
So Salesforce has built a religion or a movement. It’s quite amazing. I only went to Dreamforce a couple of times and just to see the energy there around it. so Salesforce has done some amazing things in the last 20 years. And when I saw, I didn’t actually attend the 2024 Dreamforce, but I saw Mark Benioff talking about agents and the energy that came from that. It was quite a show. It wasn’t Dreamforce, it was Agentforce. And as I saw the energy, the positivity, and the objective of a billion agents, it’s textbook how they launched this product. It was pretty amazing. As I was watching it, I’ve been in the game long enough. I’m kind of old in this game of enterprise software.
And what I could see was, you know, all this hype, it’s going to be fantastic. Everyone’s going to be really pumped. They’re going to come out of the show and they’re to go, I want agents. And everybody’s going to kind of bowl along in the first couple of quarters. Everybody’s going to be building agents and the SI partners are going to be pushing it. And my gut feel said, hang on a sec. This is never so simple, you know. And it’s the old Gartner hype cycle. We go up. We have maximum hype, which should be about now by my reckoning. And then we hit a trough of despair. There’s always something that goes sideways in this type of thing. so my gut feel said September, by September, we’re going to be in that trough of despair. An agent force isn’t going to live up to the hype. So that was really why, I mean, I felt really excited, but I knew there’s going to be a fall before it all picks itself up. I mean, it’ll resolve itself. Like come 2025, we’ll be back on track, but we’re going to hit a period of depression.
Sergio Maldonado
So something that you say is that the underlying data infrastructure and the data governance layer, everything has to be in order and that that will condition everything. So revisiting that prediction (a billion agents), or your assumption, versus where we are now — do you have any data?
Gam Dias
I’ve seen numbers saying about 5,000 customers. And I’m reckoning probably about 5,000 customers might be providing, you know, 50,000 agents at some point. So, you know, by the end of the year, they’re gonna, I mean, let’s think about this. know, an agent, don’t deploy an agent. I mean, it’s not gonna be like that. When I share something with you, I’m not going to share a document anymore. You know, actually, you’re a lawyer, you’ll appreciate this. When I share a set of contracts, well, you know, I’m, you know, I’m a client, I share some contracts with you that I’ve had previously written, I’m not going to share contracts, I’m not going to share documents, I’m going to upload them to some sort of GPT or LLM, and then share the LLM. And you can query that. And so if you think about that notion, you know, the way I’ve put my resume, in terms of a chat, you can chat to my resume now. So if every document turns into an agent, we’ll hit a billion agents really, really fast. It’s not going to be a problem. Then so in the Salesforce world, I think, you know, they’re starting slow. It’s going and people are taking the agents. But it’s a little bit more trouble than you see on stage. On stage, it’s all very simple. You know, you just spin up an agent, you kind of write in text what you want it to do, you hook it to the flows that you already got. Fantastic. And all of a sudden you’ve got a working agent and you’re off to the races. In reality, you know, there’s some bumps. And so that’s really what the book is trying to look at resolving or getting ahead of those bumps and obstacles.
Sergio Maldonado
Were they too eager? Did they get ahead of themselves with the pricing model? Asking for $2 per conversation, has that put anyone off? Has that had an impact?
Gam Dias
I think it has. But that’s hindsight talking. It’s very easy to say here. If you step back and think about, OK, I’m looking at the most popular, most useful agent is probably going to be the call center agent. If your pricing is $3, $4 a call for a resolved call, $2, half the price. Great. Sounds good. It’s not quite working out like that. I think, though, this is a time of discovery for absolutely everybody. We look back at how Sam Altman said he’s priced ChatGPT, now, $20, $200. It was a guess. And it’s not enough. Even they understand that, you know, if you think about the processing costs, you know, all of the investment money is subsidizing, you know, everybody buying into becoming a way of life. That’s what they’re buying into. It is the whole generative AI is massively subsidized. Now, in Salesforce case, $2 is great for certain circumstances, you know, if you’re putting them into a call center. If I’ve got a query engine that a high volume sales rep is going to use as a sales coach, they’re going to be hammering on that, you know, and so your sales rep might be running up a couple of hundred dollars a day in asking it questions. You know, it’s, there’s a lot of different pricing models out there.and that we’ve seen, you know, that there is, you know, usage, there is by the server, there is by the call, there is Play, no, it Paid? Paid have just launched a Pay On Success.
As you know, if you’ve come from consulting and are trying to get a consulting project to be, you know, we’ll do it and we’ll take a percentage of your success. It’s like, that’s really difficult to measure. and so you’ve got that. And then if it’s a different sort of problem, you know, that perennial attribution problem of clicks, if you’re doing it differently, you’ve got that. So I think pricing is something that we haven’t figured out yet. My hunch is that Salesforce, within the next three months, is going to revise their pricing and come up with a better model.
