Real Agentic ROI: Turning Signals into Revenue — Fireside Chat with Rishi Mallik & Blake Williams

Julia Nimchinski:
Thank you so much. Rishi, thank you. Rishi Malik from Workato, welcome.

Rishi Mallik:
Hey, Julia, how are you?

Julia Nimchinski:
Super excited to feature you bringing on Blake Williams, the moderator of our next session. How have you been, Rishi?

Rishi Mallik:
I’ve been great, I’ve been great. The weather’s been nice and sunny here in San Francisco, so I can’t complain.

Julia Nimchinski:
Awesome. Can’t get… can’t wait to get into this session. Blake Williams, welcome. Take it away, the stage is yours.

Blake Williams:
Alright, thank you, Julia. Nice to meet you, Rishi. How are you?

Rishi Mallik:
Good, Blake, and yourself?

Blake Williams:
Good, good, good, good. Well, let’s get right into it. So, When you think about MCP, Model Context Protocol. What is it? Let’s start there. And then we’ll rapidly accelerate.

Rishi Mallik:
It’s a good question, and I’m sure, you know, I’m… like any model, I would have multiple answers, depending on when you ask me. But, MCP, the way that… I think the best way that I look at it is, you know, compared to APIs that we had in the past, MCP just lets us take automation to a whole different level. So it’s sort of like basic API calls, it was very much a point in time.
You could ask… a good example is, like, asking what is the weather? And it’ll tell you what is the weather. And you ask it again, it’ll tell you what it is. With MCP, you’re able to actually deliver some context.
So, whether that’s personal context, business context, along with the query, and while staying in tune with the previous context, and so then you can make… or the model can help make some more intelligent decisions. And so. That, with… combined with business context, inside the workplace, just takes automation to a whole different level.

Blake Williams:
Yeah, yeah, I can see that. And you guys have actually distinguished, two different types of MCP. connections, right? You have a basic and then an enterprise. Can you talk to me just briefly about the differences?

Rishi Mallik:
Yeah, I think, you know, when we were going down this journey as well around, you know, how can we become… an agentic organization. How can we become the most optimized by leveraging the tools and technologies that are out there?
One thing that we saw with leveraging MCP, first off, amazingly powerful protocol, but when we were trying to build out, you know, the semblance of agents, and giving our agents access to different MCPs, whether it’s for different tools.
What we were seeing, and some people, you know, that are watching this might have probably experienced this, which is… you give it the context, but it’s still stochastic. That question is probabilistic.
And when it comes to mission-critical business actions, we want to make sure that there’s some things where you want that sort of human, stochastic, probabilistic sort of response, and then for part of a workflow, you may want it to do the exact same thing every single time.
So I’ll give you an example, you know, like, when sending an email based off of some information inside of Salesforce, every company has a custom instance of Salesforce, and… you know, which field to use. If we just give our, you know, basic MCP access to all of these systems of record and ask it questions, it may… you know, we have 17 ARR fields in our Salesforce, for example.
How does it know which one to use, right? How does it know to check one system before the next? And sometimes it’ll get it right, and sometimes it won’t. So, you know, MCP very much gives that context protocol for us to interact with multiple systems.
But where we’ve seen that it’s been a little bit sort of problematic alone without Enterprise MCP is that you’re mashing, you know, in a question, you might have part of it that is stochastic and some that’s sort of deterministic.
And with Enterprise MCP, what we’ve found is that we’ve been able to, really hone in on that deterministic part, and give access to our agents, to be able to do, sort of, orchestration in conjunction with the stochastic. So, basically, in any sort of request.
there’s, you know, a strategic answer that you would want, sort of human intelligence, and then there’s something where you want it to be repeatable every single time, and Enterprise MCP helps us get us there.

Blake Williams:
Yeah, yeah, and I can imagine, you know, it’s… it’s both how do we shrink the pool that we’re ladling into for the agent, and then, just from a safety standpoint, and I don’t know if people are playing with, like, open claw, or open claw, or whatever it is. when you have multiple agents digging into that, then you absolutely want that refined scope.
Just a… if we can’t limit risk, limit impact, all of those things.

Rishi Mallik:
Exactly. We call it sort of, like, verified skills, almost. It’s, it’s almost like how the human brain works. Like, the human brain, when you’re picking up a glass of water, your mind knows, okay, I just need to make the call to this arm, to this finger, to go and do this, and you don’t want it to relearn that path every single time.
Now, when it needs to drink water, that should be a sort of stochastic intelligence to make that decision that I need to drink water. But the act of lifting up the cup It’s a repeatable thing, you want it to happen every single time, and if we can give with Enterprise MCP those verified skills.
To an agent that is brilliant and stochastic and can make those decisions of when to use those skills, that’s when we see the magic happen.

