Investment operations teams burn through easy AI wins quickly. Repetitive tasks such as standard data pulls and routine reporting are well-suited for early experimentation because they meet just two criteria.

What typically happens is that teams tackle the easier tasks first and then hit a wall. The rest is harder: positions data becomes dense, rules vary by instrument, counterparty relationships add another layer. It’s not surprising that McKinsey data show that 88% of organizations report regular AI use, but fewer than 10% achieve scale.ⁱ

To change your operating economics with AI, you need to move toward 100%. That means deploying hundreds of agents, each handling real complexity across instruments, counterparties, and regulatory requirements. Getting there requires tackling the same front-to-back problems that make investment operations challenging.

Why the traditional build cycle can’t get you there

Building operational tools once meant cycling through a loop that starts with an operator identifying a problem. Then a product manager writes requirements, and an engineer builds a solution. Engineers follow software development lifecycle (SDLC) processes to catch errors before they reach production and create an auditable trail of what changed and why.

That loop works when you need five tools. When you need hundreds of agents solving different problems across your portfolio operations, the cycle can’t keep pace. But you still need comparable discipline.

SMEs understand the details of their jobs best. They understand the nuances of corporate actions workflows, break resolution, and the day-to-day complexities of middle- and back-office processing. The scaling path runs through people who actually do the work.

The wins come from operational people who tackle the problems they face every day and solve them directly. Practitioners stop asking how to grind it out the hard way and start asking how to take the faster path instead.

“These days, front-office teams are exerting more influence as portfolio managers and strategy groups push for integrated systems that deliver a unified data model, including performance, accounting, and market data in one central location. That, in turn, is forcing firms to reassess governance and access controls.”ⁱⁱ

But they aren’t developers. What they need is a platform where they can work with an LLM to solve a problem, test it against real data in a local environment, show the evidence of what they’ve built, and submit it for review before anything touches production. Your AI infrastructure team’s job is to give them the right tools so they can build their own agents effectively.

The foundations

Agents come for free if you do the foundational work right. That infrastructure has four components, and each one has to be built with your investment operations in mind.

Your data is the starting point, but LLMs aren’t trained on this contextual goldmine. They don’t know how your positions relate to your instruments, how your instruments relate to your counterparties, or how any of it maps to your regulatory obligations. Structured data with a clear ontology that describes how your entities relate to each other is the foundation on which everything else depends.

The second component is your access layer. Model Context Protocol (MCP) helps by giving LLMs a clean, structured interface to your portfolio data and risk systems. It’s a mistake to throw existing APIs built over 20 years at an LLM and expect it to reason about your positions and counterparties. Investing in defining a layer that’s easy for the model to interpret is crucial.

The third component, skills, is the scaling multiplier. Skills are how you teach an LLM what your investment operations involve, from your concepts and ontology to the subject-matter expertise that lives in your operators’ heads. That means encoding knowledge about the day-to-day details. The LLM won’t figure them out on its own. That’s the difference between an agent getting the answer right 99.9% of the time and getting it right 15%. You build skills once, and they help you build hundreds of agents quickly, with full context on the business, the data, and the domain.

Trust through determinism: solve once, run forever

Scaling agentic AI in investment operations also requires solving a trust problem. We know LLMs can hallucinate and aren’t perfect. Allowing an agent to reason in the background about wire movements with no controls would lead to disaster. Error rates compound quickly, causing friction across your positions, instruments, and counterparty obligations.

To trust an agent in production, you have to know it will produce the same result every time. That’s determinism. You engage the LLM when you’re solving a problem for the first time. You work with it to figure out the right approach, test it against real data, and finally, iterate until the output is right. Then you can have the LLM write a Python script or SQL query that produces that answer deterministically going forward. Once it’s structured code, it always behaves one way.

For example, consider a reporting requirement in your portfolio operations. You might want to extract a dataset, merge it with a second source, check whether a threshold condition is met, and trigger an alarm to a broker if it is. All of this can be done deterministically. You’re enabling users to create a Python script with an LLM quickly, then run it with confidence that the same input produces the same output.

That’s how you earn trust, and human-in-the-loop remains a must-have. Once an agent is about to take consequential action, it should engage a human reviewer before proceeding.

The orchestrator as control center

The foundations need one other essential to hold them together. You can’t just drop an off-the-shelf LLM into an investment management environment. You have to build your own scaffolding around it – that is, an orchestration layer. A purpose-built orchestration layer functions as your control center, bringing tools and skills together to solve the problems your agents face across your investment operations. It’s the difference between a working agent environment and a liability.

That layer carries a specific set of mandates that it has to observe.