Before You Scale AI, Fix the Data

Key insights from the “What Does a Strong Data & AI Strategy Look Like for Hedge Funds and Asset Managers in 2026?” breakfast briefing panel discussion

“AI is, at heart, a ‘garbage in, garbage out’ engine. If the underlying data isn’t clean and trusted, no model will save you.”

That line landed early in the room and didn’t really leave. It came from a buy-side operations leader, and it did what good observations do: it gave people permission to say out loud what many were already thinking privately.

The breakfast briefing brought together COOs, Heads of Operations, technologists, and portfolio-facing executives from hedge funds and asset managers. The tension in the room was familiar. On one side, attendees expressed real pressure to show progress on AI. On the other, the daily reality of legacy systems, fragmented data, and manual workarounds have quietly accumulated for years.

The panel included product and platform leaders from CWAN/Enfusion, a Deputy COO from Broad Reach Investment Management, a Principal Architect for AI/ML Financial Services at Snowflake, and a head of quant analysis and engineering. Rather than debating where AI is headed in five years, they focused on what has to change in operating models, architectures, and ownership for a data and AI strategy to hold up under scrutiny.

Four themes came through clearly:

1. Your AI strategy is only as strong as your data foundation

Picture this: a portfolio manager asks for yesterday’s P&L. Three people give three slightly different numbers. Nobody is wrong, exactly. They’re just pulling from different sources. Now imagine layering an AI model on top of that.

This is the problem many firms are living with, and it’s the one the panel kept returning to. Data strategy, for too many organisations, has become shorthand for platform choices: which cloud, which warehouse, how fast, how cheap. Those decisions matter. But they are not the strategy.

What operations leaders actually need is a layered, investment-grade data stack:

• Clear, agreed golden sources for P&L, risk, performance, positions, and client exposure
• An architecture where raw data is refined into a trusted, auditable layer that teams can consistently rely on
• A small set of core data products, typically around 10 to 12, that PMs, risk, and client teams use every day rather than each team maintaining their own version

Most firms have grown tactically: a spreadsheet added here, a new vendor feed there, a workaround that became permanent after the person who built it left. Different desks may still treat different vendors as their preferred source of truth, without anyone having formally decided that.

The result is that AI, when introduced into this environment, simply automates inconsistency.

The question worth asking next: Can your organisation name the trusted source for each key metric and trace it back through trade and reference data? If your answer is “it depends who you ask,” that is the starting point.

2. The conversation has moved on, and so should the business case

A few years ago, the question firms were asking was: “We need AI. What should we do with it?” The panel described a noticeably different conversation happening now.

Firms are no longer debating whether to use AI. The question has shifted to how to deploy it without creating unnecessary complexity, and where it actually pays.

Open-ended pilots with copilots and chatbots are giving way to specific, measurable use cases. If there is no clear value metric attached, many initiatives are not moving forward.

Some of the concrete early wins the panel described:

• Using AI to synthesise research inputs such as earnings calls, sector commentary, and frontier-market documentation, so analysts spend more time on judgement and less on first-pass reading
• Compressing spreadsheet-heavy analysis into minutes by allowing decision makers to ask questions directly of trusted data
• Giving a data-hungry CFO controlled self-service access to validated data ahead of board meetings, meaningfully reducing ad-hoc requests to reporting teams

What makes these examples useful is that they started from a problem, not a tool. And increasingly, firms are quantifying the gains: hours saved in reporting cycles, reductions in manual reconciliation effort, time-to-decision for investment ideas.

The question worth asking next: For each AI initiative currently underway or being considered, is there a defined baseline and a target? Without a defined baseline and a target, an initiative is unlikely to secure sustained sponsorship.

3. If the front door is broken, people will use the back window

Here is a scenario that will sound familiar to many operations leaders. An IR team has a due diligence meeting in two days. They need a clean exposure summary. The official system has most of what they need, but it takes three exports, a VLOOKUP, and a phone call to reconcile. So they build their own version in a spreadsheet. They’ve done it so many times the file has a name.