Sergio Maldonado
Okay, very good. So before trying to understand what people that use Salesforce can do now, what’s happening with your analysis? So you wrote this book, there’s another volume coming up. Why? What is it that you left out?
Gam Dias
I left out half. What was happening was I started writing in December and by the time I got to like the first five chapters, I realized things were changing. One of my chapters is about headless business and Salesforce hadn’t launched it. And then in about February, I see announcements for headless and it’s like, okay, so I’ve got to get it in, and the book was already 200 pages.
So thought, I’ll just publish now, I’ll publish it in time for the Trailblazer event in March. So that’s what I did. I put out five chapters and I’ve got another five coming. the five coming deals with, actually the first volume was all about Salesforce here and now. And the second volume goes outside of Salesforce. And the premise, why it’s going outside is because You can build and deploy agents inside of Salesforce. These agents are going to go off to the web and, well, primarily, they’re going to use as leverage the flows inside of Salesforce. And those flows are going to be able to send emails and carry out actions. However, what we need to understand is agents are going to actually negotiate with other agents.
And so an agent might visit another website and be received by an agent. And the agents will have a conversation without the humans. And then they both come back to their results. So agents are a really good word. It’s like an assistant that goes off. Your people will speak to my people. That’s kind of what’s going to happen. And so the book is really covering that. It’s covering how these agents need to have interoperability, how they need to have needs to be provenance on what it is they do. If I’m exchanging information, if my agent is exchanging information with your agent, there needs to be provenance on that exchange. What information can you see? How long can you see it for? And what can you do with it? So there’s another piece to that. There’s ethics. There’s probably the economics of agents. When I’ve got thousands of agents executing, how do I measure them? How do I measure the swarm?
Because I can’t make individual actions anymore. The agents need to kind of moderate themselves. But the act of the swarm, how do I ensure that it’s aligned with what the business is trying to do? If I’ve got agents in sales and agents in marketing, this is how I generally explain it like this. If I’ve got an agent sending out nurture emails, and I’ve got another agent that is taking the process from lead to close that or part of that. So I’ve got one agent that is traditionally optimized for opens and clicks, which is great. So I’m going to maximize those. And now I’ve got another agent on the other side that’s receiving all of this stuff and receiving qualified leads. What can the agent do, what can that agent do to close? And I’ve got such a large volume of leads coming in that I’m on it. Well, I know agents are sort of infinite, but the type of activity I might not be closing them all. So both agents should actually work on the same metrics. So how do I orchestrate across functional agents? And that was a problem that we’ve seemed to manage to solve in human teams. You know, what’s the handoff? What KPIs do we share? As we start to introduce agents, we need to rethink our KPIs and our structure and how successful the agents are, because the agents are not going to be static. They’re going to learn as well. So how do we moderate the learning? So it’s a new era in organizational design.
Sergio Maldonado
Yeah, exactly. Yes. So what I was thinking, being very practical now and before even getting into any sort of legal compliance, data protection, etc. So, for example, in the B2B space, there’ll be people using generative AI to create content, to build all of these papers that, you know, downloadable papers to generate leads. I guess for now, as human beings, as decision makers, we’ll still fall for some of these. So there’s a process that would be qualifying the leads. And just as per your description, you may have another process, which is going to take those leads that have been qualified and is going to do something to get them closer to the finish line. We used to buy workflow automation tools and add them to my years of enterprise sales. I would buy SalesLoft, for example, I think that was the name, to attach it to Salesforce and then it would automate an email and the second email is a cadence. We would call it a cadence of emails. So that’s gone. That’s deterministic workflow automation. So if you’re going to now apply this with the world of GenAI powered agents, do you think the world of probabilistic is ready to build a process that is reliable to the point that we entirely kill the SalesLofts?
Gam Dias
I think the simple answer is no, we’re not ready. But if we don’t start now, we will never be ready. So I think there is going to be a good six, 12, maybe 36 months of making some mistakes. We’re going to be implementing these things. And they’re going to do some non deterministic things that we didn’t expect. And we’re going to learn from that and then control some of these agents. I think, you know, if you start relying on an agent, if you start relying on an LLM’s output, and you’re going to notice that it’s going to start hallucinating every now and again, so if you check, you’ve to check it before it hallucinates, and you actually use that hallucination in something, you know, maybe you’re writing, having a white paper written, and it’s just telling blatant lies. And so I think the same is going to happen with these agents. They’re going to do some strange things. You know, I’m going to be seeing a customer like opening things and I’m going to miss, I’m going to mishandle something as a buying signal. And I’m going to send them out like a proposal form at that point. And then enough of those failures, we’re to go, okay, hang on a sec. This isn’t quite ready. So I think Salesforce, Salesforce loft has enough time. to go build their own. You know, there’ll be sales loft agents that live inside of the sales force marketplace.
