Despite enterprises' investment in data platforms, visualizations and analytics, the process of understanding all those signals and moving to a business decision remains challenging, according to Shayde Christian, chief data and analytics officer at Cloudera.
AI has the potential to change that by providing business teams with the visibility and context necessary to act: decision intelligence. But moving to that decision intelligence requires governance and security controls to ensure trusted and accountable action.
Christian spoke with TechTarget and answered questions about how enterprises can effectively progress from the data analytics era, defined by dashboards, to a future of decision intelligence that shortens time to insight and enables at-the-moment decision-making.
Editor's note: The following interview was edited for length and clarity.
How is decision intelligence different from the "data-driven decision-making" language companies have used for years?
Shayde Christian: Dashboards were extremely valuable and wonderful in the day. Back in the late 90s and early 2000s, there was no real way to bring together cross-functional data kept in different transactional systems across the company.
Since 2006 or so, the volume of data has become so extreme that even dashboards struggle to integrate data from different sources. Their main value was in integrating disparate data, but now dashboards have become just another piece of the puzzle.
Whenever our business units want intelligence from us, they're trying to make a decision. How many systems and dashboards do they have to go to get a piece of the puzzle to put it back together again is the issue. It takes a long time. Do I even have access? Do I even know where the data is? When I go to that dashboard, will I get intelligence? Will it be data that I then again have to analyze and piece together to make a decision?
Decision intelligence, in the way I use it, is about using a single pane of glass to surface the puzzle pieces we understand comprise a critical business decision, so it's more glaring, obvious and quicker for our business units to make decisions.
How can granting AI access to distributed enterprise data help businesses achieve decision intelligence?
Christian: Let me answer the prelude to that question: whether you should do that. If you should do that is really based on your level of data privacy protections, policies and procedures.
When I ask myself whether I'm going to go into decision intelligence, I have to do it in a way that ensures AI doesn't have access to my data. I've worked around that, or I've embedded models in my data platform so it's safe.
First, you can create simple user interfaces using AI with Streamlit, React and other development tools to build these single panes of glass. And you're servicing it from your data estate, but you're really doing it in a temporal way, so you're exposing the business to the decisions that need to be made.
How much groundwork do enterprises need to do to grant this type of access without increasing their risk?
Christian: You need to be able to apply the security controls, the row-level security, which exist in your data platform to AI so that it's another regulated account or regulated entity.
With the data platform I use, I can immediately provide row-level security to AI. Everybody who has access to an AI application that my team pushes into production, the platform and the AI tool know exactly who you are, and it knows exactly what data you have access to and what data you don't have access to, such that no data can be surfaced to you that you don't already have access to.
Your team uses a natural language querying (NLQ) tool to make it easier for people without data knowledge to construct dashboards. How did your team develop this tool?
Christian: We didn't do it the routine way. The trap that many companies have fallen into is placing a large language model on top of their data and asking it questions to chat with their data. This is a new buzz phrase: chat with your data. And it doesn't work.
Why doesn't it work? Because AI doesn't readily understand "machine." It doesn't understand your data sets. Does that tell you it's a finance table? It doesn't. Does it tell you where ARR is? No. Does it tell you what ARR means? Does it talk about how it's computed? Does it talk about how we use that in business language today, and how we use it as our key term in managing our business? No.
For years, we've been working on data governance and data readiness. Now, we need AI governance and AI data readiness, meaning the business context of the data needs to be clear to AI. We give it access to our business glossaries, data catalogs and business and technical metadata so it understands the data's context. That's number one: the semantic layer that helps AI understand, in plain English, what it's really looking at.
I use my row-level security to ensure I know which access is granted to which data. Then, I enable a user to prompt. Our AI application for natural language querying is secure: the public model only vectorizes your prompt into an array and compares it with the business and technical metadata I mentioned we collected and added to our solution. Then it sends it to an agentic on-prem workflow that says, "Go execute the SQL and Python code that already exists."
When we build NLQ, and we make data available to natural language speakers, the only aspect of AI that's being used is to vectorize the question, compare it to metadata, and then send instructions back on-prem to compute the answer using the same code we've always used so that we know it's highly accurate and reliable.
"For years, we've been working on data governance and data readiness. Now, we need AI governance and AI data readiness, meaning the business context of the data needs to be clear to AI."
