AI use case

AI for settlement fail prediction and post trade exception management

AI that scores each pending securities settlement instruction for its likelihood of failing, names the probable cause (unmatched instruction, wrong settlement details, lack of securities or cash), and helps operations teams work the exceptions and counterparty queries before the intended settlement date, so fewer trades fail and fewer late settlement penalties are paid.

By Len Debets · Last verified 27 September 2026 · 4 public deployments

EUR 250,000 to EUR 3.8 million
Indicative value per year
A broker dealer settling 2 million securities instructions a year in EU markets. Worked example, see how it is calculated.

What problem does it solve?

A securities trade settles only when both sides have sent matching instructions, the seller has the securities and the buyer has the cash, all by the intended settlement date. When any of these is missing the trade fails. Fails tie up liquidity and collateral, create credit exposure between counterparties and, in the EU, trigger cash penalties under the settlement discipline regime of the Central Securities Depositories Regulation (CSDR), which applies since February 2022. BNY Mellon has described how on a typical day about two percent of US Treasury transactions fail to settle.

Many operations teams still find a fail after it has happened: they work through a pending and failing report in the morning, look up each trade in several systems and chase counterparties and custodians by email and phone. The window to fix a problem is getting shorter. The United States moved most broker dealer transactions to settlement one business day after the trade date (T+1) in May 2024, and the EU plans to follow in October 2027, which leaves hours rather than days to spot a mismatch and correct it. The step change is to predict which instructions are at risk while there is still time to act, and to take the reading, looking up and chasing out of the exception queue.

How does it work?

  1. Collect the instruction lifecycle. The system reads every pending instruction with its matching status, counterparty, instrument, market, place of settlement, amount and the securities and cash positions behind it, from the firm's own books and from the status messages of the central securities depository (CSD) or custodian.
  2. Score the fail risk. A model trained on historical settlement outcomes estimates the probability that each instruction settles on time and ranks the drivers, such as an unmatched instruction, a counterparty with a poor record, an illiquid bond or a short position. Some CSDs now offer this score to their participants as a service.
  3. Estimate the cost. For each instruction at risk it estimates the exposure: the late settlement penalty, the funding cost and the client impact, so the team works the most expensive problems first.
  4. Work the exception. For the top of the queue an AI agent gathers the facts (both instructions side by side, standing settlement instructions, inventory, previous correspondence), proposes the likely fix and drafts the query to the counterparty or custodian in the channel they use.
  5. Decide and act. An operator approves any amended instruction, securities borrow, partial settlement or cash movement. The agent sends approved queries, tracks answers, chases before the cutoff and records the root cause so the same break does not return.
Audience
Back office
Autonomy
Copilot
Adoption
Early adopters
Channels
Internal tools, API and system to system, Email

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

No public deployment has disclosed a measurable outcome yet.

Value drivers: Risk and loss reduction, Lower cost to serve, Employee productivity, Speed and cycle time.

Indicative value

A broker dealer settling 2 million securities instructions a year in EU markets

EUR 250,000 to EUR 3.8 million

Fail cost avoided per year

How this is calculated

Formula: instructions * failRate * preventedShare * costPerFail. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Settlement instructions per year instructions, instructions per year2,000,0002,000,000The reference firm. Replace with your own instruction volume.
Share of instructions that fail today failRate, fraction of instructions0.0250.05The low value follows ESMA's final report on the CSDR penalty mechanism (November 2024), where fails across all EEA CSDs in June 2024 were about 2.5 percent of the number of settlement instructions for sovereign bonds, the lowest asset class, and about 20 percent for ETFs. The high value of five percent is an editorial assumption for a mixed book, not a sourced figure. Replace both with your own fail rate.
Share of fails prevented by acting on the prediction preventedShare, fraction of fails0.10.25Editorial assumption, deliberately conservative. No deploying organization on this page has published a measured reduction in fails.
Cost of one fail costPerFail, EUR per fail50150Editorial assumption covering cash penalties, funding cost and operations handling time. Replace with your own penalty and handling data.

What it leaves out: Counts only fails that are prevented. It leaves out the operator time saved on fails that still happen, penalties received from counterparties, the capital and liquidity effect of fewer open fails, client impact, and the cost of the data, the models and the integration.

Who already uses it?

4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

Euroclear

Belgium · Capital markets · 2025

ProductionGrade B

In June 2025 Euroclear announced EasyFocus+, the next generation of its EasyFocus service, built with Meritsoft and Taskize and running on a Microsoft cloud. It uses predictive analytics to flag likely mismatches, identify root causes and resolve exceptions, and gives clients a single view of their settlement instructions across the Euroclear CSDs, which represent over 60% of EU settlement. The collaboration platform of Taskize, a member of the Euroclear group of companies, is embedded for routing and resolving the exceptions with counterparties. Euroclear positions it as support for the EU move to T+1 in October 2027.

No outcome disclosed.

