Building Agentic AI Workflows: 3 Pillars for Investment Operations
Agentic workflows promise to turn time-consuming operational tasks into a stream of nearly real-time moments. To take one example, much of the cost of reconciling mismatches (transactions, prices, valuations) comes from navigation, filtering, and assembling evidence. Professionals know what to do. When the right data shows up quickly in front of them, they can handle the complex aspects of each task.
But designing a single agent for investment operations, let alone a system of interconnected agents, depends on getting three pillars right: effective controls, accuracy and predictability, and a user experience (UX) that eliminates the need for separate training.
These pillars are essential for agentic AI to deliver on its promise in investment management. Investment managers and the firms that support them should only embrace agentic AI when they have designed and built for control.
Note on terminology:
- An agent completes one or a small number of discrete tasks.
- An agentic workflow orchestrates the activities of agents into a larger result, with or without further user input.
- Agentic AI refers to the overall category of technology.
Controls: making execution safe
Controls prevent agents from completing tasks or seeing data that they shouldn’t. Without them, agents would have unfettered access to your data, your client data, and to actions with millions or billions of dollars in consequences.
Building for control means weaving six main principles into each design pattern.
Build in a sandbox
Agentic workflows belong inside an enterprise sandbox that is designed to meet customer data security expectations. In that sandbox, customer data remains separate from training, and users retain control over how the system operates on live data.
Keep agents specialized
That foundation makes orchestration practical: Instead of expecting one agent to complete an end-to-end process, you coordinate specialist agents across the workflow, each with a specific task.
Emphasize structured responses
The conversational approach to generative AI is an inherent part of its success. Yet chat-based interactions are ill-suited to questions about investment operations data. Without structured control, chat tends to tell a story and keep the conversation going, rather than retrieving the specific data a workflow needs. With structured control and an exact response format, teams can build agents that return an exact response from specific fields in a predictable format.
Create “single tenant” agents
Agents need to be single-tenant, one client at a time. That design avoids cross-client bleed when a user has access to multiple clients. It prevents a scenario where an agent response or action pulls live data from one client while answering a request about another.
Audit everything
In high-stakes workflows, each step matters. Full auditability of what the agent does, step by step, including the decisions it makes and the tools it uses, makes the workflow reviewable and accountable, rather than relying on “because the model said so.” Human accountability is non-negotiable.
Create and enforce guardrails
When each agent has a specific job, that job should stay bounded even when a user tries to push it elsewhere. This layer validates the user’s input and the agent’s response, then matches them to that objective. When there is a mismatch, the system redirects the user toward what the agent can safely do, keeping the agent aligned with its stated purpose.
Predictability: avoiding mistakes and surprises
Predictability and accuracy come from asking the model for a much more focused response and returning it in a specific format, with content that we can further process down the line. The practical bar for live usage is for the model to proceed with the operation reliably. Reliability should be defined as a 0% hallucination rate and at least 90% stated goal completion. In designing for predictability, three patterns are essential.
Build single-specialty agents
Operational workflows already move through stages. A request comes in, the system identifies the intent, gathers the relevant data, checks for exceptions, and then routes the work to the right specialist when deeper expertise is needed. An orchestrated, multi-agent approach mirrors that logic by stitching narrow agents together instead of expecting one generalist agent to handle an end-to-end process.
Enforce consistency
When output is predictable, downstream systems can pass it cleanly to the user to confirm, store, compare, and feed it into the next step. Reliability comes from repeatability, with the same consistent fields and shape and highly focused responses each time.
Accrue long-term memory
Over time, long-term memory captures feedback when the workflow misses or hallucinates, then summarizes that learning and feeds it back into future runs. It also becomes a repository of specific operational workflow nuances that differ from one client to another. The result is a more accurate, more predictable experience that improves with use, without asking users to become expert prompters. It also enriches the prompt, so downstream agents get the right context.
User experience: adaptive learning
It is not practical to expect users to learn how to prompt. If they have to learn the interface, the user experience is flawed. Instead, users type what they mean, then the system enriches that request for downstream specialists.
Offer help in context
A great agentic UX starts with natural intent. Users ask in plain language, then the system enriches the prompt so downstream agents have what they need. When the prompt is not sufficient, it asks for the missing inputs and clarifies before proceeding.
Prevent changes without approvals
The pattern for any write action is to prepare it and show it. When a user asks to modify transaction or trade data, the system does not change the record immediately. It prepares the update, shows exactly what it will do, and then asks for explicit approval, with the SME and the customer in the loop. This keeps data manipulation out of autonomous execution today, and it creates a practical path to more autonomy over time as accuracy improves and the workflow earns trust.
Managing the agentic AI promise
Agentic workflows become valuable when they compress the operational journey. In investment operations, the work often means hopping multiple screens, looking for relevant data, and filtering it. The goal is to reduce this time from several minutes to roughly a minute or less.
To get there, the path forward is execution with control. This entails more focused responses in a specific format, rather than chat, agents with narrowly defined tasks that do not exceed their stated purpose, and a “prepare and show” standard for any data modification. Following this approach, over time, the workflow can move toward increasing autonomy as accuracy improves, from hours to just a few minutes.
Authored By
Vivek Viswanathan
Vivek is a Senior Vice President and Distinguished Engineer at Arcesium. A technology evangelist and researcher, Vivek works to build frameworks and solutions that make the most out of a firm’s resources. He was awarded top Quora Writer in Software Engineering & Computer Programs in 2020.
Share This post
