Building an AI-Native Data Platform

From data to knowledge

Agents now read the same tables as the claims handlers, underwriters, actuaries and controllers they work beside. Those tables hold governed, trusted data. What still separates humans from agents is real-world context: a claims handler knows what a figure means, which table holds it, how a case links to its contract. Organisations have tried to close that gap for many years (Figure 2) through active alignment by central teams. Language and definitions keep changing, so those teams were always fighting an uphill battle to stay current. Generative AI changes that: always-on agents enrich the semantic layer and the ontology from real interactions, and the central team reviews. That puts an old dream within reach: a CEO asks for the cost per claim on a Friday evening and can get the answer independently without pulling someone else in to get the details. For the company, that means lower cost per case, faster decisions and auditable automation.

Two traditions, one knowledge layer

The semantic layer

The semantic layer says what things mean. It levels the playing field, putting the subject-matter expert and every agent into the same context: for example, how last year’s profit figures were calculated or what lies behind the KPIs. Jessica Talisman’s ontology pipeline defines the stepping stones, from controlled vocabulary to thesaurus (Figure 1). Generative AI now pulls these steps together, keeping information up to date for both humans and agents.

The ontology

The ontology adds the business objects and their relations, so agents see invoice, contract and policyholder as one case, as a claims handler does. Its links serve three jobs. Act: check the cover and approve an invoice within the agent’s limit. Detect: flag a provider who bills the same treatment twice. Trace: follow a contract change forward to every open claim it affects, or a rising cost per claim back to the providers behind it.

The knowledge graph

Imagine your city’s public transport: lines connect the stops, and around each stop lies a neighbourhood buzzing with context. Put the semantic layer and the ontology together and you get the same: the knowledge graph. Business objects such as contract and invoice are its stops, their relationships its lines and the definitions its neighbourhoods. An agent starts at a case, reads what its terms mean, follows the links and leaves with an action (Figure 1). In a benchmark by Sequeda, Allemang and Jacob (2023), a language model answered 54% of enterprise questions correctly with a curated and verified knowledge graph, against 16% with the raw tables alone.

In the knowledge graph, the semantic layer says what things mean and the ontology says how they connect, all on a trusted data platform.

Built on the platform you run

The knowledge layer builds on the platform the company already runs. Raw, curated and serving data stay where they are, and both layers read the serving data and meet in one graph (Figure 3). Strict governance allows the ontology controlled write-backs through policy gates into the carbon layer, which the platform ingests like any other source.

How knowledge flows

The knowledge layer stays up to date without a central team, because knowledge enters where people already write it, from documentation and tickets to code and the catalogue, and through the data pipelines. Each data product arrives with drafted definitions. Generative AI drafts the definitions in version control, and a person reviews each one before it enters the graph. Every definition traces back to its source, author and reviewer (Figure 4). This closes the loop from incoming knowledge to governed action by an agent and a recorded outcome.

Knowledge moves in a loop