Clearstream

Luxembourg · Capital markets · 2022

ProductionGrade B

Clearstream, the Luxembourg based international central securities depository of Deutsche Börse Group, launched an AI Settlement Prediction Tool for its clients in July 2022, together with a Settlement Dashboard. The tool estimates the likelihood that a specific instruction settles on time. The enhanced version released in July 2025 identifies potential failure drivers and at risk instructions up to four business days in advance and estimates potential penalty costs, to support clients' T+1 readiness. Clearstream's current product page adds that the tool names the three primary factors most likely to cause a fail and estimates daily penalty costs. Clients access it in the Xact Web Portal. Clearstream has not published measured outcomes.

No outcome disclosed.

BNY

United States · Capital markets · 2021

AnnouncedGrade B

On 4 February 2021 BNY, then branded BNY Mellon, announced a collaboration with Google Cloud to predict settlement failures in the US Treasury market, where it provides clearance and settlement. It trains models on millions of trades on Google Cloud's data analytics and machine learning services. BNY Mellon's Clearance and Collateral Management head described the aim as helping clients predict approximately 40% of settlement failures in Fed eligible securities with 90% accuracy. The release describes a solution in development; no production results were published.

No outcome disclosed.

BNY

United States · Capital markets · 2026

ProductionGrade C

BNY built its own AI platform, Eliza, on Microsoft Azure and Microsoft Foundry, and uses it across operations. In a Microsoft customer story, BNY's Head of AI Enablement says more than ten percent of the inquiries clients raise about BNY's transactions were resolved or assisted by AI, with eighty percent faster processing of those inquiries. Microsoft's summary calls them client settlement inquiries; BNY's own words are broader, so the link to settlement exception work is adjacent rather than direct. The same story describes a digital employee that repairs incomplete payment instructions and an agentic workflow for client onboarding research.

  • Cycle time reduction: 80%, client transaction inquiries resolved or assisted by AI
    "More than ten percent of these inquiries were resolved or assisted by AI and that has resulted in eighty percent faster processing of these inquiries."
    Claimed by: organization

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • At least a year of instruction history with matching status changes and the final settlement outcome
  • Standing settlement instructions and counterparty static data
  • Securities inventory and cash positions per account and depot
  • Penalty reports from each CSD, to price the cost of a fail
  • Past exception cases with root cause and the correspondence that resolved them

Systems to integrate

  • Settlement or post trade processing system (pending and failing instructions)
  • CSD and custodian status feeds, and their prediction services where offered
  • Inventory, securities lending and collateral systems
  • Email and post trade query platforms used with counterparties
  • Case or exception management tool for the operations queue

Complexity: High

The prediction needs clean, joined history of instructions, matching statuses and outcomes across every CSD and custodian the firm uses, and the model falls under model risk management. Using the score a CSD already provides lowers the entry cost; the exception work still needs integration with the settlement system, inventory and counterparty communication.

  1. 1

    Measure the fails you have

    Take six to twelve months of fails and penalties and group them by cause, market, counterparty and asset class. Late matching, wrong settlement details and lack of securities usually need different fixes, and the size of each group decides where to start.

  2. 2

    Use the CSD score before you build your own

    Where your CSD or custodian offers a settlement prediction, feed its score and drivers into the operations queue first. Build an internal model only for flows the service does not see, such as your own internal settlements and positions.

  3. 3

    Rank the queue by cost, not by age

    Combine the fail probability with the penalty, funding and client impact so operators start with the instructions that matter most, and show the drivers next to each score.

  4. 4

    Automate the gathering and the first query

    Let the agent assemble both instructions, the standing settlement instructions and the inventory, propose the fix and draft the counterparty query. Operators approve every outgoing message at first, then allow routine information requests to go out directly. From that point, tell recipients in each message that it was written and sent by an AI agent.

  5. 5

    Close the loop on root causes

    Record the confirmed cause of every fail and every prevented fail, feed it back into the model and fix the static data or counterparty set up that caused it.

Guardrails

  • Maker checker approval on every amended instruction, borrow, partial settlement or cash movement
  • The agent never changes standing settlement instructions or static data on its own
  • Every score shows its drivers, so operators can see why an instruction is flagged
  • Outgoing counterparty queries use approved templates and contain only the data needed for the trade
  • Model monitoring for drift, with a documented fallback to the manual pending report

KPIs to instrument

  • Settlement efficiency by value and volume, before and after, per market
  • Late settlement penalties paid per month, net of penalties received
  • Precision and recall of the fail prediction on a holdout period
  • Median time from flag to resolution for instructions at risk
  • Operator minutes per exception and share of queries sent without manual drafting

Human in the loop

Operators decide on every action that changes an instruction or moves securities or cash, and own escalations to the front office and clients. Team leads review a weekly sample of predicted and actual fails, and model owners validate the prediction model on a schedule.

