Getting to Production: Six Lessons from FT's Future of Asset Management North America
Appetite for AI in asset management has never been higher. Here's what is actually standing between firms and production.
Enthusiasm isn't the constraint
Two days of sessions at FT's Future of Asset Management North America left little doubt about the industry's mood. Appetite for change is at an all-time high. As one opening session put it, “firms that don't evolve won't survive, and the next five years will decide which asset managers win.”
The more revealing insight wasn't the enthusiasm. It was where firms say they are stuck.
Nobody described a shortage of ideas or a lack of capable technology. Firms already map their day-to-day work to find where agents fit, and agents already take on routine work so people can focus on higher-value tasks. What slows them down is training, regulation, governance and data. Underneath all of it sits one question that came up in nearly every session: will firms trust agents with production data, not just experiments?
That question is the dividing line between AI that stays in the lab and AI that changes how operations run. Here are six lessons from the conference about getting across it:
1. Data is the first wall
Fragmented data came up in almost every session. Information is sprawled across systems, files and people's heads. Lineage, technical debt, inconsistent standards and data that differs by source all add risk. Service providers face the same problem as the managers they serve.
The response across the industry is a push toward a single book of record and one unified data model. Some firms want one vendor and one data model that represent the whole firm.
Fragmentation matters even more for agents than it did for traditional software. Agents need shared context across teams to do multi-step work, and data sharing between teams is often a bottleneck in its own right.
Our view is that firms shouldn't wait for a perfect data estate before they start. Consolidation is a multi-year effort. The better starting point is a process where inputs are predictable and the rules are well understood, so agents can begin earning trust on live data. Every exception an agent investigates also tells you where your data disagrees with itself, so that the data estate can be improved along the way.
2. Governance decides what reaches production
If data is the first wall, governance is the gate. Speakers named it as the main hurdle keeping agents out of production, and they described data governance as the foundation for AI governance.
The standard the industry is converging on is to govern agents like employees. That means traceable data, outputs validated by both people and technology, guardrails that are clear rather than improvised, and an owner accountable for each agent's quality and cost. For breaks and exceptions, the expectation was explicit: AI should be conservative and show its work.
This distinction matters, because it represents the clearest break between agents and the black-box systems of the past. An agent can record its plan, its steps, and its reasoning as it works. The difficulty isn't tracking what an agent did; it's capturing and scoring that record so operations leads, auditors, and regulators can rely on it.
Firms are balancing guardrails against staying at the cutting edge, and too many constraints limit the value agents can deliver. The answer isn't fewer controls. It's controls you can see.
3. Infrastructure matters more than the model
A recurring warning was not to bet the operating model on any single large language model. The model itself is not the advantage. The infrastructure beneath it is. An architecture that doesn't depend on one model makes it easy to switch as models improve and to keep cost under control. A flexible stack also stops a technology choice from becoming calcified and stops legacy systems from limiting future options.
One phrase captured the mood well: boring is better.
In heavily regulated operations, the goal isn't the most impressive demonstration. It's work done correctly every day, on a foundation of reliable data, predictable inputs, and a sound framework. Several sessions stressed choosing tools fit for the stack, the data architecture, compliance requirements, and the business problem. Sometimes deterministic logic with thresholds is all a task needs.
Reconciliation showed how the pieces fit together. Firms described start-of-day recs and deterministic thresholds working alongside agents that triage breaks and recommend fixes, so analysts only see the breaks that are material. Rules handle what rules can handle. Agents take on the investigation that rules can't finish.
4. Measure outcomes, not AI quotas
The firms furthest along don't measure progress by the number of pilots or licenses. They measure business outcomes: funds covered, throughput and cycle time, before and after AI.
Their method is disciplined. They break the work into steps, establish a baseline and build the agent pipeline step by step. They are realistic about difficulty. Multi-person, end-to-end workflows are the hardest to automate and need subject matter experts to design them.
They also judge success by whether what they build can be repeated and shared across teams, as opposed to a one-off. They hold to a simple rule: agents only get credit in production.
5. Humans stay accountable, and they need the skills to do it
No session argued for removing people from the loop. Routine work will move to agents and new AI roles will appear. Accountability for what agents produce stays with people.
One point stood out above the rest. Analysts who rely on AI before they build core skills can't validate its output. That creates a quiet risk for firms that automate too quickly: the people meant to check the work lose the expertise to do it. Success needs both a willingness to use AI and the institutional knowledge to judge its results.
Firms are adapting how they build that knowledge. Training is moving from large classroom sessions to individual learning inside the tool. Adoption tends to take off when senior associates pick up a new tool, and smaller teams often force the change faster than mandates do. People learn to manage agents through exposure and coaching, not slide decks.
6. Technology partners are becoming design partners
Firms are building more in-house, and subject matter experts now iterate directly with technologists, which changes who owns what gets built. Yet the same firms increasingly treat specialist vendors as design partners on governance, compliance, and workflow design, not simply as suppliers.
The question firms ask of partners has changed. It's less "what can your product do?" and more "will you help us get to production, and can you stand behind the results?" A partner earns its place by bringing experience internal teams don't yet have: deep domain knowledge, a track record of running agents in production, and controls that are built in from the start.
What this means for operations leaders
Taken together, the lessons point to a practical agenda:
1: Start where your data already flows. Don't wait for a single book of record to begin. Pick a process with predictable inputs and a clear baseline.
2: Make traceability a precondition. If an agent can't show its work, it shouldn't act on production data.
3: Give every agent an owner. Accountability for quality and cost shouldn't be shared by everyone and owned by no one.
4: Measure before you deploy. Agree on throughput, cycle time, and coverage up front so the return can be proven, not just claimed.
5: Invest in judgment, not just tools. Build the skills your people need to validate what agents produce.
Where OnCorps fits
These lessons closely match how we've built OnCorps. Our agents have run in daily production for years, inside the operations of large and complex asset managers and fund administrators:
• 120+ agents in daily production, covering more than $13 trillion in assets.
• 1,000+ exceptions triaged by agents every day, with every decision logged.
• 250+ pre-built accounting checks for financial reporting review, with a 60%+ reduction in manual review hours for clients.
• ISO 42001 certified, with governance built into the architecture rather than added later.
Our agents investigate breaks, explain root causes step by step, and write to source systems so your teams don't have to. Analysts see what every agent is doing in one place, and people stay in charge of the decisions that need judgment.
The takeaway
The industry has moved past asking whether AI belongs in asset management operations. The question now is how to trust it in production with traceability, transparency, human accountability, and results that can be measured.
Firms that address that question in the next few years will set the pace for the industry. Waiting for perfect data or for the perfect model is the surest way to fall behind. The firms that move now, with controls they can see, will be the ones still standing when the winners are counted.
