Sector note · Capital markets

In post-trade, the money is in the small share of trades that fail

Settlement fails run in the low single digits of volume. Almost the entire post-trade operations effort concentrates on that remainder — which is exactly the work software can carry.

Alphaweb sector note

The exception queue is the business case

Fails are a small share of volume but absorb most of the operations effort, and that residue is still manual.

The American move to T+1 is done, and the industry's own account of it reads well. The after-action report published by SIFMA, ICI and DTCC records that the "average CNS Fail Rate for July 2024 was 2.12%", with non-CNS fails at 3.31%, and notes that the industry "continues to affirm nearly 95% of transactions by the 9:00 p.m. ET cutoff". Prime broker affirmation reached 98 per cent, up from 81 per cent in January.

Read those numbers as a compliment to the pipes. Then read them again as a description of where the people are. Settlement fails run in the low single digits of volume. The operational effort — the settlements team, the asset servicing team, the reconciliations team — concentrates almost entirely on that remainder. That is the inversion worth holding onto, because automation budgets in post-trade have historically gone the other way, towards matching engines, connectivity and standard settlement instructions. That work is finished. The residue is not.

The residue resists conventional automation for a good reason: each instance differs. A fail whose cause sits on a counterparty's screen rather than yours. A rights issue announced as a PDF from a local agent at six in the evening with an election deadline three days out. A reconciliation break that turns out to be a custody fee booked to the wrong ledger in a currency nobody watches. None of this fits a rules engine. It fits a system that can read a document, look the same fact up in four places, notice they disagree, and write down what it found with the evidence attached.

And the clock is now fixed. Both the EU and the UK are moving to T+1 on 11 October 2027. The UK's Accelerated Settlement Technical Group put the staffing question beyond doubt in its February 2025 report: "Relying on the expediency of adding manual resource should not be considered as anything other than a short-term work around whilst automated solutions are developed."

Settlement fails run in the low single digits of volume, and the operational effort concentrates almost entirely on that remainder.

Why this particular work suits software that acts

Diagnostic work, authoritative fact holders, textual evidence and answers checkable against history.

Post-trade exception work has a shape that suits an agent unusually well, and it is worth being precise about why, because the same argument does not transfer to trading, to client coverage or to most of the front office.

First, the work is diagnostic rather than creative. Nobody is asking for an opinion. They are asking what went wrong, and there is a right answer. Second, every fact involved has an authoritative holder — the CSD knows the settlement status, the custodian knows the position, the paying agent knows the record date, the accounting book knows the posting. An agent's job is retrieval and comparison across systems that already publish the truth, not inference from thin air. Third, the evidence is textual and semi-structured: an MT548 reason code, a sese.024 status advice, a counterparty's confirmation, a portal screen, an agent's notice. This is exactly the material a language model reads well and a rules engine reads badly.

Fourth — and this is the part that makes the economics work — correctness is checkable after the fact. You can take last quarter's closed breaks, run the agent over them blind, and compare its conclusion with what your settlements analyst actually concluded and what the CSD record eventually showed. Very few AI deployments in financial services come with a free, labelled, historical test set. This one does.

Fifth, the volume is spiky in a way that punishes human staffing. Index rebalances, a heavy dividend season, a market disruption: the queue triples on a Tuesday and the team is sized for an average Wednesday. Software absorbs that. What it does not absorb is judgement about a client relationship, or a negotiation where the counterparty has to be persuaded rather than informed. Those stay with people, and should.

The three processes that go first

Settlement breaks, corporate actions and reconciliation triage go first; T+1 is a deadline, not a process.

The sequencing below is not arbitrary. It puts first the processes where the work is highest in volume, where the right answer is externally verifiable, and where the agent can produce real value without ever being given authority to move securities or cash. Everything in the table is a drafting and investigation role. A person releases the instruction, sends the client election, and books the journal.

T+1 does not appear as a fourth process because it is not one. It is the deadline that changes the economics of the other three. Compressing the post-trade day does not create new work; it removes the overnight window in which today's manual work quietly happens. The EU T+1 Industry Committee's roadmap is explicit that allocations and confirmations "should be communicated intraday and as close to real time as operationally feasible", and no later than 23.00 on trade date. A process that currently clears at nine the next morning by someone rekeying from an email does not survive that.

What actually stops these projects

Access, decision ownership, reference data and irreversibility, not modelling, decide the outcome.

Almost nothing that kills a post-trade agent project is a modelling problem. Four obstacles account for most failures, and only one of them is technical.

The first is entitlements. An agent that cannot see the custodian's portal, the CSD's status advices, the recs platform and the OMS is a chatbot with opinions. Getting a non-human identity provisioned into those systems — with the right permissions, logged separately, reviewed by information security, and agreed with a third-party custodian whose contract did not anticipate this — routinely takes longer than building the thing. Start it on day one, not after the pilot works. Budget three months and expect the custodian to say no once before saying yes.

