How do you ensure that agents are able to get the structured data and the context correctly?
There are two dimensions. The first is how do you make sure all your data is actually accessible by agents. If you think about data estates [places a company keeps its digital information], there is a bunch of it in Databricks or SAP applications. Solving that for customers is actually a pretty important piece to have a starting point where they can be successful, including getting to the context. If you don’t have access to all the data, you can’t build the context for it. The hard part here is we also have to make sure that legal contracts and images can also participate in this borderless lakehouse. The second part is for the agent to understand the business and which data to access. And once we are able to see all the data, we can build that context. This is because enterprise data is too expensive for agents to process from scratch. This is where building up the right knowledge really helps; it tells the agents which database they should go to.
What are some of the guardrails enterprises must put in place?
Enterprises need to give the agents the minimum amount of permissions needed to get to an outcome. And so one of the benefits of us actually having integrated the agent platform we have with Gemini Enterprise and with the Agentic Data Cloud is that we have a consistent security model across all the systems. It has an extremely high trust and governance with our customers because they know how to control it top to bottom. The second part, from a context perspective, is we have very high-quality context and also it is well governed. The agent can only get to the context that it is allowed to see.
We also work with customers depending on the use case. For some customers, we work with humans in the loop. In some cases, they [AI agents] can be more autonomous. For example, in customer support, you will typically see a bit more autonomy. In financial services, you will usually see humans in the loop.
Does that mean trust is becoming a data problem?
I would say that it’s a must-have but it’s not sufficient. That means it is probably where customers are spending a majority of their time because the trust in data ultimately influences how agents are reasoning and the outcomes they are delivering. It is not only from a security perspective; you also want to have trust that the answer or the action is the correct one. Data is definitely where customers are spending a lot of time. But there are other elements where Google has a very good track record and capabilities in making sure that we create a very tight security model for agents and then, of course, Gemini as a model I would say has a lot of guard rails. Not every vendor has the same amount of discipline and guard rails as Google has with Gemini.
How do enterprises strike a balance in getting adequate returns on their AI investments?
I think customers really care about trust, cost, and user experience. The return on investment is usually going to be in one of three dimensions. The best model is the smallest model that drives the outcome you need — and if you look at models that we have like Gemini Flash, it’s an extremely capable model compared to its cost. Most of our customers don’t even need to use Gemini Pro because Gemini Flash is so capable. Context will matter a lot. The less context you have, or the less good your context is, the more reasoning loops the agent will do, the more tokens it’s going to consume, and the more compute it’s going to consume. This is where a lot of our differentiation comes with our context.
Between structured and unstructured data, which gives you a better context and more desirable outcomes?
Structured data is something we have dealt with for the last 40 years. We now have an opportunity to take this unstructured data, where a huge amount of the enterprise knowledge is actually locked and unleash that for the practitioner directly and through models. I think one of the most exciting things is how we make that 90 per cent data — that traditionally has been dark data — light it up and help businesses deliver better outcomes along with traditional structured data.
For example, the use of Gemini in BigQuery (Google’s autonomous data and AI platform) has increased 30 times year-on-year. Gemini is usually used on unstructured data. We have built into the data platform capabilities that help customers unlock the value of unstructured data, and the demand is just growing like wildfire. We are just seeing a huge appetite from customers right now in getting the benefit of that.
