[Image from OpenAI’s image generation software]
That’s the stark assessment from Mikael Hagstroem, CEO of LabVantage. While many labs possess vast amounts of data, more isn’t always better if scientists struggle to access and contextualize it effectively.
For Hagstroem, the ability to converse with data represents “a shift towards a more human centric approach to data analysis, or AI, if you will, where AI then becomes the bridge between the user and the data.” This vision builds on traditional LIMS functionality. “In the beginning, a LIMS was to connect the data to the workflow, contextualize it. Now the contextualization and the data sources have just grown exponentially, but the challenge remains the same.”
LabVantage aims to strengthen this connection by focusing on what Hagstroem calls the “connective tissue around the corporate data.” One pillar is its use of biochem-specific ontologies to ground AI and provide “guardrails.” Another element addresses the operational burden labs face managing complex informatics systems.
Hagstroem noticed many labs found system upkeep secondary to their core focus. Clients often would state something to the effect of: “‘it’s not my core competence to manage the master data… I need you to do it’,” he recalled. To meet that need, LabVantage developed what it terms ‘SaaS 2.0’. This approach moves beyond traditional Software-as-a-Service, inverting it to what Hagstroem terms “Services-as-a-Software.”
Mikael Hagstroem
Under SaaS 2.0, LabVantage taps its cloud platform to actively “support the use of the environment” by taking on essential data-centric tasks—like connecting instruments to establish a “golden source” of data, performing configuration, and managing master data—on the client’s behalf. These “data-centric services,” Hagstroem argues, help labs navigate complexities from data fragmentation to regulatory demands, allowing them to concentrate more on core scientific objectives.
Ultimately, Hagstroem emphasizes, the technology and services are successful “the moment the lab workers feel they can have a meaningful conversation with their data.”
Data, data everywhere: When data lakes don’t quench the lab’s thirst
Data fragmentation persists in R&D labs, even with data lakes. “Data is fragmented in all sorts of ways,” Hagstroem states. He highlights challenges with specialized data types: “Think of a DNA structure with various omics fitted to it. You’ve just modeled it… Where do you save that? There is no RDBMS [Relational Database Management System] where you can save that,” he elaborates, explaining that complex structures or images often defy traditional storage.
For decades, the response was often adding another silo—”if you have 10 databases, create number 11,” Hagstroem quips. Yet even data lakes don’t guarantee usability. “The reality is that whether the data is fragmented in or outside the data lake doesn’t really change life in the lab because it’s hard to get data out of the lake,” he observes.
Often, data in the lake remains inaccessible to scientists lacking SQL skills. This points to a fundamental disconnect:
If you ask an IT person, they would think of data fragmentation differently than a lab person.
From untangling data to taming AI
Given these persistent challenges in making stored data usable, attention naturally turns to AI as a way to bridge the gap. Yet implementing AI, especially agentic AI designed to perform tasks, can come with hurdles.
Hagstroem identifies three AI-related challenges when data isn’t properly connected and contextualized:
Hallucination: Models can invent incorrect information. Hagstroem uses the amusing, yet pointed, example of an AI identifying a picture of his dog: “…I took a picture of our dog playing in front of the TV and it said, ‘this is a cat snugging up with blankets.’ And I said, that’s a problem. Is it a cat or is it a dog?” In a lab context, where hallucination rates can be high (as are the stakes), such inaccuracies are unacceptable.
Verifiability: Regulated lab work demands traceable proof, but many AI approaches use probabilistic methods (like vector databases) that clash with scientific and regulatory rigor. “Working with probability doesn’t sit well with scientists – they tend to favor absolutes,” Hagstroem states. Consequently, “The regulator is not going to accept that you say it’s probably this data source,” he warns. “They’re going to want to see evidence… and the lineage the whole way back to the source data. FDA doesn’t accept that approach.”
Competence: Public LLMs often lack deep biochemical knowledge. “When we talk about names of omics and their number in gene sequence numbers… there’s not enough data in the public domain to train an LLM,” Hagstroem says.
Grounding with biochem-specific ontologies
LabVantage tackles these AI hurdles with ontologies and a semantic platform, focusing on data connections rather than just the AI agent itself. This focus differentiates their strategy, Hagstroem observes: “Most companies focus their agentic AI work on the agent itself, as opposed to the data connection or connective tissue around the data.”
LabVantage’s approach to deep linguistic understanding is grounded in biochemistry – a capability cultivated over decades. “That work started in the university 30 years ago… that is LabVantage owned and LabVantage specific,” Hagstroem explains. This biochem specialization distinguishes them from broader data platforms, whose linguistic foundations might lie in other fields.
The ontology creates “guardrails” using the lab’s own data to maintain integrity. “Through the ontologies we put guardrails so that your internal data becomes the control,” Hagstroem explains. “We want to make sure that if you have a truth… that truth remains the truth. And therefore, if an agent is drawing a conclusion that’s not in line with the truth, that agent gets killed right away.” This method turns verified data into the control mechanism, grounding AI in scientifically valid reality.
Laying the foundation for the data-savvy lab of the future
While the underlying semantic technology is complex, Hagstroem stresses that practical, straightforward applications are key to building user trust. He points to “…small use cases like the agents that tell you this test has already been run… They are very popular and they quickly gain acceptance,” noting these wins drive cultural adoption. For instance, agents might flag duplicate samples or experiments, prompt about missed tests, or automatically surface relevant research papers.
Ultimately, these technologies point toward Hagstroem’s “lab of the future,” where scientists can have contextualized conversations with their data through empowered human workers and “digital research assistants” that augment the workforce. The core business goal is clear: “How can we provide more products and more effective drugs to the market, faster, without increasing resources? And that’s where the AI bots come in.”
