Testing GenAI in Asset Management: Automating Compliance Review
One of the more useful ways to test GenAI in asset management is to put it against a real operational process.
At Optimus Consulting Partners, we have been looking at practical applications of GenAI for asset management firms. One example is compliance review: a process that is often document-heavy, judgement-based and time-consuming, but where the output still needs to be controlled and explainable.
The Question
Could we take complex regulatory disclosures and check them, clause by clause, against a firm's internal sustainability guidelines?
That sounds simple enough. In practice, it quickly becomes more nuanced. The documents are long, the wording is not always consistent, and the issue is not just whether two pieces of text look similar. The question is whether the disclosure intent is aligned with the relevant rule, policy or guideline.
What We Built
We built a modular GenAI proof-of-concept that:
- Extracts legal and regulatory text into structured clause objects, using LLM output in JSON format
- Parses internal guidelines in the same way, so that both sides of the review can be compared consistently
- Uses a staged LLM review process to match disclosure clauses to relevant rules and guidelines
- Produces a draft compliance review report, highlighting potential issues, supporting rationale and areas requiring human review
What We Learned
- Document parsing has improved materially, but it still needs careful configuration. Complex fund, legal and regulatory documents rarely behave like clean input data.
- Simple matching is not enough. Naive similarity scoring can over-apply rules or miss important context. A second reasoning step is often needed to test whether the match is actually meaningful.
- Human review remains essential. The useful output is not an automated pass/fail decision, but a better-prepared review pack that helps a compliance or legal reviewer focus on the right issues.
- The value is in workflow reduction, not magic. Used properly, GenAI can remove a lot of repetitive manual checking, but it still needs controls, evidence and clear ownership.
The Economics Are Shifting
There is still a cost/quality/speed trade-off. Higher-end models generally perform better on structured extraction and nuanced review, but they are also more expensive to run. For a proof-of-concept this may not matter much. For production use, it does.
That said, model costs continue to fall, and hybrid approaches are becoming more attractive: use cheaper models for simpler extraction and triage, and reserve stronger models for the judgement-heavy steps.
Technology Used
- MistralAI and OpenAI for structured output, parsing and review logic
- LlamaIndex as the orchestration and document workflow layer
- FastAPI backend and React frontend for the application layer
Where This Applies
The same pattern is relevant beyond sustainability disclosures. It can be applied to fund documentation, legal agreements, investment mandates, ESG policies, operating procedures and other areas where written obligations need to be checked against internal rules or standards.
For me, the main lesson is that GenAI is most useful where it is embedded into a controlled process. The goal is not to replace judgement. It is to reduce the amount of low-value manual review, improve consistency, and give the human reviewer a better starting point.
If you are looking at practical GenAI use cases in asset management, compliance review is a good place to start: specific enough to test properly, but broad enough to show where the technology can add real operational value.