Closing the Gap Between AI Ambition and Lab Reality Key learnings from the Think Tank at SmartLab Exchange, Philadelphia

Information architecture, metadata strategy, governance: none of this is “sexy”. Yet without it, AI initiatives stall.

Automation is not AI: clarifying the real bottleneck

A recurring theme throughout the session was the confusion between automation and AI. Many organizations already automate parts of their workflows, including the generation of reports from structured data. This is not new, and it is not where the main challenge lies.

The difficulty begins when data lacks context. Converting outputs into tables or summaries is relatively straightforward. Understanding what that data means within the experimental framework, and translating it into a coherent scientific narrative, is far more complex. This is where AI can create value, but only if the underlying data is properly structured.

The vision described by several participants is compelling in its simplicity. A scientist runs an experiment, data is captured directly from the instrument, and systems automatically process, contextualize, and generate a report ready for review. Technically achievable, yet still difficult to implement consistently across real laboratory environments.

The missing layer: context and metadata

If there was one concept that surfaced repeatedly, it was the importance of metadata. Analytical results without context have limited value. A curve or spectrum on its own does not support decision making unless it is linked to experimental conditions, parameters, and intent.

Participants highlighted the challenge of balancing raw and processed data. Both are needed, but neither is sufficient in isolation. The absence of metadata creates a disconnect that prevents effective reuse and limits the potential of AI.

This challenge is particularly acute when dealing with legacy systems. Extracting data is already complex, but reconstructing context is often even harder. Efforts to digitize through IoT and internal pipelines are progressing, yet the lack of standardized, structured metadata remains a critical barrier.

Fragmentation across systems and vendors

Another dimension of the problem lies in the diversity of instruments and data formats. Different vendors expose data in different ways, creating inconsistencies in how information can be accessed and integrated.

This fragmentation makes it difficult to build scalable data pipelines. Some participants pointed toward emerging solutions, such as local agents trained on specific data sources and the adoption of common data models. These approaches show promise, but they also highlight the need for stronger collaboration across the ecosystem.

Without a shared foundation, every integration becomes a bespoke effort, slowing progress and limiting impact.

From vision to execution: aligning strategy with reality

Beyond data, a second gap emerged: the misalignment between strategic aspiration and execution. There is, as participants put it, an "elephant in the room":

Organizations feel they cannot afford not to do AI.

And yet, they struggle to articulate:

1. Why they are doing it
2. What success look like
3. How it translates into business or scientific impact

Without this clarity, organisations risk pursuing solutions that fail to translate into measurable impact.

One analogy resonated strongly:

Deploying advanced AI without the right foundation is like placing a high performance car on an unpaved road.

The capability is there, but the environment prevents it from delivering value.

This misalignment also plays out across the organisation. Middle management is often asked to implement AI initiatives without a clear narrative of value, while scientists remain focused on delivering experiments rather than transformation programmes.

At the same time, a disconnect persists between data producers and data consumers. Scientists generate data to answer specific experimental questions, while organisations seek to reuse that data to accelerate timelines and improve decision‑making.

Bridging this divide requires more than technology. It requires alignment in objectives, incentives, and expectations.

Adoption and culture: the silent enablers

Although culture was not the top voted barrier, it remains a critical factor. Introducing AI into scientific workflows is not simply a technical exercise. It requires trust, clarity, and relevance to the task at hand. Adoption depends on demonstrating value in context, not on imposing technology from the top down.

Senior leaders often see only the final dashboard or model output, not the complexity underneath. Change agents or champions play a critical role in translating complexity into understanding, and ambition into achievable steps.

Participants noted that generic AI tools often fail when applied to specialized scientific problems. The right tool must match the use case, whether it involves scientific reasoning, regulatory reporting, or operational optimization. Language matters too. Terms like digital twin can obscure more than they clarify if they are not grounded in real use cases and domain constraints.

As one participant put it: “Using a generic AI assistant for all problems is like forcing a square peg into a round hole”.

Different AI models are good at different task (e.g., scientific, legal, operational ...) and need to be applied accordingly.

Closing the gap: where to start

By the end of the session, one message was clear: AI‑readiness is not about having more data, bigger models, or shinier tools. It is about: context, alignment, purpose and the willingness to invest in the foundation before chasing outcomes.

The gap between ambition and reality does not close with another platform or pilot, it closes when organisations align data, people, and intent around problems worth solving. It requires capturing data with context, enabling seamless exchange across systems, and aligning initiatives with clear business objectives.

When these elements are in place, AI becomes a natural extension of the workflow rather than an isolated capability.

The path forward is not about adding more intelligence, but about making existing data intelligible.