Michael Driscoll would argue that he is like many technology entrepreneurs. The co-founder and CEO of agentic business intelligence (BI) tools provider Rill Data says he has spent his career obsessed with data and analytics.
After receiving a PhD in computational biology at Boston University in 2008, Driscoll moved to Silicon Valley, pivoted from biology and chose ad technology for his computational pursuits. As he explained to tipsheet yesterday, “that’s where the data was at and it was the early days of programmatic.”
In fact, the seeds of Rill Data were created in his first programmatic startup, Metamarkets, which he co-founded with David Soloff. The company made sense of the massive volumes of data pouring from programmatic ad auctions, including the data firehoses of its first two customers: ad exchange OpenX and mobile ad exchange MoPub.
“We essentially built the leading analytics platform for programmatic media platforms. We worked with almost all the major exchanges, SSPs, and over time, we started to bring in DSPs. And then the dream I had was, ‘Start in digital media and then move beyond it,’” according to Driscoll.
But one thing led to another in a rapidly growing programmatic space and Metamarkets exited to Snap in November 2017.
And that’s where Rill Data and its agentic analytics story began.
Mr. Driscoll spoke to tipsheet about Rill Data yesterday, including:
- Rill Data’s origin story at Snap
- What Rill can do for advertising and marketing
- The transition to solutions for humans and agents
- Plugging into large language models (LLMs).
- Observability’s role in advertising today
- What is Conversational BI?
- MCP for Conversational BI
- Are APIs going away?
- The reasoning behind ‘local-first’ analytics
- AI and ‘local versus cloud’
- The state of mediation in today’s ad ecosystem
- Agent-to-agent future at scale — when?
Scroll down for the interview, which has been lightly edited for clarity.
TIPSHEET: How did the idea come about for Rill Data?
MICHAEL DRISCOLL: The origin story of Rill Data occurred while I was at Snap.
We saw an opportunity which in some ways was a continuation of the mission of Metamarkets — accelerating the path from data to decision for every organization. That means going from data to insights to decisions — this is the essence of business intelligence (BI) and what people have been trying to do for decades.
The world of programmatic advertising was very amenable to bringing technology to bear on some of these real-time, high-frequency, low-latency decision opportunities — micro-transaction decisioning.
That was exciting to us. So, while at Snap, with the same technology that we were using to provide observability into Snap’s nascent ads platform, we also rolled that out to other parts of Snap. Essentially we used the same platform we built for Metamarkets: an in-memory database that was scalable and could handle billions of events a day.
We started bringing observability into Snap’s app monitoring — at that time, Snap was having trouble releasing an Android app and so the ability to get observability and crash analytics for Snap’s native Android application was an important mandate for them.
We watched our tech stack replace a much more expensive tech stack that Snap had built previously. And that’s when I thought, “Maybe there’s a business here that can serve more than just programmatic advertising platforms.” I had always talked about ad technology as being the tip of the spear for technology innovation.
With Snap’s blessing, we spun the intellectual property out of Snap and, ultimately, that became the basis for Rill Data.
We started in February 2020, right before COVID hit, with a group of engineers that I brought with me from Snap — the same technology team from the original Metamarkets. We spent the next few years rebuilding that stack from the ground up.
Initially, we were focused at the database level where we spent a lot of effort modernizing and building a real-time analytics database that could run in the cloud. We ultimately realized that there were some others that were doing that — and, arguably, doing an even better job at that.
Then a couple of years ago, we realized that we didn’t need to build the entire stack ourselves. We started to focus entirely on the business intelligence application layer which would be powered by this emerging class of real-time databases. DuckDB is one, ClickHouse is another.
And at one point early on, one of our former Metamarkets customers came and said, “We’ll pay you for the old Metamarkets dashboard. We miss it.” They were building database infrastructure to support other BI tools at the time.
It was then that I remember going to the team and saying, “The market has spoken.” What it wanted was a unique, exploratory, interactive, immersive business intelligence tool that was loved by thousands of folks in the programmatic space.
And for me, that’s when the light bulb went off: We should give that to every vertical. So that’s when we started Rill Cloud’s business intelligence tool.
Today, we’re approaching 100 enterprise customers and growing more than 100% year-over-year in terms of “logo growth” and revenue growth on Rill Cloud.
We’ve raised $15 million from institutional venture capitalists such as Bloomberg Beta, Sierra Ventures, True Ventures. We also have a number of angel investors who are founders including MoPub co-founder Jim Payne (now CEO and co-founder of CloudX).
