AI and T+1 Settlement: Where Machine Learning Helps and Where Rules Do the Work

Updated October 2026, with two years of T+1 data. First published March 2024.

In March 2024 this post said artificial intelligence would sit at the heart of the move to T+1. It would match trades, predict fails, allocate collateral, and learn from every transaction. T+1 took effect on May 28, 2024, and there are now more than two years of numbers to check that against. Most of the improvement came from rules-based matching and automatic affirmation. Machine learning has a real job in settlement, but it's a smaller one: telling people which exceptions to work first.

What Got Affirmation Rates to 96%

DTC's cutoff for affirming an institutional trade is 9:00 p.m. ET on trade date. At the end of January 2024, about 73% of trades were affirmed by that time. By July 2024 the figure was nearly 95%, according to the T+1 After Action Report from SIFMA, ICI, and DTCC. In 2025 it averaged close to 96%.

The clearest account of how that happened is a report DTCC ITP Matching (US) now files with the SEC every year under Rule 17Ad-27. Its report for 2025 breaks affirmation out by workflow:

The top rows are deterministic. A central matching service compares the investment manager's allocation with the broker's confirmation, and when they agree, the trade is affirmed without anyone touching it. The bottom row is people affirming trades one at a time, many of them at custodians and managers outside the U.S. The ten-point gap is what it costs when a person has to be at a desk before 9 p.m. The report doesn't mention artificial intelligence or machine learning at all. For what each of these tools does, see our post on what central matching automates under T+1.

Where Machine Learning Has a Job

A match has to give the same answer every time, so it runs on rules. The work around the match is messier, and that's where a statistical model can help. Two uses have public support.

The first is predicting fails. In its March 2025 consultation report on AI in capital markets, IOSCO said member regulators had seen exchanges, clearing houses, and other market infrastructures use machine learning to predict the likelihood that a trade will fail to settle. In practice a model like this scores open trades by risk, and the desk works the highest scores first.

The second is reading what arrives unstructured. Settlement instructions, letters, and confirmations still come in by email, PDF, and fax. A model can classify them and attach them to the right account, so they land in a queue instead of an inbox. IOSCO also reported firms testing large language models to write transaction summaries and draft documentation.

Both are triage. A prediction doesn't fix anything by itself. A model that flags a bad allocation at 8:45 p.m. is worth very little if nobody can correct it before the 9 p.m. cutoff. The value depends on how early the score arrives and whether someone owns the queue.

Where It Doesn't Belong

Keep models out of the steps you have to explain the same way every time: the match, the affirmation, and the release of a delivery. If an examiner asks why a trade was affirmed, "quantity, price, and settlement date matched the allocation" is an answer. "The model scored it 0.93" isn't one you want to give. For those steps, a written rule is the better tool.

The Rules Don't Change for AI

Neither the SEC nor FINRA has written settlement rules specific to AI. FINRA's Regulatory Notice 24-09, issued June 27, 2024, says FINRA intends its rules to be technology-neutral and that they apply when firms use generative AI as they would to any other tool. If a model sorts your exceptions, your supervisory system under FINRA Rule 3110 has to cover it. At the SEC, the 2023 proposal on predictive data analytics, which dealt with how broker-dealers and advisers use such tools with investors, was withdrawn on June 12, 2025.

That leaves the existing rules. Under Rule 15c6-2, a broker-dealer that relies on policies and procedures has to describe the systems it uses for allocation, confirmation, and affirmation, set target times on trade date, and measure its completion rates. An AI-assisted step is one more system to describe and measure.

Questions to Ask About an AI Settlement Tool

1. What does it predict or classify, and how accurate is it on your trades, not on a demonstration data set?
2. How early does it flag a trade? A score after 8 p.m. ET leaves little time to fix an allocation.
3. What happens to items it scores as low risk? Someone should still review a sample.
4. Can you reconstruct, a year from now, why it flagged a particular trade?
5. Does the tool decide anything, or does it only rank? Prefer a tool that only ranks.

Our own use of AI is narrow in the same way. FVD-R uses it to index Letters of Free Funds that arrive by email, fax, or PDF and match them to the right account. Items it can't match with confidence go to a review queue, and an authorized person still confirms or denies every request.

This post is for informational purposes only and does not constitute legal advice. For guidance on specific regulatory obligations, consult your counsel or compliance advisor.