Most firms start with AI by automating a few painful workflows without reengineering them. This misses the point. When a process is inherently inefficient, you have to change the way it works before you try to shift the balance between what models and humans do.

In the field of AI, the Turing test says that AI is intelligent if it’s indistinguishable from a human in conversation. For investment management, the better test is the “Junior Analyst Test.” AI proves its effectiveness if its work is indistinguishable from that of a new hire. But to get there, you need to address data, skills, tools, and process redesign in a context-aware manner.

Three ways firms get AI wrong

Teams get process thinking wrong in three consistent ways:

1. They fall into magical thinking. They try to automate judgment-heavy work without tightening the process around it. Corporate actions processing shows how wide the gap can be. The process is often subjective. Notices may have inconsistent language, missing fields, and different nuances for each instrument type. The rules for elections, deadlines, entitlements, and downstream booking often live in precedent and judgment calls.

AI enablement would have to decide how to handle ambiguous or incomplete notices, since that’s how humans do it. It needs the equivalent of the domain context and reusable skills that experienced humans would use to complete tasks. If those elements are missing and there’s no human in the loop, the first signal you get is severe operational fallout. For example, you might see wrong elections, entitlement mismatches, breaks, or P&L impact. Oversight must be designed into the process.

2. They reach for AI when the job is deterministic. The conflict is that AI works probabilistically. When a task has definite rules and always takes one path with fixed steps, transforms, and outputs, dropping a probabilistic layer on top adds variance to something that should never change.

Classic ETL is an obvious case. If you pull vendor XML files, you only need to parse them, transpose fields into a target schema, and run standard validations. You don’t need AI to predict what the next probable string should be. If you can do the work deterministically, you should keep solving it deterministically. But AI would excel at helping you write code and complex queries or at breaking down your steps into a procedure that it can run over and over again.

3. They chase the shiniest strategic project and skip the operational grind. Trying to put the full trade lifecycle on autopilot might look great in a demo, but it fails in the real world. You get better outcomes starting with tedious, repeatable tasks where validation is quick, and the feedback loop is tight. Something like pulling trading comps fits perfectly. An agent could auto-assemble a peer set, recent multiples, and a short “what moved” note using approved data sources, then let humans focus on exceptions. AI delivers value immediately without turning into a moonshot.

Passing the “Junior Analyst Test”

If we want AI to matter inside real workflows, we must set it up to pass the Junior Analyst Test in the first place. Whether new hire or new agent, we start by giving that resource clear instructions, the right data access, and enough context to execute the task the way the team expects. We don’t sit them down and say, “go reconcile the book,” hoping they will figure it out. We give them documentation, we explain the rules of the game, and we show them the decision tree that lives in the team’s head.

Take daily cash and position break triage between your IBOR and a prime broker, for example. A new analyst would need materials like the reconciliation report, the break taxonomy, reference guides for common codes, and a playbook of past breaks that were resolved. They would need to follow a decision tree for break thresholds, resolving stale price issues, diagnosing timing breaks, or finding missing trades. They’d need a breakdown of how to trace it from OMS through execution, allocation, and booking, confirm which system dropped it, and determine whether and how to escalate anomalies. That’s the setup AI needs as well, following clear steps, rules, and branches to do the same triage work a junior analyst does, faster, and route the true exceptions to the right human reviewer.

Building context into workflows

Looking at workflows this way forces you to put processes in three buckets. Some work should be eliminated outright. Some work belongs in deterministic automation because there’s a script, and the next step is always the next step. Then there’s the slice where AI earns its keep: It does the dirty work fast, pulls evidence together from the systems we trust, and outputs an answer that a human can review and sign off on.

However, the more we look at AI-enabled workflows, the more we see that AI is too broad a category. If there’s an assembly line where individual steps do different valuable work, you create AI agents in that assembly line trained to do those individual steps. Every step in the assembly has a dedicated, purposeful agent trained to do one task and do it well. The steps may not map one-to-one to how humans achieve the same outcome, but there are clear handoffs between steps and points at which a human reviews and validates the agents’ collective output.

For example, when an analyst works with front-office traders to resolve a trade break, traders might say, “Handle it the way we did last time.” In a purely human process, the human is doing, deciding, and checking all at the same time, versus breaking it down into, “First I’ll write it, and then I’ll check.” Agents can do the same thing systematically by breaking it down into granular steps. They can read the email, identify who the trader is, identify all trades where we had an issue, and give a concrete answer with choices to validate and select.

It’s a high-pressure task, but it’s repetitive work. There’s judgment involved, and every problem is slightly different. You need context to understand what kind of problem it is, who the experts are, and who you should assign it to. With process redesign, AI can work on data discovery, do the reasoning, and come up with recommendations. That cuts the workload by 80%. The human reviews at the end and remains accountable.

That means data flows through a chain in which each step has a contract. That is where model context protocol (MCP) comes in. Standardizing that contract through MCP captures it in steps that are exposed as inputs and outputs, so downstream agents can use them effectively and know what to provide and in what format when they need to speak to the next step.

The readiness question and talent implications

One early signal of AI readiness is whether you can express what you do in steps and decision trees. What is the decision tree in your head? You make those decisions implicitly because you’re the expert. If you can break it down and write it down, and if it is clear to you, AI can solve that problem. The processes where you can clearly describe what you do and how you do it — not necessarily why you do it — are the ones to start with. The second part is whether your processes are repeatable.

But as you progress, a structural question starts to come up. The roles that will get eliminated the most are these junior roles. You no longer need 100 junior staff because you can do the same with a fraction of that now. But what happens 10 years from now when you don’t have a pool of talent that has built up experience and human judgment to validate the output? That’s a genuine business continuity risk. If you eliminate the junior roles today, where do your next mid-level people come from?

The industry needs to come to terms with this implication. As agentic AI reshapes how work gets done, how do we preserve the pathways that create seasoned judgment? The practical answer, for now, is to be deliberate about the mix: 50 juniors and 50 agents, not 80 agents and 20 juniors. The harder question is what comes after that, and whether the industry can sustain the expertise it needs if the traditional apprenticeship model breaks down entirely.

We can build the context and redesign the processes to make AI work today. But we haven’t yet figured out how to build the people who will make those systems work tomorrow.