Last week I attended Everpure’s annual Accelerate conference in Las Vegas, where the vendor laid out its data primacy thesis – where CEO Charlie Giancarlo argued that it is now imperative that enterprises push applications downstream and operate with data at the core. The idea is that 50-plus years of application-centric IT has the hierarchy the wrong way round, and that AI is the thing that finally makes the change necessary.

However, as we know, announcements and intentions are one thing, whilst the reality of enterprise technology and data use are another. diginomica’s own research on this topic lays out how inaccessible, unusable and fraught with risk effective data use is for most senior technology and business leaders. And so during a panel last week, I was keen to hear from two companies about how they are thinking about centering data and their approaches for data use in the coming years. What was interesting was that the companies are thinking about this in two somewhat different – albeit valid – ways.

On one side was Eric Frost, Director of Infrastructure Services at UGI Corporation and AmeriGas – a US energy holding company commonly known for the propane cylinders you pick up at Home Depot, alongside a regulated gas and electric utility business. On the other was Pradeep Bandaru, Head of Platforms & AI at Sanofi, the global biopharma group largely focused on immunology medicines. One is trying to pull scattered data together for more organized use. The other is pushing compute back toward where data already sits – on-prem, at the edge, or in regulated environments where movement creates latency, governance and sovereignty issues.

Keeping the lights on

Firstly, UGI’s Frost explains that he runs three teams across five business units, and his key problem is fragmentation of data. Frost said:

It’s been a very large challenge, especially in the utility business, where a lot of our data is very much siloed. We have industrial, we have operational technology, we have a lot of different data hidden all over the business. So trying to bring that together in an organized fashion has been an ongoing challenge

We’ve partnered with Pure, with Nutanix and a lot of other companies, and there are a number of initiatives happening right now to start bringing all our data together. 

That task – bringing that data together – is not one, tidy data-warehouse exercise. Much of it sits in operational systems that were never built to talk to anything else – and in a utility the cost of getting it wrong is not minimal. Frost said:

We’re responsible. If someone has a gas leak, we have to be available, online and responsive, because lives are in jeopardy.

The route through it, for now, is using AI for the unglamorous work first. UGI began a partnership with AWS earlier this year. He said:

We started about six months ago. We partnered with AWS, and it has been a really strong partner. They came in, met with the business, tried to understand what our core challenges are, and looked at where some really good opportunities are to leverage AI as a force multiplier to improve the business’s performance. 

Starting in January, we had really deep meetings, and they powered us with onsite training to help bring AI into reality. They didn’t just show us how to do it, they taught us how to do it. So I now have program managers who are actually working in AI, developing small use cases that are really impacting our business.

The focus is deliberately on customers, not technology. Nobody, as he put it, wants to be on the phone about a bill or to reorder propane. And as such the organization is looking at automation and optimization opportunities, to move away from ‘old-fashioned’ ways of doing things (namely, data scattered across the organization). Frost said: 

We’re trying to break that mold and modernize – not only for efficiency for the company, but also for customer satisfaction. The better we can serve our customers, the happier they’ll be. So we’re looking at AI as a force enabler to better serve and better assist our customers.

This is early-stage work, and Frost did not pretend otherwise – programme managers are running small use cases that are only starting to become visible in the business.

The centralization tax

Interestingly, Sanofi’s Bandaru is trying to achieve similar outcomes – getting access to usable data – but is approaching it somewhat differently. Where UGI is consolidating – thinking about how scattered data makes for fragmented decision making – a distribution change is coming to Sanofi. The global biopharma has spent a decade pursuing centralization, building structures for data under the industry’s FAIR principle. Bandaru said:

We’ve invested pretty heavily over the past decades to build both the organizational structures and the technologies to enable what we call FAIR data – a principle in our industry around making data findable, accessible, interoperable and reusable.

That decade of investment, he said, came with a bill. Bandaru said:

We’ve basically built out entire data engineering and infrastructure engineering arms to build centralized data management platforms, both in the cloud, on-prem and at the edge. Historically, a lot of the energy was put into the varieties of use cases we have – whether we acquire companies, or get data from consortiums or academic partnerships – all directed towards that centralized infrastructure model. 

The tax that’s been incurred over the past decade or so is: the more you push to centralization, the more it often competes with your time to outcome. So what we’re really refocusing on now is decision latency. 

As such, Sanofi is heading the other way now and thinking about an approach that embraces federated compute – bringing the compute to where the data already sits. He said:

That’s important to us for a variety of reasons: regulatory, data residency and sovereignty, as well as practical reasons. It gets away from that centralized model. 

Practically, what that means for data being in the driver’s seat is that as you adopt more of these federated approaches to compute – and what becomes more and more important is being able to track the lineage and the provenance of your data as it flows across different systems and compute platforms and is processed. All of that needs to be stored effectively as a provenance trail, and that also allows us, very practically, to audit our data and workflows.

Context is the new compute

Sanofi is arguably further down the AI road than most, running agentic workflows in production against more than 20 petabytes of R&D data. From there, Bandaru’s argument is that the model is no longer the hard part. Bandaru said:

Models are increasingly a commodity. The real rate-limiting step always comes back to the data layer.

What the agents actually need, he argues, is not more models: 

We build a lot of models internally, we use a lot of external models, and over time there’s not really a lot of thinking required to choose the right types of models and deploy them. The real rate-limiting step always comes back to the data layer. What these agents and agentic workflows really need is context – and what I like to say is that context is effectively the new compute.

It is a useful anecdote for anyone deciding where to spend (especially as the ‘token-maxxing’ discussions rumble on). If the models are converging as a commodity, the differentiator – and the cost – moves to the data layer and the plumbing that feeds it: the context, the predictable latency for an agent’s tool calls, and the governance to keep autonomous systems honest. On that last point Bandaru said:

As agents start to do autonomous, independent activities, they need guardrails, harnesses – a real governance layer – to make sure they’re not going off the rails.

And all of this is in aid of the use case – which in this case can have life changing consequences, given the nature of Sanofi’s business. Bandaru said:

Whether you’re running an early discovery process – actually discovering molecules and trying to test hypotheses in a lab – or in other areas closer to the clinical development side, where you may be running a clinical trial and testing the effect of a drug molecule on a cohort of patients. 

I think ultimately it comes back to that same thread of data lineage and data provenance being the thing that ties all of these use cases together – [from] the moment the data is actually created, whether that’s a physical environment like a laboratory, a manufacturing plant or a clinical trial site, and then as it gets processed, and ultimately as that processed data drives a decision somewhere. 

So that’s how we think about faster time to science: really just accelerating that entire life cycle of data, from being produced all the way to being processed, and decisions being made.