Common failure modes

A score nobody acts on
The prediction lands in a dashboard outside the operations queue and changes nothing. Put the score and the drivers inside the tool operators already work in.
Too many flags
A model tuned for recall floods the team with instructions that would have settled anyway. Tune the threshold on cost and track precision per market.
Stale static data behind the prediction
Wrong standing settlement instructions cause fails the model cannot explain. Treat reference data fixes as part of the programme.
A model that learns from a different cycle
A model trained under T+2 behaviour misjudges risk after a move to T+1. Retrain and revalidate around every change in settlement cycle or market practice.

What are the risks and rules?

EU AI Act

Depends on design

Predicting settlement fails and handling post trade exceptions between professional market participants is not a use listed in Annex III and is not a prohibited practice under Article 5, so the tier depends on how the agent communicates. While an operator reviews and sends every message, the system is minimal risk: the messages are the firm's own correspondence and the firm as deployer owes AI literacy for staff (Article 4). Once the agent sends queries or chasers to counterparty or custodian staff itself, as the playbook recommends for routine information requests, it interacts directly with natural persons and Article 50(1) requires telling the recipients they are dealing with an AI system. In both designs the provider of the text generating system must mark its output as AI generated in a machine readable format under Article 50(2). Model risk and operational resilience controls apply on top.

Guidance

Controls to put in place

  • Model inventory entry and periodic validation for any internal fail prediction model
  • Documented fallback to the manual pending and failing report if the model or feed is unavailable
  • Full trail of scores, proposed fixes, approvals and counterparty messages per instruction
  • Monthly review of penalties and fails against the prediction, per market and counterparty
  • Third party risk assessment for CSD or vendor prediction services under DORA

Frequently asked questions

Can AI really predict which trades will fail to settle?
Central securities depositories already offer such predictions. Clearstream's product page says its Settlement Prediction Tool calculates the likelihood that an instruction settles on time and identifies the three primary factors most likely to cause a fail up to four business days in advance. In February 2021 BNY Mellon said its model with Google Cloud aimed to help clients predict about 40 percent of settlement failures in Fed eligible securities with 90 percent accuracy, a goal rather than a result. Neither states a measured accuracy or a measured reduction in fails on the pages cited here.
Do we need to build our own model?
Not to start. Clearstream offers its prediction to clients through its Xact Web Portal, and Euroclear offers EasyFocus+, announced in June 2025, which gives each pending instruction a matching score (the predictive likelihood that it will be matched) and shows the CSDR penalty impact across its CSDs. An internal model adds value for flows those services do not see, such as your own inventory and internal settlements.
Where does generative AI help if the prediction is a classic model?
In the exception work: reading both instructions, drafting counterparty queries and chasing answers. At BNY, more than ten percent of client inquiries about its transactions were resolved or assisted by AI on BNY's Eliza platform, and BNY says this brought eighty percent faster processing of those inquiries. The inquiries are not limited to settlement exceptions.
Is this high risk under the EU AI Act?
No. Settlement fail prediction between professional market participants is not an Annex III use, and the tier depends on the design. With an operator sending every message it is minimal risk. Once the agent sends queries to counterparty or custodian staff itself, Article 50(1) requires telling those recipients they are dealing with an AI system, and the provider of a text generating system must mark its output under Article 50(2) in either case. Beyond that, model risk management, operational resilience under DORA and good records of every approved action are the controls that matter.

How to cite this page

Blits.ai AI Use Case Library, "AI for settlement fail prediction and post trade exception management", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/settlement-fail-prediction-and-exception-management. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 27 September 2026: First published

Related use cases

BankingPayments and cards

AI for ledger and payment reconciliation

AI that matches entries across nostro and vostro statements, card and scheme settlement files, the general ledger and suspense accounts, proposes matches and clearing journals, and routes only the genuine breaks to an operator with a plain language explanation.

Deployments
4 public, best grade B
Autonomy
Supervised agent
BankingPayments and cards

AI for payment investigations and exceptions

AI that works the payments that fall out of straight through processing: it reads the failure, repairs or enriches the message, drafts the ISO 20022 or SWIFT investigation, chases the counterparty bank and proposes a return, recall or correction, while an operator approves anything that moves money.

Deployments
2 public, best grade B
Reported automation rate
at least 10%
BNY, organization claim
Wealth and asset managementBanking

AI portfolio drift monitoring and rebalancing proposals

Continuous monitoring of every client portfolio against its mandate or model, which detects drift beyond agreed bands and prepares a tax aware, low turnover rebalancing proposal with its rationale for an advisor or portfolio manager to approve before any trade is placed.

Deployments
3 public, best grade B
Autonomy
Copilot
Cross industryBanking

AI for continuous controls testing and control self assessment

AI that moves control testing from periodic samples to continuous, full population assurance: it collects evidence from source systems, maps each artefact to the control it supports, tests every transaction or record against the control's rule, flags exceptions for a human to judge and prepares the risk and control self assessment from incident and loss data for the business to review.

Deployments
3 public, best grade B
Autonomy
Supervised agent