Ranjith Raghunath

AI agents create a strange tug-of-war between security and uncertainty. We try to control non-deterministic systems using deterministic rules.

Every AI roadmap I see assumes one thing: Data readiness will be quick. In investment management, that assumption rarely holds. You may have automation in place already, but when AI starts generating transformations or classifications, the tolerance for ambiguity drops fast. If definitions aren’t versioned, validation isn’t enforced, and lineage isn’t automatic, you may have a real mess on your hands. Before layering AI on top of your estate, ask a simple question: “Can we explain today’s outputs clearly and consistently?” If the answer is unclear, AI will magnify that uncertainty. If the foundation is disciplined, AI becomes a powerful multiplier.

Everyone wants AI and data products, yet almost no one has the foundation to support them. This is where most data strategies quietly break down.

Governance tends to sound abstract until something shifts unexpectedly. A definition changes, an exposure metric drifts, or a tolerance boundary moves quietly. In asset management, small shifts like these can compound quickly. But commitments to versioned definitions, conversion metric logic, and clear ownership aren’t overhead. They’re preconditions for scalable automation. Without them, AI-enhanced systems behave unpredictably. With them, platforms remain stable even as complexity grows. Governance is what makes controlled innovation possible.

AI systems don’t just reflect the data they learn from, they amplify it. If the data isn’t clean or governed, the problems scale with the system.

Time-to-first-analysis is more revealing than most KPIs. If analysts spend days reconciling data before modeling, something upstream is unstable. The goal is to reduce preparation drag. A faster dashboard can’t do that by itself. Set up a clear, repeatable chain to ingest, standardize, persist, and audit. When that path is predictable, analysts regain time to evaluate ideas instead of repairing inputs. That shift compounds across portfolios. Speed in investment management is about removing friction wherever possible.

Maybe the problem isn’t the agents… it’s the environment we put them in. Systems built for humans may not be the right architecture for AI.

We’re rushing to build AI agents without teaching them how to think. Annotations and grounding are the care and feeding we’re skipping, and it matters more than people realize.

Scaling data operations usually fails at 1.5x growth, not 10x. A new fund launches. A new vendor sends files. A new reporting requirement appears. If ingestion depends on bespoke logic, every addition increases cognitive load. Reusable patterns change that equation. When new sources follow known contracts, onboarding becomes configuration instead of rework. Scale tests discipline long before it tests infrastructure.

Humans solve problems by identifying the type of problem, applying known rules, forming a plan, and executing it. AI agents may need the same structure, with a human checkpoint before action.

The work becomes boring in the best possible way. And the more vendors and investment vehicles you support, the more this discipline matters. Scale will always test your systems. Unchecked variety is what eventually breaks them.

Standardized patterns change that dynamic. When ingestion and transformation are treated as reusable patterns, new sources stop being special projects and start looking like configuration.

Variety is the hidden cost driver in most data platforms. It rarely shows up all at once. It sneaks in through “just this one” integrations. A new vendor. A slightly different format. A special exception someone promises to clean up later.

Risk conversations tend to spiral when answers feel subjective: “Why did this number change?” “Is this still within tolerance?” “What’s different from last week?” Without lineage and telemetry, those questions trigger debate instead of diagnosis. People rely on memory, intuition, or authority instead of objective data. Measured systems change the tone immediately. You can show what moved, where it moved, and why it moved. At that point, explanations replace opinions. And that creates the kind of room for productive conversations that another dashboard never could.

Ownership breaks down fastest when systems are noisy. When pipelines behave unpredictably, even strong teams slip into firefighting mode. Responsibility becomes reactive, and decisions get deferred. Clear systems restore ownership by making outcomes traceable. When inputs, transformations, and outputs are observable, accountability stops feeling personal and becomes structural. Good systems make ownership sustainable and scalable.

In investment management, speed rarely breaks because analysis is slow. It breaks because preparation is unpredictable. When analysts don’t know how long it will take to get data into a usable shape, planning becomes guesswork, deadlines feel tighter than they should, and confidence takes a hit before analysis even starts. Predictable ingestion and standardization change that dynamic. When every source follows a known path, timelines stabilize. Analysts stop buffering their schedules “just in case.” This kind of stable predictability creates room for ambition.

In buy-side environments, analysts are rarely short on ideas. They are typically short on time to test them. When sourcing and preparation consume most of the day, creativity becomes theoretical. Ideas stay untested because the path to validation is too slow. Infrastructure that shortens that path changes behavior immediately. Analysts experiment more, iterations increase, and learning accelerates. Alpha doesn’t just come from insight, but from how quickly that insight can be validated.

One of the hardest questions asset managers face right now is deceptively simple: Can I trust this pipeline without constantly watching it? Most teams are already automating ingestion and transformation. The real differentiator shows up after that. What happens when inputs change? When a vendor schema shifts? When AI generates logic faster than teams can review it? Hands-off automation only works when quality gates, validation rules, and lineage are built into the system itself. Otherwise, automation just moves risk out of sight. The moment you can step away and still trust the output is the moment automation starts compounding value instead of anxiety.

They turn “are we okay?” into something you can actually answer. That’s what good telemetry does. It builds confidence for leadership long before it ever prevents an incident, and that confidence is what enables better decisions under pressure.

Showcasing clear boundaries and exceptions help take the risk from being opaque to understandable.  Risk always feels bigger when it’s opaque. Measured systems replace guesswork with evidence.

Speaking to my senior technical leaders ( CIOs, CTOs, Sr. Managers, and Risk leaders ) here… A way to make room for collaboration is through “right sized” visibility.  When we are asked: