AI adoption in rail is already well underway.
Predictive maintenance, computer vision, machine learning in timetabling and operations, IoT across infrastructure and rolling stock: the sector has moved quickly in these areas, and rightly so.
The more specific question this article addresses is whether that same readiness extends to HSEQ, and more particularly to incident investigation and root cause analysis. It is a different context, with different data challenges and higher risks if the foundations are not right.
In some parts of rail, yes. In others, the foundations are still being built.
The language coming from GBRX, from rail leadership, and from the reform agenda is consistent: digitalisation, data-driven decision-making, AI-enabled operations. These are not aspirational themes any more. They are active priorities, and organisations across the sector are being asked to think seriously about what they mean in practice.
But ambition and readiness are not the same thing.
In a sector where the consequences of getting things wrong are real and well understood, evidence and proof of operational value are non-negotiable. The more useful starting point is not which platform or provider to choose. It is whether the organisation is ready, and if not, what it would take to get there.
The data quality challenge
The rail industry generates substantial volumes of safety data: incident reports, near-miss observations, hazard records, and close calls. Most of it is captured, but far less of it is fully understood. A significant portion usually sits across legacy systems in inconsistent formats, recorded before any structured investigation methodology was in place, and has never been properly interrogated at all.
That is the most concrete expression of the gap between AI ambition and AI readiness. AI systems are only as good as the data they work with, and in a safety-critical context, outputs built on inconsistent or incomplete data are not a neutral outcome. They are a risk in their own right.
The good news is that this is fixable. But it requires organisations to approach the question of data quality deliberately, rather than assuming the technology will resolve it on their behalf.
The gap between data volume and data value
Having a lot of data is not the same as having useful data. Rail organisations that have been capturing safety records for years without a consistent investigation methodology behind them can find that their datasets are difficult to analyse at scale, not because the volume is insufficient, but because the underlying structure is inconsistent.
When root causes are coded inconsistently, near-misses captured in free text with no shared taxonomy, and investigation reports vary in depth and format across regions and teams, the data cannot be meaningfully aggregated or analysed at scale. Each of these is a reasonable reflection of how safety work has historically been done. Collectively, they are a significant barrier to the kind of AI-assisted analysis that rail leadership is trying to enable.
Manual review of large datasets is also inherently limited. A safety team working through thousands of records cannot realistically identify the patterns, recurring themes, and latent risk clusters that sit across the full breadth of the data. Individual incidents get investigated, but systemic risk remains largely invisible. AI is genuinely well placed to close that gap, but only when the data is structured well enough to support it.
Five things to consider before adopting AI for investigation in rail
Here are some questions worth working through before making an AI investment in the investigation space, whether you are starting from scratch or building on an existing system.
1. Understand what problem you are actually trying to solve
AI means different things in different contexts. In investigation, it might mean reducing the documentation burden on investigators, improving investigation quality and consistency, surfacing patterns across a large portfolio of historical data, or all three. These are related but distinct problems, and the right solution depends on which one is most pressing for your organisation. Starting with a clear problem statement is more valuable than starting with a technology.
2. Assess your data quality before you commit
Before deploying any AI analytics capability, understand what your existing HSEQ data actually contains. How consistently has it been captured? Across how many systems? In what formats? How much of it is structured versus free text? What investigation methodology, if any, underpins it? The answers to these questions will determine what AI can realistically deliver from your data, and they will shape the implementation approach significantly.
3. Standardise your investigation methodology first
AI applied to inconsistently structured investigation data will produce inconsistently structured insights. If your investigation methodology varies across teams, regions, or contractor organisations, the data those investigations produce will reflect that variation. Establishing a consistent methodology, with a shared taxonomy for root causes and contributing factors, is a prerequisite for meaningful AI-assisted analysis, not something to address after deployment.
4. Check whether the AI you are evaluating was built for investigation
There is an important distinction between AI that has been applied to investigation and AI that has been built for it. Most AI capability in the broader HSEQsoftware market focuses on data capture, reporting, and general analytics. These are useful. But they are not the same as AI that is embedded inside the investigation workflow, designed around root cause analysis, human factors, and the specific demands of safety-critical investigations. Ask the question directly: was this built for investigation, or was it adapted from something more generic?