Sergio Maldonado
Also, if it is more chaotic or more language dependent and more probabilistic, it’s in a way more human. But if we extend that, right? So now going deeper into this, into what people can do, right? And you wrote about this. So if I use Salesforce, what is the lowest hanging fruit for agent force now?
Gam Dias
Yes. So, I mean, actually I’ll tell you about the thing I wrote, but I’m going to tell you about the thing I haven’t written. You know, one of the reasons that I was kind of saying that there’s going to be that dip after the hype is that, agent force isn’t complete. It’s a, it’s very much a work in process progress and they had to launch it when they did and they had to hype it when they did because that was when the ability to create codeless agents was going to happen. They hit the timing bang on. They came out before OpenAI Operator. So it was a really good call to do that. They’re not going to do that by halves. They’re going to come out guns blazing, which they did at Dreamforce. And it’s fantastic. However, there’s lots of things that aren’t quite there in Edge in Force yet. Multi-agents, they’re not quite there. I can’t orchestrate multiple agents taking over things. It’s a little bit clunky here and there. And there’s obviously going to be a lag because I don’t think that there are that many Salesforce orgs that have all their flows correct. People are still looking at workflows and other automations. And so they’ve got to migrate those to the flow. Now, there’s a lag there.
But what is there and what is robust are probably the out of box agents. The BDR, the business development agent, the sales coach, and the customer service agent. They’re out of the box. They’re probably pretty close to deterministic in terms of an agent. There’s some set flows, set behaviors. They kind of work on the base implementation of sales cloud and service cloud. They’re relatively low risk and they can actually alleviate some of your kind of human resource bottlenecks. If you haven’t got a BDR like doing follow up and that’s a really big problem inside of Salesforce. Yeah, we’ve got all these BDRs, but you know, we haven’t got enough to follow up on every lead that we’re generating. So that kind of stuff, it’s very easy. It’s very simple. So my advice is that if you are in a Salesforce org and you know, your account manager or your rep is probably very enthusiastic that you start to build with agent force. So these are really great starters. I’d begin there. And as you start to see the limits of those agents, that’s when you can get a bit more sophisticated, but just get used to those. And even if it’s almost even if you don’t need it if you don’t think you need them yet, it’s probably worth doing a little bit of a cost to it. Because if you haven’t got a data cloud, then that’s a prerequisite outside of the dev environment.
Sergio Maldonado
And then we enter the final, the world that you were anticipating, which is the world of agent to agent, which seems to be natural, especially of course in B2B until the consumers are up to speed with something like this. And so just as a little teaser, because I know we have a separate episode about the model context protocol, the MCP. But how will this change the world and what is your vision for that part?
Gam Dias
Tough one. So my philosophy has been that agents are really good at some things. They’re really good at remembering an awful lot of information and just retrieving it very fast. And they’re actually, they’re very good at, what’s the word? Just kind of moving data between systems, just data transcription. I mean, if you’ve got a job today, that is, if your job involves moving data from one system to another, whether you’ve got a spreadsheet in the middle or you’re doing something, you need to think about what you’re going to do when you are going to be… How do you get to manage a team of agents? If you’re in that job today, you will be managing a team of agents doing it. And if you’re not, you will probably be out of a job and you need to retrain. But one of the things that agents don’t do very well is compassion. I think an agent can emulate empathy, but they can’t do compassion. Because I read somewhere that compassion is a genuine desire to help. And because an agent cannot feel, cannot reciprocate with pain, one day they might, but today I don’t see them doing it. They don’t have that genuine desire to help. So I think there’s going to be a role for humans. We need to work out where to deploy humans.
In the book, I laid out a set of scenarios, whether it’s a human talking to an agent, a human talking to an agent with a human behind it. And those scenarios are documented as to where you’d want an agent. Perhaps if you’re seeking relationship advice, then you might want to actually just talk to another human. If you’re calling emergency services, then you want a human at the front end with an agent behind them. You don’t want just a human. You need that agent to be optimizing whatever service you’re getting. But if I’m just rerouting an Amazon package, I don’t need a human. So let’s be aware of the context. We can’t just apply agents willy-nilly. We’ve got to think about that. And so I’m really hoping that people start to consider the user experience. So experience design is a really big piece of this. We’ve got to really think about it, we’ve got to do some good design thinking about our processes, because the processes today, particularly in Salesforce, are optimized and streamlined for how Salesforce wants the tables filled. And so we’ve got to think about a slightly different experience as to what the agent brings to the party, because one of the things you don’t want, you don’t want a human and then an agent and then another human and agent. You want to kind of bunch up the human activity into one contiguous conversation. Otherwise, the customer is going to end up talking to an agent, then talking to one human, talking to an agent, and then a different human. And that’s probably the most frustrating customer experience where you have to end up doing that.
