PWM Ontology 0.1: a cited data model for wealth management and its AI

What was the problem?

Kognita's controls talk about use cases, owners, delegations, claims and evidence. Each control defined those words for itself. Nothing tied them to the business they govern: clients, mandates, suitability and products.

I wanted one shared vocabulary for that. It had to be generic, so any wealth manager could use it, and it had to say where every term comes from: a standard, a regulation or a Kognita roadmap item.

What did I build?

PWM 0.1 lives in docs/ontology/pwm in the Kognita repo. It has 10 modules, covering parties, due diligence, suitability, mandates, accounts, products, the client journey, AI use cases and authority, and evidence. In total: 126 classes, 148 properties and 37 code lists with 188 codes.

Business terms reuse or specialise FIBO where it has them. 37 FIBO terms are specialised or mapped, and each shows whether FIBO marks it Release or Provisional. External terms are declared, not imported, so the file loads on its own.

Every term cites its sources with a locator such as an article, a paragraph, a roadmap item or a slide. There are 456 citations to 46 references. Each citation is marked by source type: standard ontology, ISO, regulation, Kognita, my own frameworks, or a definition I wrote because nothing published fitted.

The rules are SHACL shapes, with 28 written rules plus one generated rule for each coded property. Examples: an AI agent can never hold an accountability role. A live use case needs a materiality rating approved by the control function. A client communication must trace back to a final decision.

The site PR publishes a reference page at /ontology/pwm. Every term has an anchor, so each pwm: IRI opens its entry. The Turtle files are downloadable, and the version IRI redirects to the file so that owl:imports works.

How did I check it?

• The ontology is in the OWL 2 DL profile, and HermiT finds it consistent with no unsatisfiable classes.
• An integrity script checks every term for a label, definition, module and citation. It also checks every FIBO, OMG Commons and DPV IRI against its source files.
• Synthetic data for a fictional bank conforms to the shapes and has an unbroken evidence chain.
• A negative fixture raises exactly the 36 violations it should, covering all 28 written rules.
• Review on the PR found real gaps. One example: a claim marked as a verified fact could pass without being verified. I fixed those over four rounds and added regression tests before merging.

What's still open?

This is a draft. Some regulatory citations are marked instrument-level: they cite the published instrument, and their paragraph numbers still need checking against the official text. Sustainability preferences, orders and execution, and tax reporting are not modelled yet.

The ontology has not yet been used on real data, and Kognita does not yet emit PWM data in its own tests. A roadmap from 0.1 to 1.0 is drafted but not yet merged.

What would I tell another builder?

Write the negative tests at the same time as the rules. A shape that never fails proves nothing, and two of my shapes did nothing until a broken record showed it.

Keep firm-specific detail out of the core. Thresholds, scales and product shelves belong in an extension, so the shared model stays generic.

Sources

Kognita is an open-source alpha. This note comes from Kognita PR #160: Private Wealth Management Ontology 0.1 (draft), PWM Ontology 0.1 in the Kognita repository.