From Automating Breaks to Predictive Ops: The Next Horizon for Investment Operations

Investment operations have made real strides in automating how breaks are handled. AI-enabled operations are already climbing an important ridge line toward agentic capabilities. The throughput improvements from automation have been significant, and agentic agents are gaining traction by drafting operational narratives, assisting teams under time pressure with exception queues, and supporting specific tasks.

It can feel like the industry is nearing the summit, but anyone who has climbed in the mountains knows the illusion. Predictive operations are the next ridge. They demand the ability to see the full storyline of an event. Many firms aspire to that vantage point. Few have the data foundation to reach it. McKinsey has pointed out that the biggest constraint on AI is lack of high-quality, well-governed data and end-to-end visibility into processes.ⁱ

But predictive AI is qualitatively different from agentic. Predictive needs to be able to see the storyline of a particular event, from beginning to end. If AI can continuously observe this, then it can provide early signals of potential issues that may arise in the future. Otherwise, it's challenging to see predictive ops adding strong and safe value if the underlying data remain as discrete signals coming from siloed modules or applications.

What predictive ops will require

The idea of a storyline is the ability to see a particular event, like a trade break, from beginning to end. Each operational event must be continuously observed as it progresses from one frame to the next. State transitions must be preserved across systems and time.

All of that information from any given point in time becomes input that a predictive AI application would need to predict the likely impact of future occurrences. But without those inputs, AI would stumble, and predictions would lose accuracy and relevance. Trading‑desk research, for example, highlights the impact of addressing data gaps for prediction in a highly predictive workflow. The research described poor data and fragmented infrastructure as major blockers to optimize algorithms and venue selection in real time, two highly predictive areas.ⁱⁱ

This shortfall affects more traditional workflows, as well. Most platforms lack such lineage. There may be timestamps that show at what point in time data moved from one place to another, and what was captured, but that trail does not offer the same quality as a first-class, transparent data view. Full lineage means tracking where and when data is moved, transformed, processed, and passed to the next system.

Knowledge graphs, which link data fields and flows to business processes and system dependencies, play an essential role. They provide structured representations that map how data, systems, and business logic interconnect, capturing an understanding of what moved when and how everything relates.

Beyond lineage and knowledge, predictive systems need to learn from historical events, as well. For example, to predict breaks, it would need to analyze past breaks to identify patterns where things break downstream. These causal patterns include a range of inferences. When a certain configuration appears, a particular failure typically follows.

And beyond pattern libraries, there will be other kinds of breaks that happen but don't follow the pattern shown to the agent earlier. Addressing these calls for reinforced learning throughout the entire process and adapting as new failure modes emerge. This sophistication is a very steep climb from what we have today.

Why most stacks are not predictive-ready

Current investment management technologies create many of the barriers to the development of future predictive AI. According to the 2025 Data Integrity Trends Report, 64% of organizations cite data quality as their top challenge to AI adoption, and 77% rate their current data quality as average or worse.ⁱⁱⁱ

Even if there is integration all the way from OMS to the financial statements, there are still siloed modules. Individual systems may work well within their boundaries, but module-to-module processing happens without shared context. Having the entire system connected from front to back would be one way to climb higher.

These problems affect hedge funds and institutional managers, but they affect other financial firms, too. Think about investment banks built from multiple mergers and acquisitions, resulting in many disparate systems. Systems tend to be completely siloed. Simply keeping them in sync with each other is already a big lift. Legacy platforms resist the kind of instrumentation that lineage requires.

Even when some lineage exists, it's often incomplete, missing the modeling of transformations as data flows through the stack. Discrete signals lose their causal relationships. Backdated changes, such as modifying the way certain trades are tagged to a particular strategy, create breaks that are difficult to trace without full lineage visibility.

In other words, the predictive power of AI is tough to leverage. It will either take significant effort to connect platforms or a “black swan” advance in AI that has not yet been developed. Over time, as more systems have AI embedded in them from day one, predictive power could accelerate, but traditional, legacy, and even today’s “modern” applications and systems would always struggle to catch up.

Economic realities compound the technical challenges. All this potential would come at a cost of burning tokens and retrofitting legacy systems to be reachable by AI. Theoretically, you could extract predictive signals from decades-old data, but it would cost a lot of money for debatable value. Without foundational cleanup, the ROI of using AI to try to overcome weak or absent lineage remains unclear.

Pathways to building predictive-ready operations

Despite the challenges, there are routes up the mountain. Firms that begin now can start to make incremental progress towards the predictive ridge line, while others remain stuck at basecamp. The work breaks into three parallel tracks.

Engineering continuous operational stories. The foundation is instrumentation. Systems need to capture the end-to-end state as a norm, not an exception, in order to predict impact and likelihood. This means:

• Re-architecting connectors to reduce siloed flows and create true visibility across modules
• Instrumenting each hop from order capture through settlement to reporting so systems can observe the complete flow
• Tracing end-to-end state so that a system can capture what's happening at any given point in time

Teaching the system from historical breaks. Predictive AI must learn from past instances where things break downstream. Then, it needs to watch and detect issues upstream based on that historical pattern recognition. Consider pattern libraries built from real exceptions intended to catch future issues. For example:

• When a new executing broker is used for trading but not set up in time across various systems and with the prime broker, trades could fail and create settlement issues.
• Complex corporate actions (like stock splits, cash dividends, and spin-offs going ex on the same day) typically result in issues because the priority in which these events need to be executed can be interpreted differently by market participants.
• When counterparties misinterpret data, breaks occur in cash reconciliations, position reconciliations, and other downstream processes.

Targeting early predictive use cases. Not every operational process needs predictive capabilities on day one. Focus on areas with clear value:

• With short sales, have they located borrowers? Are those hard-to-borrow cases? What is the cost associated with it? The value would lie in predicting whether there is a likelihood of this being a difficult security to borrow if it is being short-sold.
• When corporate actions have multiple election possibilities, what is the optimal election for investors in multi-option voluntary corporate actions based on price of the relevant securities? The ability to reason from precedent would help make the choice more confidently.
• With tax and regulatory triggers like wash sales, can we identify and mitigate what happens (i.e., preventing loss disallowance) when the wash sales are triggered? Detecting similar security repurchases within regulatory timeframes before order release could support a more reliable decision.

The leadership agenda for moving beyond automation

We are all on this journey together. Predictive will come as the next step once we have solved the agentic challenge, but the path requires strategic choices now. Which operational storylines must become fully traceable? What budget and risk appetite should executives allocate to predictive design? When does it make sense to shift from legacy platforms to AI-ready architectures?

Any organization that can reach those heights could have an edge in predicting operational issues before they materialize. Yet, there is healthy skepticism around achieving agentic benefits near term, and the industry is still discovering how best to utilize it to generate ROI.

The timeline remains uncertain because innovation is happening at wild speeds. AI improves dramatically in months, not years. What seems two years away today may look achievable in a year, just months from now. The ridge line is closer than it appears, but only for companies already climbing towards it.