Blake Williams:
Yeah, it makes it completely reliable. And when you think about signals and interpreting those signals, and whether that’s just inside of the GTM or outside of the GTM, how are you thinking about leveraging that with your agents and MCP?

Rishi Mallik:
Yeah, it’s a great question. I think, you know, as Chief Growth Officer at Rocado, I’m very concerned about, or our team focuses a lot on signals for our customers and prospects, right? And we’ve sort of bucketed them in a lot of first-party signals, you know, somebody coming to our website, somebody dropping off, you know, a sign-up or a demo form.
These are all real high-impact signals That we can see first party. But then there’s third-party signals. Like, if one of our existing customers were to change a job, the new company that they’re going to, we’d probably want to have a discussion in terms of, like, is our product applicable to where they’re going? And that’s a third-party signal that’s very valuable to us.
And so, with all of these signals, today, our sales team just combs through them all, right? They spend hours, if not days, combing through. call transcripts, combing through product usage data, and then matching those signals with the right message to the right buyer. It’s important, but it’s a very arduous task.
And, you know, a lot of… using Enterprise MCP to help build some agents, it’s been very impactful to manage that.

Blake Williams:
I can imagine, I can imagine. That’s the hardest thing when I talk to sales leaders and their teams, they’re all customers of Common Room, and… you know, they’ve subscribed to, you know, the maximum payload of signals, but when it gets down to that individual person, people are like, what do I do with my hands? You know what I mean?
Or maybe they find one one path, and they’re able to signal stack, and then actually get to outcomes. But it’s very temporal, right? Because it’s not… people don’t change jobs all the time, and That kind of thing, so you’ll see these peaks and valleys.

Rishi Mallik:
Yeah, it’s the, you know, we always talk about internally, it’s that, it’s not that… it’s the diligence of being able to do it every single time accurately, and it’s hard. It’s hard for a human to manage all of those signals.
It’s not just the, getting the signals, but it’s sort of the pattern matching in terms of, hey, most successful customers will do… I mean, that’s… that’s the important thing, is that, you know, what we’ve seen for our successful customers, I mean, it’s on average of 7 to 12 different touches or engagements, in some way, across the entire user journey.
And so… there’s so many different types of signals that lead us, that there’s, hey, there’s potential, but when we start stacking those, that’s when we know that this is actually heading into a direction, a positive direction. But the mind can only hold so much data, right?
So for relying on are asking the sales rep to do that for every single one of their customers, inevitably, something’s going to drop, and that’s where leveraging these tools has become really helpful.

Blake Williams:
It is, it is. Now, when I think about the total opportunity, I think it… from NFX, James Currier was just on talking about these, you know, three… three-person… what did he call them? Decathletes, or something like that. But these three… three-person unicorns that are… that are now possible. And when I think about the signal surface area.
across an entire business, not just GTM, because that’s where, you know, people have kind of gotten that toehold in signals, but really, what we’re always trying to do is manage signal from workflow and outcomes, across the entire business. And a lot of those are managed by people today, and some of them are using Claude.
How does Wercado, I think it provides a genuine opportunity, because you guys you integrate with so much, right? And there’s just so much, and the more that you can integrate with, the more of that surface area signal you can kind of bring to the table. Do you see teams asking those questions yet? Are they already working down that path?

Rishi Mallik:
They are. I think they’re asking the right questions, Blake. I think that, they’re still trying to figure out how to get there, but what we’re seeing is that all of our customers are running into the same issues, and they have the same needs, which is… You know, they figured out where they could potentially leverage the power of a model or an LLM.
And that LLM may even have potential connectivity into these systems, and where they’re running into more of the issue is, one, like you mentioned before, the trust. Just because it has access to these systems doesn’t mean that it’s… Not even just trust in how it’s acting, but trust in terms of the data pulling. You know, it’s only as smart as what you feed it.
And so, you know, what we found is, something that I was mentioning earlier around these trusted workflows. It’s not just the ability to integrate, but the ability to. you know, like I said, I’ll go something basic as sending an email.
I need to find, the right first name field in Salesforce, I need to find, you know, if you have some if-then rules in terms of, I only want to send this email to somebody over a certain amount of ARR, you need to find the right ARR field. So it’s not the actual point-in-time act.
Those agents can do that really well, but it’s the… ability to… how does it have the context of what are these trusted workflows that I need to run? And if we can package that up into verified skills and give that access, that’s really where we’ve seen this sort of, kind of take off.

