Solidatus CEO: Governance will speed up AI

Phil Dutton thinks most enterprises have the AI equation backward. Governance isn’t what’s slowing them down, he argues — it’s what will let them go fast.

“Some short-term pain, effort and cost will ultimately unlock their ability to go fast, and to go fast safely,” the Solidatus cofounder tells TechInformed. The organizations that skip that step are betting the data foundations that got them here will carry them into the next phase. He doesn’t think they will.

Dutton trained as a software engineer before moving from Australia to London early in his career. He spent almost two decades as a financial services consultant, “parachuting” into complex organizations to solve problems, and worked on regulatory transformation programs, system upgrades and implementations.

That work taught him two things. Big organizations are hard to change because everything is connected to everything, but most people know only their own silo. And transformation programs fail, more often than not, because the people running them lack the context to see those connections.

Solidatus, a data lineage platform Dutton cofounded in London in 2017, was built to supply that context: a blueprint, as he describes it, of how data moves through an organization. Its customers include BNY, HSBC and the London Stock Exchange Group; HSBC and Citi are also investors in the company.

In this interview, Dutton tells TechInformed how ready organizations really are for AI, why it has been slower to reach high-risk areas like credit decisioning and what would make customers trust an “AI-powered” company.

How ready are organizations’ data foundations for AI?

I think it depends on the industry, and it also depends on the leadership in those industries.

We work with some very forward-thinking organizations. They can see that some short-term pain, effort and cost will ultimately unlock their ability to go fast, and to go fast safely.

Others say, “It got us here, and it’s good enough to get us to the next phase.” Though, [the truth is] what got us “here” is not going to get us to the next phase.

Over the last 40 years we’ve been trying to get to grips with technology and the pace of change, and the constant new capabilities, and we didn’t have time to step back and ask, “If I was doing this properly, how would I do it?” We always had a deadline to meet or an opportunity to attack.

I think we’re at a real inflection point with AI. If we get it right, it’s fantastic and what’s possible is going to be amazing. If we get it wrong, it could be a really terrible existence for a while until we decide to get it right.

So a lot of organizations are now focused on foundational activities to unlock that. It’s a bit of a tale of two cities: some are, and some aren’t.

How much AI is actually in production at the enterprises you work with?

There’s been lots of activity, and everyone’s really excited, but we haven’t seen a large amount move into a full production enterprise state in a high-risk category.

There are low-risk categories people are working on and happy with, like chatbot-style tools that give you information. But then there are the riskier areas, like credit decisioning, where someone gets a loan or a mortgage, or doesn’t, based on an AI’s assessment. There are regulations requiring fair lending, so that’s a riskier scenario than “tell me about my account” or “what transactions have I made in the last few months?”

There’s now a realization that we’ve got to change how we’ve been operating. There’s lots of activity around how we govern not only the data going into the AI, but the AIs themselves.

What we’re seeing organizations realize is that they don’t want 100,000 agents running around, half of them orphaned because they have no owner and no one who’s accountable for them.

We need to understand not only the data that feeds the agents, but the agents themselves, and what those agents can operate over. Is it over this set of data, for this type of reporting? So there’s a lot of activity figuring out what enterprise AI adoption really looks like in the organization.

What’s new about the risk of running AI on bad data?

In the past, AI was very much controlled at the center of an organization, by people who were very proficient and skilled at interacting with models, whether machine learning models or true AI models.

They understood exactly the data they needed to put in to get the outcome they needed.

With the LLMs we’ve got now, we’re pushing AI to the edge, out to everybody. Not everybody has the same context over that data. So how do I know the answer it gives me is correct? How do I know it hasn’t introduced a bias? How do I know the question I’ve asked has been framed in the right way? The challenge is that you don’t. The risks are greater because you don’t have the context to pick whether it’s right or wrong. If you’re an expert in an area, it’s instantaneous: the sniff test, and questioning: does that look right? When you’re working on data you’re not familiar with, it’s much more difficult.

What would make a customer trust an “AI-powered” company?

I actually did my honors dissertation on trust in electronic commerce, back when the internet was being born and rolling out to people to start transacting on. How do we create trust? How do we get someone to put their credit card into what was a bit of a black hole and not have it stolen?

One is transparency: can I see and feel the artifacts that create trust? Understanding how a set of flows is connected creates trust. I can see it with my own eyes.

Another is evidencing that this isn’t the first time, that it’s consistently done. It’s a well-worn process, and I trust it more than the first time you do something, when you’re always a bit nervous.

It’s like jumping out of an airplane. The first time you’re absolutely freaked out. If you’ve jumped 100 times, it’s like stepping out the front door of your house. You’re very familiar with the process. It’s the same in this space: being able to evidence, with hard facts, the flow of data, what we actually have and how things interact, and having that through time is a powerful factor.

It’s also about giving people the context. It’s not just this “thing” in isolation; it fits into this thing, which fits into this thing. I can see the flow, the dependencies and the impacts laid out in front of me, and that lets people trust the data they have.

It also makes the needle jump out of the haystack, as I like to say. [It’s important that] you can find where the issue is really quickly instead of having to search for it.

Where have you seen banks adopt AI successfully?

We typically work with very large, very complex, heavily regulated organizations. At [BNY], one of the world’s oldest banks, the organization has in the past been seen as slow-moving and risk-averse.

We did a lot of work helping them build out a foundation of data, mapping it through so they understood their processes, products, regulatory obligations and impacts across their estate. That has enabled them to adopt AI probably much more quickly than they would have without that foundation.

They’re one of the ones that saw that what got us here is not going to get us to the future.

For those still untangling their data, where should they start?

The biggest thing I see is that we often get stuck in analysis paralysis. We think about it deeply, but we don’t do anything. The best advice is to start. If you don’t start, you don’t make progress. There are lots of opportunities in that starting, and as you go, to pivot and course correct, because the first answer isn’t always the right answer.

How do you have your coffee?

I’m one of, I think, less than a third of the population who doesn’t drink coffee or tea. I think a lot of people want that kickstart in the morning, but I wake up and I’m ready to go.

This interview has been edited for length and clarity.