AI Exception Management in Post-Trade

Share this email with a colleague now so they’re on the same page. We only grow through word of mouth and your referrals - Brennan

Table of Contents

• Where AI Fits

• The Data Problem Underneath

• Key Terms

• Sources

Most of the cost in post-trade operations is not in straight-through processing. It is in the exceptions. Up to 10% of operating costs in asset servicing come from errors and breaks (ISSA/ValueExchange, 2024). STP rates for mandatory and income events sit at about 71% (ISSA, 2024). Two thirds of the errors trace to bad data, not bad process (ValueExchange/ISSA, 2024).

Those inefficiencies carry a heavier penalty today. Volumes rose 25% in 2025, automation levels fell at more than 60% of brokers over the same period (Broadridge, 2025), and T+1 compressed the US exception window from overnight to two to five hours (Tokenovate, 2024). The workload got bigger and the time got shorter.

Where AI Fits

Exception management follows patterns. The details vary but the shape of the work repeats: something is wrong in the data, someone works out what, someone fixes it.

AI can speed up the early parts of that. Catching data gaps before end-of-day reconciliation rather than after. Sorting breaks by type and severity in seconds rather than minutes. Some systems now propose fixes based on past resolutions. Early data points to 45% lower handling time and 35% lower costs where AI is applied (Tokenovate, 2024), though those numbers come from limited samples.

Where it gets stuck is the last step. Agentic AI, systems that can reason through a problem and act on it, would make the biggest difference here. But automated resolution requires firms to let the model act, and most will not, because the governance is not there. Who owns the model? Who is accountable for a wrong resolution? What gets escalated and what does not? Without answers to those questions, the model triages but a person still fixes.

The Data Problem Underneath

Vendors rarely talk about the next part. Two thirds of exceptions come from bad data. The AI is trained on that same data. If the training data is full of the errors the model is supposed to catch, the model learns the errors as normal. It does not flag them. It treats them as baseline.

That makes the data quality problem the actual AI problem. Firms that want AI to work in exception management need to fix the data first, or at least fix it at the same time. That means reference data, static data, market data feeds, and the way corporate actions are coded across systems. It is slow work and it does not make for good conference slides, but without it the model has nothing clean to learn from.

The governance problem sits on top of the data problem. Even if the model is accurate, nobody will let it act without clear accountability. And even if the accountability is clear, nobody will trust the outputs if the data underneath is unreliable. One does not get solved without the other.

That connection between data quality and governance is what determines whether AI in exception management actually ships or stays in pilot. The technology itself is the least interesting part.

I write about AI adoption for operations leaders at Getting AI To Work.

Key Terms

Exception management: Identifying, investigating, and resolving breaks in post-trade processing.

STP (straight-through processing): When a trade or event processes without anyone needing to touch it.

T+1 settlement: Trades settle one business day after execution. Adopted in the US and Canada in May 2024.

Reconciliation break: A mismatch when comparing records between two parties or systems.

Agentic AI: AI that can reason through a problem, use tools, and act with limited human oversight.

Sources

• ISSA/ValueExchange, Asset Servicing Study, 2024

• Broadridge, Operations Survey, 2025

• Tokenovate, Post-Trade AI Analysis, 2024