Loading...
Pascal AI Labs
July 29, 2026

AI Agents for Institutional Finance: How to Build the Right Infrastructure

Enterprise AI has made remarkable progress over the past few years. Most institutional investors today have access to frontier language models, yet many AI deployments still struggle to move beyond isolated use cases.

 

The reason is not a lack of model capability or a shortage of information. In fact, investment firms already sit on enormous amounts of data: market terminals, licensed research, internal financial models, investment committee notes, emails, meeting transcripts, and years of accumulated institutional knowledge. The challenge is that this information lives across disconnected systems that were designed for human users rather than AI agents.

 

Before an AI system can analyse a company, answer a research question, or automate part of an investment workflow, it first needs reliable access to the firm's knowledge. Building that foundation is an information architecture problem which comes down to three things:

 

  • Signal acquisition: bringing together public, licensed, and proprietary data into a single information layer.
  • Synthesis: connecting those sources so AI agents can reason across them instead of treating every system as an isolated repository.
  • Organisational memory: preserving research, decisions, and institutional knowledge so every investment process benefits from everything the firm has already learned.

 

This article explains each of these layers, why they matter, and how modern enterprise AI platforms are being designed to solve them.

AI Infrastructure Problem 1: Signal Acquisition

Before anyone talks about research processes or AI, there is a simpler question worth sitting with. Does the firm actually have access to the right information?

 

The quality of a fund's analytical output stays permanently bounded by the quality of its inputs. An analyst cannot analyse information the firm never acquired, and no amount of sophistication downstream compensates for a blind spot introduced at the point of collection.

 

Two funds covering the same market and the same companies often arrive at completely different conclusions, and the reason is rarely that one team is smarter. Usually the two teams are looking at different versions of reality. One might be working from filings, earnings calls, and broker research. The other might have all of that plus alternative data, internet signals, credit documentation, proprietary research archives, and years of accumulated institutional knowledge. The gap between those two information surfaces compounds long before either analyst opens a model.

 

Research stacks across the industry still look remarkably similar to each other. A substantial portion of the industry runs on what would have been a world-class setup twenty years ago, generally some variation of a terminal, a handful of broker relationships, research folders, and a collection of spreadsheets. That stack worked well for a long time. The difficulty now is that the volume of information available to investors has expanded enormously while the infrastructure used to capture it has stayed roughly where it was, so plenty of firms operate with a narrower view of the world than they realise.

 

Firms that solve this treat signal acquisition as a set of distinct problems rather than one system that is supposed to do everything. The information generally falls into three buckets. There is the public domain, meaning filings, transcripts, news, internet signals, and sector datasets that exist openly but still have to be discovered and filtered. There are licensed sources, often valuable precisely because they are expensive or operationally awkward to acquire. And there is the firm's own proprietary information, meaning internal research, analyst notes, investment committee discussions, emails, models, and meeting notes, which for a lot of funds ends up being the most valuable dataset they own.

 

Top-performing funds rarely cover every category equally well. What they tend to understand is where their information edge actually comes from. A technology-focused hedge fund may spend aggressively on internet signal and alternative data, while a distressed credit fund invests heavily in covenant monitoring and legal document intelligence, and a long-only investor prioritises research management and institutional memory. The useful exercise is identifying which categories matter most to your process and building exceptional capability there.

AI Infrastructure Problem 2: The Synthesis Gap

Buying good tools across every data category is only the first phase of the build, and a surprising number of firms have already done it. What this often creates is the impression that the data problem is solved, because the subscriptions are in place. Acquiring data and being able to use it are two different problems.

 

The gap shows up most clearly in the daily work of analysts, who end up acting as the interface between systems that were never designed to talk to each other. Someone researching a semiconductor company might pull transcripts from one provider, consensus estimates from a second, broker research from a third, and supply chain data from an alternative data vendor, then compare all of it against internal notes, models, and prior investment discussions. Answering even a straightforward question means assembling the context manually before any actual analysis begins.

 

AI changes this because a system can work across all of those sources at once, connecting information that previously sat in separate places and assembling it into a single answer. Where firms run into trouble is in connecting frontier models directly to a stack that was never built for machines. In those environments you might see something like 50 or 60% of tasks executed extremely well, which feels like the problem has been solved. The remaining 40% is where things break, because the underlying information was structured for human eyes. The model searches transcripts and summarises research successfully, then fails to reach internal knowledge, misses context sitting elsewhere in the organisation, or cannot trace where a conclusion came from.

 

The load-bearing point here is that an agent does not understand charts, dashboards, colours, or carefully designed interfaces. What it understands are entities, relationships, permissions, and the connections between them. Once a firm reorganises its information estate around agentic consumption rather than human consumption, a single query can draw on eight or ten sources at the same time, and the analyst stops being the connective tissue between systems.

AI Infrastructure Problem 3: Organisational Memory

A firm can acquire every category of data and connect it through a common inference layer and still fall behind over a long enough horizon. The most valuable thing an investment firm produces is not the information but the intelligence generated from it, and common practice tends to discard that intelligence the moment a trade is executed or an idea is dropped.

 

Every memo, meeting note, diligence call, rejected thesis, and post-mortem is intellectual output the firm paid to generate. In principle it should compound over decades into a proprietary asset competitors cannot replicate. In practice much of it becomes unreachable, buried in mail threads, shared drives, and folders nobody revisits, or held as unwritten context inside individual analysts. When those analysts leave, a portion of that context leaves with them and the firm ends up rediscovering conclusions it had already reached, effectively paying twice for the same knowledge.

