Sponsored by: Retool | Responding to Signals That Drive High-Stakes Decisions with Databricks Genie

Summary

Retool and Databricks Genie work together via MCP to close the gap between answering data questions and taking action, enabling a Genie query about at-risk accounts to trigger Retool agent workflows that bring in stakeholders and execute responses across enterprise systems.
This video presents a live demo showing how conversational AI answers in Genie Spaces can drive automated follow-up steps—such as routing signals to the right people and triggering downstream actions—before revenue or operational targets are missed.
The architecture is governed, auditable, and secure-by-default with human-in-the-loop approval for high-stakes automated decisions, making the Retool-Genie integration suitable for enterprise production deployments.

Chapters

FAQs

How do Retool and Databricks Genie work together for enterprise action?

Retool agents connect to Databricks Genie Spaces via MCP, allowing an agent to query governed enterprise data through Genie's natural language interface and then trigger automated follow-up actions across other systems. This video demonstrates an agent-to-agent pattern where Genie handles the data question and Retool handles the workflow execution, closing the gap between insight and action.

What is the core problem with conversational AI that Retool solves?

The core problem is that getting an answer from a conversational AI system is only the first step—a response on a screen is not itself a decision or action. This video frames Retool's role as operationalizing those answers by connecting them to the workflows, approvals, and system integrations that need to happen for the business to actually respond to a signal.

How does the human-in-the-loop approval process work in the Retool-Genie integration?

When a Retool agent identifies a high-stakes action—such as sending customer outreach or updating a record—it can be configured to pause and surface the proposed action for human review before executing. This video shows this pattern as a key governance mechanism that keeps humans accountable for consequential decisions while still automating the detection and preparation steps.

Why does Retool use MCP to connect with Databricks Genie?

MCP (Model Context Protocol) provides a standardized interface for agents to communicate with tools and data sources, including Databricks Genie Spaces. Using MCP means Retool agents can query Genie on governed enterprise data without custom integrations, and the connection inherits Genie's existing security and governance model within the Databricks Data and AI platform.

