Financial institutions have spent considerable time exploring the potential of AI. Where do you think the industry currently stands in moving from experimentation to meaningful implementation?
The industry is reaching an important transition point. The discussion is moving away from whether AI has a role to play in investment operations and towards how firms can deploy it successfully at scale.
Many organisations have already demonstrated compelling individual use cases through pilots. The greater challenge is integrating those capabilities into core operational processes. There is a significant difference between an AI model performing well against a curated dataset and that model operating reliably against full transaction volumes, live accounting workflows, and established control requirements. This is particularly important in investment accounting because many of the most valuable applications of AI are inherently time-sensitive. Anomaly detection, reconciliation, and exception management become considerably more valuable when they can influence a process as it happens rather than analyse the outcome the following day.
Moving AI into production therefore requires firms to look beyond individual models or applications and consider the operating architecture supporting them.
Why does the underlying accounting architecture matter so much when implementing AI within investment accounting?
AI can only operate effectively against the data and processes available to it. Many investment accounting platforms were designed around batch processing, with information collected during the day, processed overnight and, reported the following morning.
That architecture creates an inherent limitation. If the accounting state is only available after processing has completed, AI is largely restricted to retrospective analysis. A model may identify an anomaly, for example, but by that point the issue may already have affected reconciliations, valuations, or downstream reporting.
Asset-class silos create another challenge. Equities, fixed income, derivatives, and private assets may reside in separate systems or subledgers. This fragments the context available to AI and means that every additional book, asset class, or jurisdiction can create another integration requirement.
A real-time, event-driven architecture changes that relationship. Trades, price updates, cash movements, and corporate actions can be processed as individual events allowing validation, accounting, and other workflows to take place as information enters the system.
What does embedding AI within the accounting engine enable firms to do differently?
The principal difference is that AI can become part of the operational process rather than an analytical layer sitting above it.
One example is the control environment. Traditional controls tend to examine predefined data points and identify exceptions that subsequently require investigation. AI can broaden those controls by considering historical patterns, peer funds, securities, market conditions, and related attributes simultaneously. That allows firms to identify inconsistencies that deterministic controls may not detect and, importantly, surface them before they propagate through downstream processes.
The same principle applies to exception management. Rather than presenting accounting teams with every reconciliation break or tolerance failure, AI can help distinguish between historically benign exceptions and those more likely to indicate a genuine problem.
For exceptions requiring investigation, it can also provide context around the likely cause — such as a late corporate action, missing FX rate, or incorrectly booked trade — before the accountant begins reviewing it.
Could this eventually allow firms to automate the resolution of accounting exceptions as well as identify them?
Yes, but this needs to operate within clearly defined controls.
There is effectively a progression from filtering exceptions, to explaining them, and ultimately resolving appropriate cases. Where an exception is high-confidence and low-risk, a system can potentially resolve or approve it automatically within predetermined thresholds.
Items requiring judgement would continue to be escalated for human review. The objective is therefore not necessarily to remove people from the process, but to reduce the volume of low-value investigation undertaken by accounting teams.
This changes the role of the accounting professional. Instead of manually triaging every exception, teams can increasingly focus on defining controls, reviewing judgement-based cases, and improving the rules governing automated resolution. Crucially, any automated action needs to retain a complete audit trail and operate within defined permissions and approval thresholds.
NAV production remains one of the most important controlled processes in fund accounting. What role can AI realistically play here?
The opportunity is less about removing formal NAV approval and more about making validation continuous. Most funds still operate on a lagged NAV cycle because prices, trades, corporate actions, cash, and other inputs arrive at different times and in different formats. The accounting book is consequently finalised at defined points even though the underlying portfolio continues to change.
With continuous validation, incoming information can be assessed against expected ranges, historical patterns, and related portfolio activity as it arrives. Potential problems can therefore be identified before they affect downstream valuation.
Formal NAV approval and publication remain governed processes. What changes is the readiness of the data going into them.
For asset servicers, this can also strengthen oversight and contingency capabilities by maintaining an up-to-date view alongside the primary accounting process. Asset managers and asset owners similarly gain earlier visibility into issues that could affect valuation, reporting, or investment decisions.
FundGuard places particular emphasis on bitemporal data. Why is that important when introducing AI into accounting processes?
Explainability requires more than knowing the final answer. Firms need to be able to reconstruct the sequence of information, corrections, and decisions that produced it.
Traditional systems can overwrite historical state when corrections or restatements occur. Bitemporal accounting instead preserves those changes, allowing a firm to establish what was known at a particular point in time and how that information subsequently changed.