Blake Williams:
It’s beautiful. That’s beautiful. The, because I’ve, I’ve tried to… not tried, I’ve done it. And I stumbled along the way, right? I’m building my own agent platform, and I’ve used MCP protocols, and some of them are very rich, and some of them are, you know, skim the surface.

Rishi Mallik:
And…

Blake Williams:
hit stumbling blocks along the way, and obviously, I’m just using Cloud Code, like everybody else in the CLI. How does… how does that experience differ from doing it inside of Workato, or leveraging Workato to get… get that done.

Rishi Mallik:
Well, the cool thing is, Blake, is that I think… I think we’ve all kind of gotten into that muscle memory of, you know, do we leverage OpenAI, or do we leverage Claude for doing this, and using that almost as our front end, if you will. And so, one of the things that we’ve done internally for our team, when we looked at how we were going to roll out AI to the entire org.
One was… Simple… use cases that are AI for all. Somebody could actually go and ask a question about something with enterprise context within our company. Then the next level was getting some of our employees actually building, sort of, agents and workflows. The third were, like, departments building out, sort of, end-to-end department-type use cases. And then larger custom deployments with IT.
And for that first bucket around, you know, how do we get everyone sort of leveraging this, people were already used to using Claude and using ChatGPT, and so what we did for our internal instance is that we fed Claude access to a number of our enterprise MCP servers. So, an enterprise MCP for requesting PTO, or for… a really cool one is license optimization for IT.
So, you know, hey, you haven’t logged into Salesforce for 2 months, like, these are the types of reports, can I build these reports for you elsewhere? So on and so forth. And now, actually, we all… all of our employees now go into Claude every day, and we’ve just seen that Claude usage spike up, because Before, they could ask Claude simple questions that didn’t have as much enterprise context.
Now that all of our existing automations and workflows are sort of, reskinned as enterprise MCP servers that Claude has access to. That’s how we… that’s how we’ve actually been able to really amplify the impact of Cloud in our enterprise.

Blake Williams:
That’s huge, that’s huge.
You know, what’s interesting is, like, the… just the… the peaks and the valleys, the heat maps of where that… That spike is occurring around different departments, kind of… well, it lets you know where AI is being deployed, but I think about from a value capture standpoint, or for a friction recognition standpoint, you can kind of start to… it becomes a surrogate You know, not for observability, what are they doing?
But are we capturing productivity value, from these things, having access to this deeper, richer insight, just from when it is starting to occur, and how often it occurs, and what the outcomes are? From those workflows. Are you guys… are you guys getting into mapping any of that internally? Like, tracking it back to, you know, FP&A?

Rishi Mallik:
We do, we do. We’ve, you know, the beauty of MWCP is that we’re able to give our agents a lot of context, and so one of the… One of the things that we do when we build an agent is that we give it a job description, kind of, and it also has KPIs. So, for example, you know, pre-MCP, we had a lot of automations that were running on the Workato platform.
For example, for automating the outbound process. If there’s a certain signal, that comes in, like, for example, if an existing customer changes a job, we have a really nice, tried and tested message, that first message that goes out to that person. And so.
what we did is then we actually sort of tied that after post-MCP, we gave a job description to say, hey, you’re a demand gen manager, or an SDR manager, here’s your KPIs, you want these emails to hit a certain open rate, and you want these emails to have a certain reply rate.
And so, you know, sort of design a process to where, for different cohorts, you’re testing out a similar message, and then you’re tweaking along the way. And, Blake, it’s been, it’s been mind-blowing to see how powerful that’s been.
I mean, basically, you know, we’ve launched this thing now over the last 6 months, and we’ve been able to source 222, sales-qualified meetings, automatically from this sort of, like, signal tracking. But going back to your question, you know, so then we have… now it has, you know, real KPIs where we can track real impact.
And I think that being sort of KPI-driven when we’re thinking about these agents has really changed You know, before we felt like we were spinning our wheels, I’m sure a lot of people felt this way, and we weren’t sure if there’s actually any sort of ROI coming from it, but now, that is a primary source that goes… a source that goes down to the bottom line.

Blake Williams:
Is that… is that a… is that use case specifically one that you find customers coming for… coming to you and asking for as well?

Rishi Mallik:
We do, we do. That’s a really big one. A similar agent on the sales side is also a really big one. The license optimization agent is one that all CIOs are clamoring for, because, you know. Again, you can write a simple automation that says, you know, hey, Blake hasn’t logged into this app for 2 months, you need to check out if he’s gonna do this.
But, you know, there’s cases where it still makes sense for you to have that. And so, for you to be able to have that back-and-forth dialogue. And explain reasons why, why not, and it’s able to give you alternatives, and also, you know, look at all the contracts and let you know what the impact of your license is gonna be.
That’s… that’s been… so even that agent that is, you know, for IT has a KPI of sort of, like, cost savings and optimization of licenses, and so we have a dashboard of all these agents that are displaying where it’s at with these KPIs.