Full transcript

[00:08] All right. Well, thank you all for joining. It's just about to start in the next 5 seconds here. My name is Tom Kinuoka. I'm the solutions engineering leader. So, in 3, 2, 1. Perfect. Okay. Everyone, thank you
[00:21] for joining today. My name is Tom Kinuoka. I'm the solutions engineering leader here at Retool. And today I'm going to be talking to you about responding to signals that drives high-stakes decisions with Databricks Genie.
[00:33] First of all, quick poll from the crowd, who has used Databricks Genie? Who's heard of it? Okay, you all kind of Okay, you're all aware of it. Perfect. So now, it's the end of the day. It's been a long day. Thank you for coming through
[00:46] with this session. So, I will keep it a little bit interactive. I'm actually going to do a demo with you guys. It's not going to be a pre-recorded demo. It'll be a live demo. So, I'm a bit playing with fire, conference Wi-Fi, lots of systems, but we're going to see
[00:58] what happens. So, with that, let's actually get into the presentation. First off, you know, who is Retool? What is Retool? What do we do? We make AI development fast for builders
[01:11] and safe for business. Right now, it's very quick, very easy to develop with AI, but how do you actually secure, govern, and deploy to production quickly? So, together with Databricks, we're actually making vibe coding safe
[01:24] for everyone. Now, conversational AI has been solved, but what is changing? An answer on that An answer on a screen
[01:37] simply isn't a decision anymore, right? It is and it isn't an action. So, today, we can go into Databricks Genie, we can go into our agentic chat assistants of the day, and we can actually get answers
[01:50] about our data. We can chat with our data, but that's only the first step. It's very critical and key to actually be able to action on it. And really that is really the next step, the next generation, next evolution of where all
[02:02] of these systems are going, including Retool and Databricks Sheena. So right here we have an example, you know, which enterprise accounts are at churn risk for the week, right? Great, you got the info. You're maybe one step
[02:14] out of the next 10, 12, 11 further along the way. But what can you actually do? How can you architect this? How can you use Retool and Databricks to maybe perform some actions, really get you
[02:26] those next three, four, five steps along the way? Uh apologies, the clicker has having a little trouble here.
[02:39] So that that the challenge, right? The challenge that we're actually doing it seeing is the gap between knowing and doing, right? Things like missed targets come up, operational disruption, and compliance exposure, right? The challenge here is
[02:52] time, time to action. And in a world right now where we we all as operators, as engineers, as analysts have to do more with less, it is critical to use tools and systems and platforms that
[03:05] give you that leverage. So in churn, in the example that we're going to be showing, the cost is time. Why acting early is the whole game, right? So we have
[03:17] recoverable account value after a churn signal graphed here, right? Day zero to three window where outreach still changes the outcome, right? You can see massive there's a 60% or higher outcome if you actually reach out and action
[03:30] upon, let's say in this case a churn, uh within 30 days. By decision 30, the decision uh by day 30 the decision to leave has usually already been made. So what what can you do? Like actually like making this real for your world,
[03:43] your uh your day-to-day. As you're actually building with Databricks, as you're surfacing data, as you're working with teams at your organizations to bring that data life, allowing them to make decisions, what are other systems
[03:57] and tools within the Data Bricks ecosystem and within Retool to help even move it quicker along the way so you can actually reach out in this case to customers within the first 3 days.
[04:10] So, what happens today? This is the status quo. Someone spots an issue, right? If they happen to be looking at Genie or really any other agentic system about that day. They paste it into Slack, Teams, what
[04:23] whatever DM platform you're using, and then you hunt for an owner. Who's actually responsible for this? Who can action on this? Uh who can actually help me out here? So, it's always this kind of context switching discontinuous
[04:35] process. Now, why uh why agents and why not just an alert? A static alert, like you see here on the left-hand side, fires when a threshold uh trips, lands in a channel,
[04:47] has no view of onto what to do next. But, a Retool agent, what it can do, it can actually read the Genie result in context. It'll decide which follow-up fits the best, and it'll convene the
[04:59] right people on its own. And this is what we're actually going to be demoing here very, very shortly. And from there, you can actually take action because the alert is just step one.
[05:13] So, how do we do this? How was it accomplished, right? With Retool agents and Data Bricks Genie spaces, we actually uh we can actually uh exchange and talk to them through MCP or REST, right? MCP is
[05:25] the preferred protocol now for effectively agent-to-agent communication, where you can actually get uh discovery and schema of tools very, very quickly. And it actually allows for natural language queries. It's much It's much
[05:38] better built for natural language queries. So, a A from one of our recent customers, right now, Anthony Perez from MIQ. Agent-to-agent communications between
[05:51] Data Bricks Genie spaces and Retool means that the right people are responding to the right signals sooner. So, with that, we're going to get into the demo, so you can kind of see what it's all about. And let's actually go right go right
[06:03] now. So, I am going to go ahead and go into the demo. Let me
[06:15] just copy that right now. Okay. So, we're going to start a new Genie chat, and maybe I'll zoom in over here. There we go. So, we're going to ask it a question. So, this is an interface you're probably
[06:27] very familiar with right now. And we're going to ask it, uh what is the distribution of order amounts, min, max, average, median in the churn orders table, right? Very common query that we'll ask uh ask our our Genie.
[06:40] After a few seconds, it'll come back uh with a response. Great. We got the alert. We have the information. But, what if we could actually action on that and uh chain that with Retool agents, and
[06:53] actually uh perform uh gather the correct stakeholders in a room, schedule a meeting, invite to actually discuss and talk about it. So, after a few seconds, it's going to come back with a list of results. And this is where the
[07:06] conference Wi-Fi will definitely be uh very very key here. But, looks like we're good. It's thinking. Let me create the visualization and distribution order amounts. And it'll take a few seconds.
[07:24] All right. Perfect. So, we have our table, our minimum, maximum, average, median, right? Uh we got the info in Genie. Great. Now, let's take this one notch further, one step further. What if we got a Retool agent to actually ask Genie through MCP that
[07:38] same question? So, now we're kind of building the layers and how we can potentially automate this end-to-end. So, just as an example, I will refresh this page. She is like a playback of a Retool agent reaching out
[07:50] to Databricks Genie through MCP. Same question, what is the distribution of order amounts min, max, medium in the churn orders table? We can see we're getting kind of a status update as to what's going on. It's doing the poll.
[08:03] It's through MCP it's asking that same question and arriving at the same result. So, Genie feeds the Retool agent these values uh minimum, maximum, median and the distribution suggests that the most order amounts are clustered below