So, it’s been several years that we’ve been focused on the product and building a business intelligence tool. There’s a lot to it and I would say this: This is the year where we finally feel like we’ve got a full suite of features that are at par with some of the other tools out there. I would say we’ve had a little bit of a “coming out” party commercially with the hire of a CRO, CMO and a couple of sellers to really drive what we see as clear product market fit.
Unquestionably, the biggest accelerant for us is our core belief that AI is reinventing every category of software. Business intelligence will not disappear as a category. It will be reinvented for the age of agents.
Analytics is one of the top five most common use cases for large language models. We think that agents will be the dominant consumer of Rill Data’s business intelligence product.
Today, we already see more than 50% of projects built by agents, not humans. So, that’s been the catalyst of growth for us — agentic analytics and the explosion of analysis that we expect to see as agents drive business intelligence.
So, what can Rill Data do today for the advertising and marketing ecosystem specifically?
At a high level, Rill Data is the fastest way to go from data to insight. The main payoff for our customers is that they can talk to their data using AI agents and get trusted, verifiable business insights in minutes.
Clients might ask, “Why is this campaign underperforming?” or if any particular campaign metric has moved, “Why?”
Historically, we have heard that the difference between using Rill and other business intelligence tools can be hours. Rill can get that answer in seconds and other tools might take hours to get to that root cause analysis of why a campaign is underdelivering or underperforming.
In 2022, when Rill Data raised $12 million to make business intelligence dashboards, there was no talk of agents back then. And now Rill is addressing “Conversational BI”. That’s been quite a change, no?
There are really two major stakeholders for business intelligence tools: the ones on traditional data teams — analysts — who build reports and dashboards; and then there are the consumers of those reports and dashboards.
For so long, the analysts and developers who were building these reports were a massive bottleneck in the insight supply chain because it takes time to pull data and do analysis.
Today, the first place where agents are extremely powerful is in developing dashboards much faster than any human analyst could. This is sometimes referred to as agentic authoring.
We’re seeing Databricks coming out with announcements about their AI BI offering. They use something called Genie Code to provide this agentic authoring path. And so on that side, agents play an incredible accelerator.
For example, one of Rill’s clients is Creator IQ, an influencer marketing platform, which has hundreds of unique, partner-facing dashboards that they’ve developed for their partners and which they embed in their product. Previously, using a human to manage those at scale was incredibly time-consuming. Now, they’re using agents to both create and edit hundreds of these dashboards for the provision of insights to their partners. Again, previously, that activity would have required an army of humans.
The cost of creating analytics is dramatically reduced by developer agents.
On the other side of the equation, we have analyst consumption. This is the biggest group of consumers of Rill’s Conversational BI, and they are often sales executives.
For example, in advance of a call with the sales team, the CEO of an SSP wants to know what the top sources of inventory in APAC are? And what are the trends in the last two weeks? Those questions from the consumption side can now be mediated by an agent. They don’t need to click around a dashboard or know how to use a pivot table in Excel. A C-level executive can prompt a question and — instead of that question getting picked up by an analyst who provides a report 24 hours later — a report comes by email from agents able to answer questions in seconds through our conversational chat interface.
Does Rill work with LLMs? Or do clients bring their own access?
Customers can bring their access to LLMs. We can plug into Claude and GPT-4.
You’ve discussed with me previously how programmatic and observability are intersecting today. How do you see observability’s role evolving?
For context, observability is a term that emerged in the application monitoring space within IT. Companies like Datadog, New Relic and Elastic have built multi-billion dollar businesses on the promise that they would help engineering teams monitor their critical infrastructure and ensure that their services were running correctly.
In the digital economy we live in today, there are these always-on platforms like DoorDash and Uber. Advertising platforms are no different. All these services are powered by servers in the cloud that need to be observed to prevent downtime. So that’s where the observability concept started — looking at the vital signs of thousands and thousands of machines in the cloud.
This required a certain kind of technology to make sense of all that data and make it useful to engineering teams in order to be successful. Sometimes they’re called observability engineers — very important roles in places like Google, Meta and Apple.
I think what we’ve seen is that, again, as every vertical across industries has digitally transformed, there is a lot of telemetry coming off not just servers but aspects of the business.
For example, Stripe has telemetry every time someone completes a checkout in e-commerce. Lyft and Uber have telemetry on every pickup and drop-off point of every one of their vehicles. Media platforms have telemetry on every device that’s browsing a particular page of content. So, the same sort of technologies and approaches that were used to monitor servers are monitoring other business processes.
Some might call this business observability. As it turns out, businesses have gotten more real-time and become always-on platforms and services. For many digital platforms, observability has become mission critical for their business processes.