Sergio Maldonado
Okay, so there’s going to be an element where a privacy program, a data protection program will be impacted by agents. There’ll be less humans in the loop. Perhaps we’re able to do things with data that is less identifiable. We can optimize processes. We can avoid human mistakes — In the end, social engineering is the primary source of data breaches. But I want to ask you something that really has huge impact, which is, going back to the economics, Salesforce, Adobe, even Oracle, they’re all talking about productivity gains from Gen.ai. And how is that going to play out? Are they going to be able to pass these savings to everyone? Are they going to be challenged by it? What’s your vision in terms of GenAI and the impact on them, on all of these suppliers?
Gam Dias
It’s a big question. I’m going to just throw it out there. I think it spells a sea change for SAS. All of the big players are claiming productivity gains, and it’s fantastic. And they’re claiming that their engineering teams are using generative AI to be much more productive when they code. When they test, when they deploy. It’s going to be fantastic. Now, are they going to pass these savings on to the end customers? I don’t think they’ve done that in the past. I don’t see them doing that yet. However, the effect of that, I think SaaS companies have grown with, as the water has risen, all the SaaS companies have grown. They came in with a promise.
We’re going to relieve you of your IT department of having to provision all the hardware and the middleware. We’re just going to give you the application, and you just pay by the use. And they’ve got big. so SAS contracts are now six, seven figures. So those are big numbers. And the thing about SAS is it’s not bespoke. It is a system that’s built for everybody, and you still need to configure it, and you still need to customize it.
And so there’s a pretty gigantic services model around it. And you know, the SI partners in the Salesforce ecosystem take the place of your IT department and they’re on tap. you know, it’s a better cost model for you. You know, it’s OPEX. And so it’s fine. But you know, when I’m now spending millions of dollars on customizations and I’m looking at this and it’s like, hang on a sec. If I stop and think about this. I can actually build this application myself. It’ll be custom built for me. It’ll have all the advantages, and I can maintain it in-house. So what I’m foreseeing is probably the rise. It’s not the reincarnation of IT departments. I think it’s a different thing. think large organizations that have typically been SaaS customers, they can actually create engineering teams of their own and actually build stuff themselves. So the irony is, you know, the productivity gains that are going to come for the SaaS vendors are going to be equally available to their customers. So maybe we need to rethink the SaaS model. You know, I think Reid Hoffman last week showed a video of him asking some GenAI development tools to replicate LinkedIn and he said they built it and it was very satisfactory.
Sergio Maldonado
Yeah, a Lovable.dev or a Bolt.new.
Gam Dias
But if he can do that, then we can all do it now. There’s some amazing applications showing up. I think we’re going to have a big app proliferation very soon. And it’s not the coding that’s going to be the advantage. It’s going to be the customer experience. It’s going to be the support. It’s going to be how they go to market. It’s going to be, we’re not going to compete on code anymore. That’s for certain.
Sergio Maldonado
Yeah, I’m sure we’ll have a chance to talk about this in After The Magic. I do think we’re far from getting a hundred percent of the code polished and ready. And we still create an optical illusion by using all of these tools but there’s still this last mile… but yes, with a much smaller team, I’m sure you can do a lot.
Is Agentic AI then the last moat for many of these platforms that can now be commoditized so easily?
Gam Dias
So I mean, you know, for the longest time, I’ve been saying, you know, AI is the last resort when you know, 2016 when I was when we were building an LP products, I just said AI is the last resort, if you can do it with stats, or just standard reporting and BI just do that. Because, you know, AI requires machine learning, requires engineers, requires training, requires loads of infrastructure. And now we’ve gone on to gen AI, if you think about going back to what I said about the processing costs.
The money that is now going in, I mean, think the LLM companies have taken $100 billion between them. And there’s no sign of that being recouped as revenue yet. We are all going to be hooked. And then we’re going to turn up the prices. And then we’re going to pay for this. And so I think SaaS companies still have a future.
But they’ve got to rethink how they do things. They’ve got to re kind of figure out their offering. Now, today there was a little question, you know, should Salesforce become, what is it? An ARM, an agent relationship management. And it’s like, well, you know, agent relationship management is actually now in three layers. There’s kind of the, you’ve got to, you’ve got to manage the deployment of agents. You’ve got to orchestrate their behavior, but then you’ve got to provide some governance. So the top of it.
So if you just think about those three layers in an agentic stack in a business, there’s whole new lines of business to be had. I think we’re going to, you know, the next five years is going to be very interesting. We’re going to be doing a lot of reinventing in the tech stack.
Sergio Maldonado
Very good, Gam, thanks again.
Gam Dias
You’re very welcome.