[08:15] the average with a few higher value raising above average to mean. Okay, so we now have a Retool agent talking to a Genie agent. We have the foundation, right? When the foundation
[08:27] is there, we have multiple systems talking to each other agentically. Now, we can get creative as to what actions we want to take with this, right? So, let's take it a little step further. Uh let's actually get even more
[08:40] creative with it. I created a an application here, basically a churn application. Let's just actually load it right up right now.
[08:53] Wi-Fi has been pretty good, so I'll I'll take it. And while the application loads, what we're actually going to do uh let's see here. There we go. So, we have the churn orders app, right? Where
[09:06] we basically took that data that was in um effectively in the churn orders table, we put it into a React app with with our new with our new Retool builder. And so, we have that data over here. We can chat with our data. We see
[09:18] that same question that was here, the order distribution amount. Great. Now, here's what I'm going to do next. Normally, this process would be automated, right? Going on in the background, but instead I'm inserting
[09:30] myself as a human in the loop and I'm going to notify with uh notify uh effectively a Slack channel and notify an agent to start performing actions on it. And this agent I'm calling is called Echo. So, I'm going to notify Echo over
[09:42] here to perform an action. All right. So, we've notified. Uh message has been sent. Let's go ahead and check Slack. So, now I'm going to go into the ask uh let's see Echo
[09:56] one sec Echo app. Okay. All right. So, now we see that Echo, please assemble the churn orders team. Uh we need eyes on the churn orders distribution. Let's coordinate next
[10:08] steps. Okay. So, now we're it's actually finding the answers. What I'll actually do, I'll give you like a a peek under the covers of what this agent is actually doing. So, let's actually go ahead into that agent right now.
[10:21] Uh there we go. So, if we go here, let me refresh the page. All right. So, now we see uh we we had the uh we had the churn orders alert
[10:34] come in. It's doing the resource call and that's now scheduling a meeting, right? Uh the sched- the meeting will now be scheduled between myself and a fellow colleague named Alex Jones. So, let's go back to Slack. Let's see if it
[10:46] actually finished. All right. We finished. So, team assembled, new meeting Thursday, June 19th, 10:00 a.m. uh Pacific time. All right. Let's go ahead and check my calendar. And there we go. 10:00 a.m. Pacific
[10:58] time, churn orders review with myself and Alex Jones. Alex is Jones is sitting right at the back over there. He has no idea what's going on, but he's about to get on a meeting uh with me to discuss churn. So, with that, you saw let's now go back
[11:11] to the presentation. And uh thank you again for bearing with me there. It's the end of the day, so I wanted to make sure you got a live demo
[11:23] instead of a pre-recorded one. But with that, what just happened, right? You were able to detect churn risk using Databricks Genie. You were able to ask it through natural language,
[11:35] govern and get a governed query through MCP. You're able to interpret that and also connect it to a Retool agent, so you can actually build that foundation of an agentic system
[11:47] exchanging context and data about your organization. From there, you're able to decide what follow-ups fit that signal. In our example, we actually just used a button to actually trigger a Slack channel
[11:59] which an agent resides in and that agent was able to perform actions like schedule a meeting for the next available date between two parties, right? So, there you're able to act. Now, if this was automatically done in
[12:11] the back end, you can imagine an agent scanning systems, reviewing certain thresholds, uh then performing actions like getting stakeholders together, writing new requirements docs,
[12:23] scheduling meetings, scheduling Zoom calls, scheduling Teams calls. All of a sudden, you are now no longer performing those tasks. An agent is performing those tasks on your behalf and you are
[12:35] simply orchestrating. So, for those in the crowd today, you're actually in a very, very, very powerful and highly leveraged position because you're interacting with Databricks and it's a key source of truth platform in this
[12:47] entire AI ecosystem and in conjunction with platforms like ours at Retool, you can really, really take action very quickly against your data for those insights that you bring up.
[13:00] So, just in kind of quick summary, with our platform, you can connect either through MCP or rest. Those are the two common protocols. Depending on the use case, MCP could be better, rest an API call could be better depending on how
[13:14] programmatic that actual call is. And this was a quick little review of the actual agent calling Genie over MCP.
[13:26] It would query the space, pull the response, and see the churn orders. So, out of curiosity, poll in the room, uh who's responsible for maybe Are Are you responsible for getting data to your
[13:38] adjacent teams? Are you responsible for building apps? We'll have to know like just hands up to see like who's actually working with teams to actually get some of these insights, get some of this data for your teams. All right, a few hands going up. Great. I really wish I could
[13:51] ask you questions right now, but you're on the headphones, so I'll have to ask for hand raises as we go along the way.
[14:03] So, some of the key takeaways is you want to respi- uh respond to high-stakes signals with agents in one governed workspace. It has to be secured by default. So, in the beginning of our presentation, mentioned that effectively code is
[14:16] cheap, that incremental cost, the marginal cost is zero. So, what really the new bottleneck is how do we make sure that this is governed, how do we make sure it's auditable, and how do we actually have the option to still have a
[14:29] human in the actual loop, right? It can be completely automated, but we can still be a part of the process to help orchestrate, approve, and audit, and govern all of these uh processes.
[14:41] So, in our example, when I said, you know, send to the Echo Slack space, you're able to see that I was a human in the loop making sure that those actions are actually, let's say, approved, but also notified to me as they go along.
[15:02] So, how are you operationalizing AI? Oh, Paul, this version did not actually have the QR code. Uh so, I'll That's That's very strange. Uh apologies there, but uh ultimately, uh if you If you like, if you can, we're still at the Retool booth by zone two at the back over there, uh
[15:16] just a couple steps away. If you'd like to see this more in action, drop by our drop by our booth right now right after. We're still here for another hour or so. And we can go into more deeper dive demos, talk around data sources that you
[15:28] have in addition to Data Bricks, and how you can build applications, uh automations, and agents on top of your data to actually surface those uh insights, but perform actions and tasks against them as well at the same time.
[15:42] So, with that, thank you very much for all your time today. I know it was a long day. I really appreciate it. I'm still here for a couple more minutes if you have any other uh detailed questions if you want to come up on any any of those follow-ups. But, with that, please uh
[15:55] enjoy the rest of your day here, and stop by the Retool booth uh right behind you. Thank you so much. >>

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.