 

This matters more than it used to. Since the 2003 SEC Global Settlement, which required banks to separate research from investment banking, a great deal of experienced analytical talent moved to the buy side and the distinctiveness of external research declined. Firms now lean far more heavily on internally generated knowledge to produce returns, which makes losing it considerably more expensive.

 

Solving this means capturing intellectual output as a byproduct of daily work, structuring it in a way that reflects how the firm actually thinks, connecting it across time and people, and surfacing it back into the workflow at the moment a decision is being made. That is an intelligence infrastructure problem rather than a storage problem, and the hard part is preserving reasoning in a representation that machines can read years later.

 

The architecture that has emerged around this problem is built on four pieces that only work together. The ontology writes the firm's own vocabulary down in a form a machine can use, defining what this specific firm means by a thesis, a catalyst, a macro risk, or a management concern. Those definitions are subjective to the firm, and that is the point. The context graph then connects what the ontology has structured, mapping relationships between entities, decisions, documents, people, and events over time, so that a management meeting links to a thesis, a diligence call links to a risk factor, and that risk factor links to a position and eventually to a post-mortem written years later. Documents record conclusions; the graph records how conclusions relate to each other. The memory layer sits inside the live query path and decides which parts of that history are relevant to the question being asked, retrieving prior research and outcomes by entity and concept rather than by matching similar words. Over time these accumulate into an institutional intelligence layer, a persistent record of how the organisation has read markets across cycles, which every new investment process can draw on.

How the Pieces Sit Together

Implementations vary, but the sequence is usually the same. Information is consolidated into a common fabric through ingestion pipelines and connectors, normalised, tagged, and governed under a single entitlement model. It is then represented through the ontology and the graph, which turns a pile of documents into a connected view of the firm's knowledge. That estate is exposed through a standard access layer, whether via MCP, APIs, or another protocol, so every application and model retrieves context the same way instead of requiring a bespoke integration. Only then does the model enter, retrieving what it needs and reasoning across it.

 

The ordering matters, because models sit at the edge of this architecture while the fabric, the representation layer, and the access layer form the system underneath. Models will keep changing. The information estate, the accumulated reasoning, and the institutional intelligence stay.

 

That ordering is also what keeps the model layer replaceable, which is becoming a practical concern rather than a theoretical one. Recent reporting around institutions restricting the use of certain external AI models in particular jurisdictions is a reminder that frontier models are systems firms do not control. Today the constraint may be geographic availability or internal policy. Tomorrow it could be data residency rules, pricing, export controls, vendor concentration, or a provider changing how its product behaves. The infrastructure underneath therefore has to be built so that when a better model appears, or when access or economics move, one component gets replaced and the rest keeps running. The question worth asking of any AI strategy is whether the firm can keep absorbing better intelligence as it arrives, regardless of which model it uses today.

Where Pascal AI Fits

This architecture is what we build at Pascal AI Labs. Our work sits in the middle of the stack, between the data sources a firm already pays for and the agents doing work on top of them. We ingest internal systems, licensed feeds, and public sources, and we transform them into a context graph that agents can query directly, with the ontology, memory, and entitlement model built in rather than bolted on afterwards.

 

A few things about how we deploy it. The system is interoperable, meaning you can reach it through our interface, through Claude or GPT, or through your own internal platform. It runs in your VPC, and where a client's constraint requires it, on-premise on hardware in their own data room using open-source models we deploy and manage. And every agentic interaction is instrumented, so feedback flows back into the context graph and the system improves the way a good analyst improves with feedback.

 

Engagements usually start with one workflow rather than a platform rollout. A team picks a process that eats hours every week, we get that working end to end, and expansion tends to come from the client spotting the next manual process themselves.

 

If you would like us to review your data and tool stack, or to walk through how this would apply inside your organisation, we are happy to have that conversation.

 

Book a demo

Frequently Asked Questions

1. What is a context graph in financial services?

A context graph is a machine-readable representation of an investment firm's knowledge as a network of connected entities rather than a collection of files. It maps how companies, theses, decisions, documents, people, and events relate to each other over time, so a management meeting can be linked to a thesis, a risk factor to a position, and that position to a post-mortem written years later. Pascal AI Labs builds this layer for capital markets firms, sitting between existing data sources and the agents working on top of them.

2. Why do AI agents fail on enterprise data?

AI agents fail on enterprise data because that data was structured for human eyes. Terminals, dashboards, research folders, and document repositories are designed to land information in a human brain quickly, while an agent reads entities, relationships, and permissions. When a frontier model is pointed directly at human-formatted systems, something like 50 or 60% of tasks execute extremely well and the remaining 40% break in ways that are hard to trace, usually by missing internal knowledge or losing the audit trail behind a conclusion.

3. What is the difference between a context graph and RAG or a data catalogue?

Retrieval-augmented generation matches a query against chunks of text, and a data catalogue records what data exists and who owns it. A context graph adds the relationships, so retrieval happens by entity and concept rather than by textual similarity, and the system can trace how one conclusion connects to another across years of research. The practical difference shows up in questions that require connecting several sources rather than finding one passage.

USA

228 Park Ave S PMB 55152, New York

UK

86-90 Paul Street, EC2A 4NE, London

India

Indiqube Edge, ORR, Bangalore

Copyright © 2026 | All Rights Reserved

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.