Blake Williams:
Very good, very good. Have, and this might be a curveball for you, but my background is largely in ecosystem activation, right? And the activity of partners going after revenue together, right, to help customers create value. And integrations, you know, everyone always says, or at least in the partner world, you say, you know, integrations don’t equal a partnership.
But that’s… that’s very rapidly changing, or at least I can see it going a different way, where, you know, I don’t necessarily need your front end, all I need to do is, you know, drop in my card into Stripe, add your integration, whether it’s through MCP or API, and all of a sudden, I can start measuring value. getting thrown off of that.
So, have you thought anything about, that second-party data? from partner ecosystems, and then being able to meter or track and measure the value. And that may be completely outside of the scope of, you know, and if so, just say so, but it’s just something that’s curious to me.

Rishi Mallik:
I mean, what I will say is that, partners have been really critical in part of this process. I mean, they’re… they’re… domain specificity has been so impactful.
We actually see that partners are taking the most advantage of a lot of these platforms, just because, you know, not only for what they can bring to the customer, but also for their internal tooling, in terms of how we bring that to market. But, Yeah, not 100% my area of expertise, Blake, but if you have… I don’t know if that answered your question, or if there’s something specific.

Blake Williams:
It’s comforting to know that they’re at least asking the question, so…

Rishi Mallik:
Thank you, thank you.

Blake Williams:
Where do you… where do you hope it goes, when we think about MCP and agents, and, like, just from your own curiosity, you know, what are the things that you think are… are out there, as far as a direction?

Rishi Mallik:
Yeah, you know, I think what was released with OpenClaw recently has been really exciting. And one thing that we’ve been toying around with is, how do we get these, you know, sort of, OpenClaw use cases working for you 24-7?
How can it be… For your work, that you go to work every single day, if there was almost, I mean, OpenCloud does great work, I don’t know if you’ve played around with it on your machine, but, you know, if we could abstract that into… Sort of a cloud-based worker that also had access to all of the tools that you have access to at work, and all the permissions that you have access to at work.
I mean, basically, everybody could have their own digital intern, or not even intern, I mean, these things are really smart. I would say that the seasoned. sort of team members that’s working for you sort of 24-7. And that’s been… that’s been really exciting.
You know, we’re, working on, you know, some tools around that for our internal team, and we’re getting customers asking the same sort of question, and so that’s… That’s where, in the short term at least, I see it going.

Blake Williams:
Yeah, yeah, I love that, because, just to follow on with that idea, I’ve been thinking about, you know, how do I get my agent to spawn a new agent with all the right policies, all the right, you know, tools, the skill set, or hatch the new skills embedded in there, and then, even if it’s not always used, put it back on the bench, and then spawn it when needed.

Rishi Mallik:
Mmm.

Blake Williams:
Just to accelerate workflows and things like that. But we have a question from the audience. They said, I’d love to know more about the data sources that you’re ingesting when you’re looking for signals. So, is it news feeds, RSS feeds, first-party data, first-party cookies, or intent providers, etc? What’s the scope?

Rishi Mallik:
Yeah, that’s a great question. So first, I’ll start with sort of first-party data. This is, you know, pretty straightforward, all product usage. Even what’s happening in recorded calls, so, you know, we leverage Gong, but if you had a call recorder, We’re also leveraging first-party data, like support requests, support tickets. That’s a treasure trove for data.
And then all of our, sort of, cookie data in terms of page visits and where they dropped off, any sort of ad click data.
So all of that, And again, when we talk about signal stacking, there’s… well, we found a lot of really interesting stuff where, again, our teams couldn’t even do this, manually, which would be combing through, call data as well as support ticket data, and sort of matching that up with what’s happening in product. Then, third-party data, we leverage a number of sources.
But, you know, we actually give the models, you know, whatever’s publicly available, they usually find it on its own. And so, it’s looking at filings, it’s looking at 10Ks, it’s looking at recordings for public companies, in terms of, you know, their performance reports.
So… It’s… it’s… it’s a lot, but, you know, the… we used to get inundated with this, but leveraging… and then we also do use intent data providers, and it’s… I wouldn’t say one source is a silver bullet, but it’s that combination across the board between first party. Usually what you see a happening third party, you’re able to confirm with first-party data.