5. Plan for governance, not just deployment
AI outputs in a high-risk environment require human review. Every output should be a suggestion, not a decision. Before deploying AI in your investigation process, establish how outputs will be reviewed, who has the authority to accept or reject them, and how the organisation will build confidence in the technology over time. Phased adoption, starting with one or two features and expanding as trust is established, is a more sustainable approach than full deployment from day one.
What AI built for investigation looks like
There is an important distinction between AI that has been applied to investigation and AI that has been built for it. In March 2026, Verdantix, the independent analyst network whose research is used by HSE leaders and procurement teams across industry, assessed COMET AI Assistant. Their view was that the USP is not the AI itself but what it is built on: investigative expertise, RCA methodology, and years of failure data knowledge embedded in the process. The end-to-end focus on investigation data quality, they noted, is a genuine and compelling differentiator.
For rail, a sector where evidence and proof of operational value are prerequisites for trust, that difference is important. The question of whether AI has been purpose-built for investigation or adapted from a broader capability is becoming a central evaluation criterion.
How COMET supports AI adoption in rail
COMET is a dedicated incident investigation and root cause analysis platform, built by career investigators, with a growing track record in rail, including work with Network Rail, HS2, East West Rail, and Enable. Our approach to AI adoption is designed around the reality that most organisations are not starting from a position of perfect data and consistent methodology. We work with organisations at whatever stage they are at.
The data healthcheck
For organisations that are uncertain about the quality or readiness of their existing HSEQ data, COMET offers a structured data healthcheck. It assesses volume, consistency, format, and analytical suitability, and produces a clear picture of what the data can currently support, what needs to be addressed, and what the roadmap toward full AI capability looks like. It answers the data quality question with evidence rather than assumption, and it is a practical starting point that does not require a full platform commitment.
COMET AI Assistant
COMET AI Assistant operates across two pillars. Input AI sits inside the investigation workflow itself, available at the right moment in the process to make the investigator's job easier, more consistent, and less dependent on individual experience. The principle is that investigators should be editing, not creating. Every output requires human review and approval before anything is committed.
Insight AI, delivered through COMET Companion Plus, interrogates the entire investigation portfolio through natural language. Safety leaders can ask which root causes appear most frequently, which human factors keep recurring, which preventive actions correlate most strongly with incident reduction, and where early warning patterns precede serious events. No fixed queries. The answers come from your data.
Phased adoption
COMET AI Assistant is designed for organisations that want to build confidence before committing to full deployment. Per-feature administrative controls mean individual capabilities can be enabled one at a time, allowing teams to adopt at their own pace. This is not a workaround. It reflects the reality of how change actually happens in rail: carefully, with evidence at each stage, and with the people doing the work brought along rather than presented with a fait accompli.
COMET Signals
For organisations with large volumes of historical HSEQ data, often captured before a consistent investigation methodology was in place, COMET Signals uses machine learning to process unstructured datasets at scale. It identifies recurring themes, risk clusters, latent root causes, and leading indicators, mapping outputs back to COMET's coded root cause taxonomy so that insights are structured, comparable, and actionable. Network Rail has been at the forefront of proving this concept, using COMET Signals to process large HSEQ datasets. We know what operational value looks like in rail, not just in theory.
Are you ready? A few honest questions to consider
The sector is moving toward AI adoption. The organisations that will get the most from it are the ones that approach it with clarity, honesty, and the right foundations in place. Before taking the next step, it is worth sitting with these questions:
• If your investigations stopped tomorrow, would you know what your data actually tells you about systemic risk across your organisation?
• If a new investigator joined your team next week, would they produce investigation data that is directly comparable to everything captured before them?
• Do you know enough about the quality of your own data today to make a confident AI investment decision tomorrow?
If the answer to any of those is uncertain, that is not a reason to wait. It is a reason to start the conversation now. COMET works with organisations at whatever stage they are at, and we would welcome the chance to help you work through where you stand.
