Last month, we ran a hands-on workshop with leaders from the fund management world around one simple question:

Where can AI agents be deployed inside a fund manager’s operating model, in a way that is actually useful, and still controllable?

We could have done the usual flipchart exercise. We did not. We used LEGO minifigures instead. Not because it’s fun (it is), but because it forces people to think in the language that everyone already understands: roles, handoffs, controls, and ownership.

Fun fact: there are apparently more LEGO minifigures in the world than people.

Building our agents (LEGO style)

The biggest mistake people make is starting with the technology. The right starting point is much more boring (and much more effective):
What are the repeatable jobs in the business that are currently done by humans, mostly by reading text, interpreting it, deciding something, and then doing something?

That is why we deliberately frame this as jobs-to-be-done, not “role replacement”. If someone tries to replace “Trader” as a role, they’ve already missed the point.

Using our LEGO figures and building blocks, we used a simple model to determine the key candidates of jobs to be performed by agents. We asked the group what processes are carried out that follow predictable paths of:

• Information input (usually unstructured, usually text)
• Interpretation (what does it mean and what matters)
• Decision (what should happen next)
• Execution (do something, or prepare an action for approval)

Once you organise the potential agent processes through that lens, it becomes much easier to spot where an agent can help and where controls are non-negotiable.

A super agent vs multiple agents

One of the best parts of the session is also the simplest: If you cannot explain the task clearly on an A4 sheet of paper, it is not a single-agent job.

If the process is not clearly understood, consistently executed, and explainable to someone brand new to your business, you will struggle to automate it effectively. That’s not how people-based organisations work, and it’s usually not how agent-based ones should work either. We discussed splitting work into multiple agents, for example:

• A validation agent to check legitimacy and relevance first
• An interpretation agent to summarise and enrich inputs
• A decision agent to route to the right next step

This is particularly relevant in fund management because it aligns with segregation of duties and governance expectations, and also means more LEGO was needed to build out what may have seemed simple to start off with.

Agent governance

AI agents are not deterministic. They can look “good enough” most of the time, and still be unacceptable when the edge cases bite. In the workshop, we used the rule “assume the agent could be right 95% of the time”, which introduced another agent into the mix: the manager agent, in our LEGO world, a police officer. Think of it as the orchestrator that sits above the specialist agents. It does not do the work itself. It coordinates the workflow, checks confidence and context, and decides when to stop the chain and escalate.

Finally, we determined where humans sit in the loop. Not everywhere. Not nowhere. The manager agent should be able to route outcomes into three paths:

• Auto-execute (low risk, high confidence, repeatable)
• Human approval required (material impact, ambiguity, policy or regulatory sensitivity)
• Exception handling (insufficient data, conflicting signals, low confidence)

The key point is that you can overdo oversight. If your human-in-the-loop step takes the same time as doing the work manually, the automation is pointless. We explicitly called this out in the run-through: the human validation cannot take as long as the underlying task, otherwise you have just created a very expensive process.

Our LEGO agents

At the end of our session, our LEGO agents were complete, and the group largely designed:

• A corporate actions agent
• An investor reporting and responses to DDQs/RFPs agents
• Risk and Compliance monitoring (flagging, reporting, escalation) agents

Build your own agents

Our workshop, whilst fun, is a starting point and framework to help guide you through a design process that identifies key areas in your business that AI agents could significantly optimise, without you needing to be an AI engineer, that Lanware can then implement, with our AI Engineers. Are you ready to start building AI agents in your business (with a little help from LEGO), get in touch.