Shayde Christian
chief data and analytics officer at Cloudera
How has the NLQ tool reduced time to insight?
Christian: For individuals to get insights out of their data platform, they need to understand machine language, know where the data is, know what it means and be able to use a visualization tool. It typically takes days or weeks for them to do that.
They're now doing it in 50% of the time because this NLQ tool is enabling them to cut a lot of the intensive work out of the process. Natural language speakers couldn't do that before. They would have to wait in the queue for my team, and might wait weeks or months. Now, they're able to simply, in minutes or hours, prompt the tool, get the answer, which also comes with a graphic, a visualization that they can then attach to their own custom dashboard and share with everyone.
In minutes and hours, they don't need to know anything about the data. They don't need to know SQL or Python. They don't need to learn a language to create data visualizations. It's available to them just using natural language. That time-to-insight improvement is at least 90%.
What kind of guardrails did your team build into the tool for safety, privacy and accuracy?
Christian: The major guardrail is that you have no access to our data. You have access to how our data might be calculated or what it might mean, but you have no access to the data.
The second guardrail that we've already examined is: We know who you are when you first come in, and the row-level security on our data platform gets applied right to AI. We did an extreme amount of testing to see what the accuracy and reliability of AI was in this case, even though it was marginally used. Accuracy is: Did you get it right? Reliability is: Do you always get it right?
With all our AI tools, we have a feedback mechanism. Explain why you think we got it wrong or verify that it's accurate.
Your team uses Shared Customer Intelligence to identify signals and bring them to the surface for easier, real time decision-making. Can you walk me through how the tool identifies which training or service opportunities to surface?
Christian: There was always a need for the business and IT to work together to understand what we are looking for and the technical skills to go find it.
Now, with data science and AI, IT is playing a larger role in identifying important data and signals. There's more thought partnership between IT and the business. My team and my data scientists are reviewing the data and identifying interesting patterns.
One example that we found: We had a customer opening multiple tickets on an application. We closed six of those cases in one quarter as NBF (non-break-fix). Our customer was doing very complicated things with data, so they may or may not always understand how to use the tool in conjunction with other tools they have.
Our data scientists said that's a signal for this tool: NiFi. It means this company may not know how to use it appropriately in their environment. They may want to get training on that tool, or engage our professional services team to figure out how they might use it more effectively internally.
We showed marketing, product and professional services that this opportunity existed. We automatically open an opportunity with an agent in Salesforce.
How do you want to deal with these signals? What do you want to do? What do you want to be agentified? Then, we create agents that will either automatically send a 30% training coupon for NiFi or open the opportunity for someone in professional services to reach out to our champion in that account to determine whether professional services are something they might need.
What kind of governance does an enterprise need to enable autonomous actions stemming from decision intelligence?
Christian: When you're going to imbue agents or service accounts with capabilities, you have to ask yourself: number one, is this something we want an agent to be able to do? If it is, secondly, there has to be guardrail definition around it. What are they able to do?
That second guardrail is really informed by what's on the menu. If you limit the menu as to what AI agents can perform for you, then agents can't take control in ways that you don't want them to.
The last guardrail we put in place is that we have to continually show the business how accurate and reliable our agents are at making the right decisions before they will agree. We need to have over 95% approval accuracy. We have to explain the variance.
What are the top actions CIOs and other enterprise leaders can take today to help guide their organizations out of the dashboard era and into the era of decision intelligence?
Christian: I think the first thing is to assess your proximity. To what degree are you, as a CIO, engaging in the business? The less you understand the business, the more harm you can do.
After that step, it becomes a matter of partnering with the business to discern if and where AI can be highly capable. What we're finding is the same thing our customers are finding across go-to-market strategies, marketing and customer service. These areas are ripe for workflow automation.
Lastly, it really involves the technology and making sure that you're making the right, safe technology choices that will enable AI to be successful, so when agentic applications are productionalized they not only do not do harm but they do maximum benefit to the top line and bottom line of the company.
Carrie Pallardy is a freelance journalist based in Chicago. She has cultivated experience writing in cybersecurity, technology, healthcare, and more. She currently covers a wide range of issues relevant to today's CIOs and IT leaders.
