Investment organisations have spent years building data warehouses, models and systems to manage increasingly complex portfolios. However, as artificial intelligence becomes increasingly embedded in investment workflows, having data is not necessarily the same as making it useful, the recent 11th Investment Data and Technology Summit heard.
The challenge is particularly acute for investment organisations, where data is spread across internal and external systems and exists in both structured and unstructured forms. Traditional investment data architectures already perform well with structured information such as positions, transactions and cash flows. The difficulty comes when AI needs to understand what those data points mean, how they relate to one another and where to find the relevant information.
Simply put, traditional data models work well for much of the quantitative information used in investments, but become less effective when AI needs to understand relationships, context and information contained in documents. That challenge becomes even greater when unstructured information is included.
An audience poll illustrated the problem. Asked about the biggest barrier to making organisational data genuinely useful to AI, 52 per cent selected data foundations, including quality, lineage, identifiers and metadata. A further 28 per cent selected governance and adoption, including ownership, trust, permissions and staff usage, while 20 per cent chose business meaning – agreeing definitions, relationships and context.
The results point to a familiar conclusion: the foundations still matter.
A further poll that asked where participants would place their next major investment to accelerate the transformation of their investment function found that 56 per cent of attendees would focus on building intelligence: improving data, context and knowledge. 26 per cent would focus on automating workflows: embedding AI and automation and 18 per cent would prefer to invest in talent/operating model: redesigning roles, processes and technology.
This is where semantic layers, ontologies, knowledge graphs and context graphs come into play. Semantic engineering is the process that establishes business concepts and definitions upfront; ontologies describe how those concepts relate; knowledge graphs represent relationships between entities; and context graphs provide the operational context needed for a particular task. Taken together – if done well – it is it is easier for AI systems to interpret and use the underlying information.
It is technically possible to skip this process – connecting a large language model directly to a database can produce useful results – but the model would have to navigate large numbers of tables and relationships to determine how a question should be answered and it would end up costing significantly more in token costs. Including a semantic layer, the conference heard, provides definitions and relationships upfront, potentially improving both the quality and efficiency of the interaction.
From context to action
While semantic engineering as a discipline is hardly new, its application in investment management is more novel. Indeed, the conference heard that one major challenge is simply determining where to start.
When asked what was most likely to derail semantic engineering implementation, 48 per cent of delegates selected adoption and governance, while 41 per cent identified scope – attempting to capture everything or model too much too early. Only 12 per cent selected technology and legacy systems.
The panel cautioned against trying to model an entire organisation from the outset. Instead, it suggested starting with a specific workflow and working backwards to identify the data, relationships and definitions required.
Private markets reporting was cited as a good place to start because it can involve multiple data sources, unstructured information and decisions that ultimately need to be handed back to a human.
Another approach was to start with questions that investment teams repeatedly ask. Panellists said that questions that are seemingly simple can expose significant inconsistencies beneath the surface. For example, questions around ‘revenue’ can appear deceptively simple, but in actual fact can include multiple options and complexities ranging from how it’s classified through to invoice and billing dates. In those cases, a semantic layer that applies certain definitions can help to maintain consistency.
Successful semantic engineering also requires collaboration between domain experts and data teams. Data teams may own the infrastructure and models, but business and domain owners need to determine what the information means and how it should be used.
An evolving discipline
The conference heard that semantic engineering is not a one-off investment. Technology, data sources and AI capabilities will continue to change, meaning organisations need to avoid architectures that are too tightly tied to a particular technology or implementation.
The broader application will also likely evolve from semantic engineering towards context engineering. As AI systems gain access to more information and tools, the need to provide the right context is likely to become increasingly important.
For investment organisations, the opportunity is therefore not simply to make more data available to AI. It is to connect the definitions, relationships and context that allow AI to use that data effectively.
The competitive advantage may not come from having more data than everyone else, but from being able to connect the data already available in ways that both humans and machines can understand.
