Reasoning Agents vs Deterministic Processes in Collateral Management

We’re currently researching the use of AI in Middle Office and Operations functions within capital markets and investment management.

This is where decisions over using Reasoning Agents vs Deterministic Processes get really interesting.

In layman’s terms, let’s say:

🔹 Deterministic Process > follows fixed rules and always gives the same result.

🔹 LLM Agent > uses probabilistic language models to reason.

🔹 Rule-Constrained Agent > reasons probabilistically but is tightly bound by rules and validation checks.

Then, let’s take collateral management for example – hypothetically we might conclude:

➤ Collateral Requirement Determination = Deterministic Process

➤ Margin Call Notice Generation = Rule-Constrained Agent

➤ Call Issuance = Deterministic Process

➤ Counterparty Response Evaluation = Rule-Constrained Agent

➤ Dispute Investigation = Rule-Constrained Agent

➤ Collateral Allocation & Optimisation = Deterministic Process + LLM Agent Assistant

➤ Settlement Instructions = Deterministic Process

➤ Settlement Monitoring = Rule-Constrained Agent

➤ Internal Report Generation = LLM Agent

This complexity plays out across the breadth of Middle Office and Operations processes.

Postscript: In response to our initial post on this topic, we were asked how we see RAG fitting into this. In our (not necessarily scientific) schema, we categorise retrieval as a capability that an agent uses (rather than an alternative to an agent). i.e. a capability that pulls relevant docs/data at run time and feeds them to the agent to ground its outputs. This retrieval capability could be utilised by either an LLM driven agent or a Rules-Constrained agent, but possibly more commonly in the latter in the scenarios we’ve been looking at here.

We’ll be producing further insights on this topic in the coming weeks, so follow us to stay in touch.