That has two important implications for AI. First, regulators, auditors, clients, and internal teams can reconstruct the information on which a decision was based.
Second, it provides a cleaner signal for AI itself. There is an important distinction between a trade that was originally incorrect and subsequently corrected, and one that was correct throughout. An overwrite-based system can lose that distinction.
As AI begins participating more directly in operational workflows, maintaining that history becomes increasingly important for governance and accountability.
How important is having a unified view across different books of record and asset classes?
The usefulness of AI increases significantly with the breadth and quality of the context available to it. A unified architecture can allow different accounting views — including IBOR, ABOR, CBOR, and PBOR — to derive from the same underlying transactional dataset rather than operating as disconnected books.
The same principle applies across asset classes. If equities, fixed income, derivatives, private credit, private equity, real estate, and digital assets share a common portfolio model, AI can potentially identify relationships that would remain invisible within individual asset silos.
This becomes particularly relevant as firms expand across public and private markets. Supporting those assets within the same operating model provides a common foundation for data, controls, and lineage while still allowing AI models to remain specific to individual use cases.
There is considerable discussion around agentic AI. How do you see this developing within investment operations?
The next stage extends beyond automating individual tasks towards systems capable of coordinating governed actions across workflows. Within investment operations, agents could potentially support areas including onboarding, reconciliation, corporate actions, oversight, reporting, and exception management. In public markets, that could include reconciling NAVs across custodians or processing corporate actions. In private markets, it could involve interpreting documentation, supporting capital-call processing, or validating complex calculations.
However, an agent can only operate safely when it has access to trusted data, understands the current operational state, and has clearly defined authority.
The accounting system therefore needs to provide both the context required for the agent to make a decision and the permission framework governing what it is allowed to do. Interoperability standards can help agents communicate across workflows, but they do not resolve fragmented or incomplete underlying accounting data.
What safeguards need to be in place as AI takes a greater role in regulated accounting processes?
AI cannot be judged solely by model accuracy or the percentage of activity it automates. It has to operate within the same control environment expected of regulated financial institutions. That includes confidentiality, data integrity, and availability, alongside explainability, governance, and auditability. Institutions also need clearly defined permissions, decision thresholds, human oversight, and complete audit trails.
There also needs to be a clear approach to failure. Firms should understand what happens when a model encounters activity outside its confidence range, when a data feed fails, or when model performance deteriorates.
In those circumstances, the system should fail safely, preserve the underlying workflow, and escalate activity to human review rather than allowing automation to continue unchecked.
Ultimately, accountability remains with the institution even where AI is making or recommending operational decisions.
Do the potential benefits of AI differ significantly between asset managers, asset servicers, and asset owners?
Yes. AI should be considered in the context of the operational problem each organisation is trying to solve rather than as a generic capability.
For asset managers, a major objective is supporting additional products, asset classes, and jurisdictions without creating proportional increases in operational effort. AI can contribute through areas such as data mapping, onboarding, reconciliation, and anomaly detection.
Asset servicers face a different combination of fee pressure and increasingly complex client requirements. The opportunity is therefore both greater efficiency and greater operating capacity — supporting more tailored reporting, asset classes, and operating models from the same platform.
For asset owners, the emphasis is more likely to be consolidated visibility and independent oversight.
AI can help interpret and normalise data received from multiple managers, custodians, and asset classes, supporting a more current view of the total portfolio and validation of externally reported information.
With so many vendors now describing their technology as AI-enabled, what should firms look for when assessing whether a platform is genuinely ready for production use?
The key question is not simply whether a vendor can demonstrate an AI capability, but where that capability operates and what it can actually do in production.
Firms should establish whether AI operates within the accounting engine or through a separate reporting or data layer; whether historical states and corrections remain reconstructable; and whether models can access information across books and asset classes where required.
They should also examine explainability. A vendor should be able to demonstrate the supporting data and audit trail behind a flagged exception, suggested match, or automated resolution.
Production evidence is equally important. Buyers should understand whether the capability has operated against realistic transaction volumes and complex portfolios in live environments, rather than only within controlled demonstrations.
Finally, firms need clarity around authority. They should know precisely what actions an AI system can take, the permissions and confidence thresholds governing those actions, where human approval is required, and what happens when the model fails.
The distinction between AI experimentation and meaningful operational change will ultimately depend less on access to models than on the accounting and data foundations on which those models operate.
Next interview →
Global Tax Recovery
Mark Friedman
Image: FundGuard 