This is what the panel meant when they talked about the “front door” problem. Data democratisation sounds like a technology goal. In practice, it is a user experience problem. If the people who need data cannot access it in a usable form, they will find another way. And that other way usually involves uncontrolled logic, version confusion, and a reconciliation problem waiting to surface at the worst moment.

The panel pointed to examples that are mundane but telling: unencumbered cash numbers that differ between calls and reports; IR teams manually stitching together exposures before every investor meeting; reports that require manual checks of totals before they can be sent.

Getting the front door right means mapping who needs to see what, when, and in what form. Dashboards, decks, IR portals, IC packs, and notebooks are all different surfaces. Non-technical teams need self-service access to reconciled data without needing to write SQL to get it.

The question worth asking next: Are the user journeys for PMs, IR, reporting, compliance, and the board explicitly documented? If those journeys are not feeding into data model and governance decisions, the official data platform and the parallel universe of workarounds will continue to grow side by side.

4. New tools, same governance problem

The panel raised a concern that is easy to overlook when the focus is on moving quickly: agentic AI, without discipline, could recreate the spreadsheet problem at a much larger scale.

The logic is straightforward. If every desk can spin up its own AI agents or automated workflows, firms will quickly end up with dozens of lightly governed micro-processes, each with different data access, different logic, and no clear owner. The opacity that made spreadsheets a governance problem does not go away with AI. It gets dressed in new clothes.

A few distinctions the panel drew that are worth holding onto:

• Deterministic vs probabilistic tasks. Generative AI is powerful where a range of answers is acceptable and speed matters (research, scenario brainstorming). It is poorly suited to official P&L, trade matching, or risk numbers where “approximately right” is not good enough.
• Agent sprawl. When workflows multiply without central oversight, no one has a complete picture of what is running, what data it is touching, or what happens when something breaks.
• Multi-vendor reality. Data is already scattered across clouds, vendor platforms, and internal systems. A realistic architecture accepts that and designs for it: leave data at source where it makes sense, use standard connectors to query across systems, and design with an explicit exit strategy for major vendor relationships.

The panel’s recommendation was to treat agents and AI workflows the way firms treat any important application: with owners, change controls, and monitoring. Central teams in risk, operations, and data should own the authoritative agents and expose them through well-defined interfaces. The closer a workflow gets to official numbers, the tighter the standards should be.

The question worth asking next: Who is allowed to create agents or AI workflows, on which data, and under what controls? If the answer is unclear, that is worth resolving before the number of workflows outpaces the ability to govern them.

Turning insights into action: a short diagnostic for operations leaders

These six questions are a practical way to test whether a data and AI strategy is genuinely grounded or still largely aspirational:

1. Sources and lineage. Can you name the golden source for each key metric and trace it back through trades, prices, and reference data?
2. Core data products. Have you defined a small set of investment-grade data products that teams consistently use, or is every report still a custom build?
3. Use-case clarity. For each AI initiative, can you articulate the problem, the current baseline, and how you will know if it worked?
4. Front-door design. Do PMs, IR, and reporting teams have a coherent, self-service way to access trusted data, or are they still rebuilding it offline?
5. Agent governance. Who can create agents or AI workflows, on which data, and under what controls?
6. Vendor posture. If you had to replace a major platform or data provider, do you understand what that would involve technically and operationally?

What’s at Stake

Done well, a data and AI strategy becomes a quiet competitive advantage. COOs can show boards and regulators exactly where numbers come from. PMs and analysts spend more time on ideas and less on reconciliation. Reporting cycles compress. AI augments rather than undermines control.

Done poorly, AI becomes another layer on top of the same brittle foundations: more tools, more opaque logic, more inconsistent numbers. Audits, investor due diligence, and regulatory reviews become harder and more costly. And someone, eventually, has to explain why the numbers still don’t match up.

Which path a firm takes depends less on its choice of model or cloud, and more on the decisions operations leaders make now about what to standardise, what to measure, what to govern, and where to say “not yet.”

Those decisions determine whether AI becomes a genuine superpower or just the latest expensive layer on a problem that was never properly addressed.

If any of these themes connect to challenges you’re working through, we’d welcome the conversation.