Safe Software’s Don Murray is one of the many tech executives betting that artificial intelligence (AI) will eat the world. But the CEO of the Surrey-based data integration company says Safe makes an important distinction: AI is an assistant, not an authority.
Founded by Murray and Dale Lutz in 1993, Safe Software offers a software platform to help enterprises share different types of information, including spatial data. Today, Safe provides a range of data solutions to more than 25,000 customers and over 200,000 users across 121 countries. A spokesperson told BetaKit that Safe is on track to achieve $112 million in revenue this fiscal year, adding to Canadian tech’s growing centaur list.
Murray discussed the emergent role of AI in enterprise with BetaKit ahead of the announcement of two new features for its FME data workflow platform at Safe’s ‘The Peak of Data and AI’ event in Seattle this week: FME Data Virtualization allows developers and data teams to create APIs enabling large language models (LLMs) and AI platforms to interact with enterprise data securely while adhering to the company’s governance protocols; FME Realize is an augmented reality (AR) tool that can connect an organization’s data ecosystem, including digital twins, to workers out in the field.
BetaKit sat down with Murray to discuss the intent behind the OpenAPI protocol in Safe’s new features, how AI killing SaaS is a full circle moment, and deploying digital twins in the field.
The following interview has been edited for length and clarity.
Can we start with the technical considerations to implement Data Virtualization and what your customers are looking for?
Organizations have many, many different systems, and in the past, when they would try to build an application, each application would have to deal with all the ugliness of stuff like authentication for each system. So if you’re connecting to seven systems, they probably have three or four different ways of authenticating, and each application would have to do that separately.
We’ve built this no-code way of building a data virtualization layer that is OpenAPI compliant, which is important. Because of OpenAPI, you can [share data with] LLMs like OpenAI and chat with it. So now you’ve opened up that data to the non-technical user.
There are experts in every field, but they don’t know how to program. So that’s been a barrier, and just one reason we’re excited about the data virtualization layer.
RELATED: US-based JMI Equity buys out Safe Software co-founder in strategic investment
The other interesting thing is that the way people share data is going to fundamentally change. We see a trend where [data will be shared] through these OpenAPIs, and then there will be applications built on top.
It’s gone that extra step of making data accessible to everybody, but also secure. OpenAPI really allowed people to redefine a secure layer. So if you build an API, you will know what people can access. They won’t be able to get any access to anything other than what you decide.
The other big thing is that [Data Virtualization is] no-code, so you can also build these layers with no code through a graphical interface, which is quite different compared to other data virtualization technology from other vendors.
As I understand it, Data Virtualization is essentially a virtual machine space where AI and data can play together and then disappear when the work is done. Is that a function of convenience or security?
With virtualization, in this case, everything’s live. We’re not storing any data. So it’s virtual in the sense that it isn’t physically stored anywhere, and it could be an AI agent [running] underneath.
You make a [data] call and an AI agent or an AI workflow underneath is analyzing and giving you data back. You could make a call asking to count the number of cars in an image, and then it’ll give you back an answer to your question.
The virtualization makes it so that multiple data sets—and you don’t know how many data sets are underneath—are pulled together. So through the API, you see a single homogeneous thing, but it’s virtual. It may not actually exist in the underlying data.
Are you doing this through proprietary work, or is there an industry standard?
OpenAPI is the standard. By making that decision, any application that understands that specification can read [your data]. LLMs are the hip example right now, but there are lots of other ones as well.
Organizations themselves want to be able to build really specific applications for their own business, and this just makes it really easy for them to build their own applications on top of a foundational virtual layer.
Imagine I built something, and part of it was on Salesforce. If I switch out from Salesforce to HubSpot, then the customers using my OpenAPI layer will not know because that’s on the other side of the interface. It gives organizations the ability to change their infrastructure without having to rewrite the applications above it. In the past, any time you changed something, it would impact everything that was directly accessing it.
Microsoft CEO Satya Nadella essentially said SaaS apps are dead because AI will be the new connecting force between a company’s data and its business logic. Does that align with what you’re saying about replacing the infrastructure?
Yeah, I subscribe to that view.
If you think about Salesforce, it’s a database with some logic on top and an interface, right? Now, with AI, you’re gonna be able to do a lot of this directly on the data. The data stores itself. We are not SaaS, we run in many different environments, so I’m not worried that that will negatively impact us. In fact, I see it as a way to be even more empowered.
If we can go directly to the data and just enable people through AI prompts—even when you have AI and an automation workflow that isn’t run by a human, at the end of the day, it’s all prompt-driven. Prompt engineering is a real thing.
Don’t you have SaaS customers that might be worried about the death of SaaS?
It’s more like we have customers who use SaaS. SaaS is interesting because, before it, people didn’t want their data owned by a proprietary interface; they wanted open databases. Then along came SaaS, and the data is locked in through a proprietary interface. We’re kind of coming full circle on that. SaaS is dead or dying, and we’re going back to the model where the data is there and [solutions] are built on top.
We’re also seeing other things like open table formats, such as Delta Lake and Apache Iceberg, where the idea is you can pick whatever database engine you want on top of those, rather than have to do a big database migration. That’s being pushed by big customers for some organizations like Netflix and Apple.
We talked a bit about Data Virtualization having visual components beyond some raw database output. How does Realize, AR, and digital twins fit into this equation?
Digital twins have always been locked on a computer screen. If you go to one of the airports that we serve, the control room will have digital twins on big monitors, but the guy in the field walking around the airport doesn’t have any access to that.
Now he’s able to look at his iPad or iPhone, and have a window into that 3D environment.
At the airport, for example, the guy in the field could walk by, point to a car, and learn that the car’s been there 30 minutes too long. A camera system is watching all the vehicles—the airports know every car that’s parked in front of an airport, they know exactly how long it’s there. That’s part of the digital twin.
The AI agent knows where the field guy is and, as they’re walking around, it’s doing analysis and suggesting things for them to check out.
With text-driven prompts, it feels like we’re resetting UX by 40 years. I’m wondering if we’ll eventually get back to graphical interfaces after a few years of chat boxes.
[Laughs] We might, right? I talk to my phone and there’s no person on the other end now. If I wake up in the middle of the night and I have an idea, I just talk to my phone, and it indexes it. I think it’s gonna be much more natural. Typing is horrible when we can just talk to it.In some ways, nailing the perfect prompt can be harder than learning to use the data tools. Do you anticipate a circumstance where the systems that receive these prompts can solve for a lack of sophistication from the user?
We’re seeing that in development, even the way we build products. We’re also revealing our new AI assistant: even though we’re in a no-code environment, you have this AI agent there helping. It makes you so much more productive, but it’s an assistant.
I think that’s an important thing that we talk about at Safe Software: AI is still an assistant, it’s not the authority.
Signal CEO Meredith Whittaker has said that giving an entire system like an AI agent root access is a really, really bad idea for both encryption and privacy. Does that criticism land for you, or are you doing something to combat it?
I’m not familiar with that interview, but the person who creates a virtualization layer decides exactly what is going to be available.
A simple use case we’re seeing over and over is that an organization has a table with 12 columns, and four of them are sensitive. So they can just use data virtualization as a really easy way to make available the eight columns that aren’t sensitive. I think it would be a very bad idea to give OpenAI, or any LLM, preferred access to all the data out there.
The other really exciting thing is we’re gonna see LLMs that are purpose-built outside of the public internet. The problem with the public internet is that most of the data is not on the public internet—like health records, finance, oil and gas, etc.
What [these companies will do] is make a service and start selling their data through the LLM. They’re gonna have an LLM that is trained on the data that they own, [data] that you won’t be able to find, and they’ll monetize that [access].
Canadian computer scientist Yoshua Bengio says there is a catastrophic risk if we build AI agents but do not prevent or control any sort of self-preservation instinct. What do you think about that?
I think it goes back to assistant or authority, right? We have a company that builds solar farms, and they have AI agents analyzing drone imagery. What the AI agent does is it sends something to a human and says, ‘Oh, you should check this one out,’ because there are literally thousands of images.
So then the human is the authority, rather than the AI agent itself deciding to call the police or report something and engage a third party.
So even with AI agents, at the end of the day, they are an assistant. They’re gonna take automation to the next level. But I think if you take the human out of the equation, we’re asking for a bit of trouble, because they hallucinate.
With files from Douglas Soltys.
Feature image courtesy Safe Software.