Blake Williams:
Right, no, that makes sense. That makes sense. And I can’t wait till you start adding in second-party data into your vocab here, which is all partner data.

Rishi Mallik:
Totally.

Blake Williams:
And that’s what I’m excited about, is, that 1, 2, 3, all coming together. But I, you know, my company’s a company of one right now.

Rishi Mallik:
And…

Blake Williams:
the companies I’ve been going into for these private equity groups, and looking at… That… just that observability layer, if we’re gonna stay on signals. how many people struggle to figure out what the beer and diaper scenario, right? If you have this massive market basket of… of signals, what’s the ones that matter to you?
And so, how do you think about enabling either those MCPs or the agents, to test, and do you… do you create those tests and figure out?

Rishi Mallik:
It’s a great question, and we do.
So, going back to the whole job description thing, you know, when we, we actually, you know, give the agents, you know, for example, if it’s doing, I recently launched, my team has launched a, an Agenta campaign, and… The campaign, you know, actually, you know, we actually have… We’ve given it sort of the KPIs in terms of what it’s supposed to hit, in terms of open rates and reply rates.
So what it does is that when it launches that campaign, it does it in cohorts, and it’ll take… it’ll look at the first hundred responses, it’ll internalize those responses, and any positive responses That’s great. It kind of confirms that that was the right signal. Any negative responses? It then does pattern matching in terms of what companies are they from?
What are the industries that they’re in? What are the titles of those folks? Is there any pattern match that, hey, reaching out to a certain title doesn’t work, or reaching out to a certain industry right now, given what’s happening in the market? has not been as impactful.
When the tariffs were out, our agent came back and found that a lot of manufacturing companies had a little bit dip in their response rates. And so, you know, that’s really nice, sort of, impact. So, we’re launching… have it launching cohorts. It looks at, sort of, the response rate, and then it sort of scales from there and continuously iterates.
So, again, it basically acts as a digital worker that’s working for us 24-7, and that’s how we confirm, you know, sort of which signals are working for us.

Blake Williams:
Very good, very good. So the, you know, the market’s generated around signals, but the honey, I guess, is inside of the pattern recognition. And what comes off of all of that?

Rishi Mallik:
And also pattern recognition from our closed-won and closed-loss deals. So we know, you know, one thing that we’ve done recently is we’ve mapped all of the touches on average, for a closed-won deal for a certain segment and for a certain vertical, and And so now, you know, as customers go through this customer journey, or prospects go through this customer journey.
our agents are trying to increase the next touch that it knows is typically successful for that industry or that size company. And so, that’s also been really interesting, is looking at our first-party data for our existing customers. Taking that pattern matching and applying it to the prospect journey.

Blake Williams:
What was the time scale that you used for, pattern recognition? Did you just let it go back as far as it could, or did you confine it to, like, the most recent 6 months?

Rishi Mallik:
We… We… first, we started with everything, and it was… it was a pretty mixed bag, and what we did then is what we realized, to your point, is we shortened it a little bit. The year seemed like a sweet spot for us, because it was able to track, sort of, trends. You know, the reason people were buying back, you know, after COVID was very different than the reasons people are buying now.
And, so all of that data was being a little bit confounding, but… What we found is that there is a sweet spot with, like, not too much and not too little, right, with the models. So the year mark, actually, was a sweet spot.

Blake Williams:
Very good, very good. And if there’s one last thing you can leave everybody with, if they want to come seek you out or try out Workato, what should they do?

Rishi Mallik:
Workato.com or, LinkedIn slash Rishi Malik, or Rishi at workado as well. Don’t hesitate to reach out, would love to work with you all.

Blake Williams:
Very good. Very, very good. Alright, no other questions from the… The audience, and we finish on time. Michelle.

Julia Nimchinski:
Thank you so much, Rashi. Thank you, Blake. Before you leave, just curious, what’s in your Argentic OS? What tools are you using, Rishi and Blake?

Rishi Mallik:
Say it again, Shali, what’s…

Julia Nimchinski:
What tools are you using? Obviously, Workato.

Rishi Mallik:
So we, you know, we’re fortunate that, we can use our own tool for a lot of this stuff. So we’re using Workato. Huge fans of, obviously, you know, what OpenClaw’s doing and, and leveraging a lot of, sort of, the Cloud infrastructure. And, yeah, so we get to drink our own champagne in that regard.

Blake Williams:
Very good. I’m, I’m knee-deep in Cloud CLI. I’ve got cursor going, and linear, and a bunch of GitHub repos that I’m, testing out and launching.

Julia Nimchinski:
Awesome. Thanks again.

Blake Williams:
Very good. Thank you.

Rishi Mallik:
income.