Historically, to achieve the kind of level of observability, businesses have stitched together many different pieces of technology: software, services, apps and data for the database. And then you turn all of the event logs into aggregations and meaningful metrics. Ultimately, you’ve got to create a presentation layer for those metrics that can be visualized.
Rill Data has tried to radically simplify that stack by having clients plug Rill into just the data. We take everything from that real-time database all the way through to insights exploration. We make that path extremely simple, easy to build and intuitive to consume.
For Rill, it’s about simplifying that process of delivering an observability solution into an organization.
Rill has recently introduced an MCP server for AI agents. You positioned it as supporting “Conversational BI.” Before we get into MCP, what is “conversational BI?”
So it’s not that different than having a conversation with a very smart analyst — that’s the essence of Conversational BI.
I think we’ve all recognized that we have a new modality for interacting with machines, which has emerged since ChatGPT, that is natural-language-based. Previously, software was not very good at consuming natural language.
If you wanted to ask an analytics question, the closest you could get to using natural language was sending a Slack message to a member of the data team.
Conversational BI is really about replacing that data team that you send a Slack message to with an agent on the other side of that.
What made you introduce an MCP server to support the new modality, Conversational BI?
Traditionally, if you were to plug in something like Anthropic’s Claude or ChatGPT, they would have to understand our APIs. They’d have to reverse engineer those APIs and actually write computer code to say, “What’s the eCPM of our APAC publishers last week?”
What the MCP does is it allows Claude and ChatGPT to speak to services themselves. It’s agent-to-agent communication using natural language. That’s what Rill Data’s MCP enables — effectively, what all MCPs enable — agent-to-agent communication using natural language.
Does this mean APIs are going away?
I think it’s an open question right now whether MCPs are the ultimate way in which agent-to-agent communication will occur. There are some advantages to it, of course — natural language is the universal interface of the world. But any agent-to-agent communication ultimately queries APIs. So if you look at Google’s approach as of late, I actually think agentic workflows will increase — not decrease — the surface area of APIs for businesses because whether it’s Rill Data’s, OpenAI’s or Claude’s agents, some agent needs to turn a natural language query into code that actually runs against a service.
You can never sidestep an API or a CLI (command line interface) — there’s always going to be code.
If anything, MCPs will drive up demand for APIs rather than removing the need for them. They will also drive up the requirements for APIs to be better documented, more reliable and, frankly, be more robust.
Given your business intelligence point-of-view, what are your thoughts regarding the future of the “black box”?
There’s a nice synergy between observability and transparency. And I think that’s one of the reasons the media and advertising ecosystem has lacked transparency. It’s very hard to provide transparency in the programmatic ecosystem, in particular.
There is just a tremendous amount of data flowing through these supply chains. And a lot of the machine learning approaches to media buying have been put into “black boxes” — in part because observability is hard. And so, it’s hard enough just to get basic reporting about a campaign’s performance, never mind extremely granular reporting into decisioning strategies at the auction level for billions and billions of auctions and ad buys.
Since agents are indefatigable and never sleep, I think agentic analytics does represent a potential breakthrough in transparency for these black boxes. I think it becomes possible to know what’s going on. We do have the telemetry, i.e. the data, and so I think we’ll need to trust a lot of the agentic decisioning that’s going on. That’s obviously driving a lot of value for buyers and sellers of advertising.
But I think we also need to verify it. And I think that Rill Data as a tool represents an opportunity to break open these black boxes and provide verifiability of what’s going on — that can be scary to some. We’ve always found that our best customers are the ones who really do favor transparency and try to win in an honest way. They’re not afraid of transparency.
Where does the lever of “compute,” or computing power, fit in this world of transparency coming to black boxes?
Since I started Metamarkets, we’ve had, let’s say, six cycles of Moore’s Law. We continue to see a massive buildout of data center capacity globally and, certainly, in the U.S. Having the level of observability that businesses desire does require significant amounts of compute. I’ve heard estimates that observability workloads today represent as much as 15-20% of all infrastructure budgets for organizations.
So you’re right, it does require massive amounts of compute. But I think it’s the market that defines this investment meaning… folks would not spend so much on observability if they didn’t think there was a payoff for it.
Agentic analytics raises the value of observability. For every unit of data that we capture, or every piece of signal, agents are now able to extract more insight and value from those events, so I think we are going to see the share of wallet spent on observability increase in the coming years along with the amount spent on compute.
There’s a case study that Rill published on its company blog recently titled, “Building an Agent-Friendly, Local-First Analytics Stack with MotherDuck and Rill.” Why is “local-first” important?
One of the most fascinating developments in the last couple of months in the world of AI has been this tool called OpenClaw. Originally, it was called Claudebot. They renamed it OpenClaw due to copyright infringement issues.
Meanwhile, agents need to operate inside of “sandboxes” or contained environments. And so, there’s a whole breed of developer tools out there today which are local-first developer tools and software.
The idea is that you have a lot of compute sitting on your MacBook or Mac Mini. The ability for an agent to build software in a local environment in some ways mirrors the way that human developers work in a local environment. This means that iteration loops can be very tight and you’re not spinning up services in the cloud. You don’t have to share your data with six or seven SaaS vendors. You can actually build some very powerful, intelligent applications that can run totally within the sandboxed environment of your local machine.
It has implications for privacy, which is something people love too. The security perimeter is the laptop that the agent runs in. It has implications for speed of development because, again, everything is running “close to the metal.” Those are the two biggest payoffs.
OpenClaw is an agent that runs on a Mac Mini — entirely contained — and it can analyze data, browse web pages and take actions. That gives you security control in addition to speed of development. OpenClaw has just exploded as a tool that a lot of developers are using to drive agentic workflows and to experiment with agentic workflows in a safe way.
For us, the “local-first” development is extremely important because it is consistent with the way that software developers build software today. One of the most important foundations for Rill Data is making it easy for developers to build dashboards and reporting. Rill has taken this view — which others have also taken — that software applications can be defined in code and so can business intelligence.
What that means is if you want to build a real dashboard, you don’t need to log into Rill and do a thousand clicks to set up your dashboard. You can actually open up Cursor or Copilot or another coding agent and write code or, more likely, prompt your way through an agent that writes the code that will define the look and feel of the Rill dashboard.
A code-first approach lends itself to a local-first approach. Because most developers in the world develop locally, they actually have their development environment running on their Mac and so we’ve built an IDE which runs locally and it allows someone to build, test and preview a business intelligence application right from their machine.
It’s about the speed of development — Rill enables it with a local-first code-first approach to business intelligence.
Is AI taking us to a more local experience versus using the cloud, for example?
In the history of technology there’s always been this pendulum which swings from server-side to client-side applications. Today, no one is running Claude’s foundation model locally. Same goes for OpenAI’s GPT-4 locally. They are communicating with cloud servers and sending data there. That intelligence is not local. However, many observers would predict that some of these foundation models will get quite small. You could imagine that when you ask Siri a question, it would make a lot more sense for Siri to be able to process and answer that question on the silicon on your iPhone without having to make a round trip to Apple’s servers in the cloud.
So, there is potential that some models will run in a local environment entirely. But, today, AI is still very much a cloud service that every application interacts with.
How is mediation evolving in today’s ad ecosystem?
The value of business observability is only possible if there are decisions which observability can improve. When we look at any business process, we have to find where the “low-hanging fruit” of decision points are.
In the digital media and advertising markets, there’s a whole slew of decisions that happen at the auction event. Pricing decisions, enrichment decisions, decisions that happen inside of the auction mechanism for programmatic advertising…. One of the key questions for the media supply chain is: where does a publisher send an impression? They can’t send it to everyone, right?
So that’s another place where observability can be extremely valuable — helping guide app publishers and traditional publishers with their decision-making around what will be the most lucrative channel to deliver their inventory. That decision needs observability — and those decisions are best updated continuously, rather than once a day or once a quarter. Publishers’ mediation decisions should be updated in real-time and continuously fed by observability metrics.
I would point to CloudX, which is proposing to build a sell-side platform (SSP) defined entirely in code. It’s an SSP for which an agent — not a human — can tune the parameters.
SSPs, like SDKs, have a lot of complexity. They need two pieces: agentic observability, which is fast, real-time and continuous; and they need agentic activation or decisioning.
It’s not enough to have agents on one side and humans on the other to close the loop. You want to have agents observing a system and then you want to have agents actually acting on that system.
(CloudX is a Rill Data customer and CloudX co-founder Jim Payne, also a co-founder of MoPub, is an angel investor in Rill).
The dream of an agentic economy is agents talking to agents and the massive shift that occurs as we remove the labor friction of humans who are fallible and slow consumers of information.
The more human involvement we can remove from these business processes while putting in agents that are observable and verifiable, the more value we unlock for humans that are participants in that digital economy.
How far away are we from that agent-to-agent future at scale?
I don’t think we’re that far. I mean… it’s already happening. At Rill, the number one question I get from our leading customers is, “… we don’t just want your Rill dashboards. We don’t even want your conversational analytics for our executives. We want our agents to talk to your agents. How can we set that up?”
The nature of the digital media ecosystem is that it is digital, of course. Change can happen fast.