The second is ownership of the decision. Operations managers are measured on break counts and ageing, not on time-to-close, and a manager whose bonus depends on a number will not hand the lever to software they did not choose. Worse, nobody wants to be the person who signed off an agent's conclusion. Unless someone senior explicitly owns the agent's output the way they own a junior analyst's, every finding acquires a reviewer, the reviewer redoes the work, and the saving evaporates. This is the most common way these programmes die: not rejected, just re-checked to death.

The third is reference data, which looks technical and is really organisational. ESMA's own diagnosis of settlement fails names "problems with quality of reference data", changes to SSIs "not being communicated or updated in relevant systems in good time", and allocations sent "in a non-STP format or through non-standard channels". Those are not bugs. They are the consequence of nobody owning the SSI master, and an agent will surface that fact loudly in week two. Firms that treat the resulting mess as the agent's failure stop there. Firms that treat it as the finding they paid for carry on.

The fourth is what it costs when it goes wrong, which is asymmetric and must shape the design. A misread fail costs a chaser email and an hour. A misread election deadline on a voluntary corporate action costs a client their entitlement, and it is irreversible. Give the agent unrestricted drafting authority and zero execution authority on anything time-barred or economically final. Accept that this caps the first year's saving, and say so to the sponsor before they hear it from the auditor.

What the supervisor will actually ask

Supervisors ask the ordinary control questions: evidence, reproducibility, validation, register, fallback.

Supervisory interest in this will not arrive as a question about artificial intelligence. It will arrive as the ordinary question about any control: show me how this decision was made, by whom, on what evidence, and prove it would be made the same way today.

In the EU, the operational rules are becoming prescriptive in a way that helps rather than hinders. ESMA's final report on CSDR settlement discipline requires that "written allocations shall be sent using an electronic standardised communication method structured so that software applications can easily identify, recognise and extract specific data", with unstructured channels permitted "only in cases of temporary technical unavailability or service disruption". Regulation is pushing the input into a machine-readable form. Firms that automate against that grain are working with the regime, not around it.

Expect separate questions about the model itself. In the UK, a bank's model risk management framework will want to know how the agent is validated, monitored and re-validated when the underlying model version changes. Under DORA, the agent and whatever it runs on land in the ICT third-party register, with the usual demands for concentration risk, exit planning and incident reporting. And under operational resilience rules, if settlement is an important business service, you will be asked what the impact tolerance is when the agent is unavailable and whether the manual fallback has been tested since the team was resized. Answer that one honestly before it is asked.

What it is worth

Fail exposure varies by an order of magnitude, so size the break population locally before building anything.

Any number offered here would be dishonest, and a senior reader knows it. The value of automating settlement break investigation at a firm whose fails run at two per cent is a different business from the same work at a firm with a long-tail emerging markets book. ESMA's own technical advice records that "ETFs' settlement fails amounted on average to 17.32% of the monthly total volume of ETFs settlement instructions at EEA level" between June 2023 and May 2024. Two firms in the same city, same regulator, an order of magnitude apart in exposure.

So do the arithmetic locally, and do it before you build. Take one break population for one quarter. Count the items. Time twenty of them properly — not the estimate the team gives you, the measured elapsed time including the wait for a custodian's reply. Multiply. Then add the part nobody counts: the CSDR cash penalties that accrued while the break was open, the client compensation for the missed entitlement, the two people who spend a rebalance week doing nothing else. That total is your ceiling. Assume you capture some fraction of it in year one, and that the reference data cleanup you did not plan for eats part of the rest.

Alphaweb is a young studio and does not have a shelf of capital markets logos to point at. What it has is a view of this work that we think is correct: that the interesting problem in post-trade is not making the straight-through path faster, it is giving the exception queue the same treatment — software that carries an investigation from the fail to the written finding, on infrastructure the firm can audit, with a person on the release button. If that is wrong for your book, the numbers above will show it within a quarter. That seems a fair test.

A ninety-day sequence that survives contact

1. Days 1-15 — Pick one population and count it honestlyChoose a single desk and a single break type — say, fails on European equities settling through one CSD. Pull a quarter of history. Count the items, then sit with an analyst and time twenty investigations end to end, including the dead time waiting on a custodian. Do not accept the team's estimate; measured time is always different from remembered time. This number is the only baseline that will survive contact with a sceptical CFO, and it is also your test set.
2. Days 16-30 — Start the access request before anything elseProvisioning a non-human identity into the custodian portal, the CSD status advices, the reconciliation platform and the order management system is the longest item in this plan and the least interesting. Open it now, in parallel with everything else. Expect information security to ask how the credentials are held, expect the custodian's contract to be silent on machine access, and expect one refusal before agreement. Nothing else in the ninety days matters if this lands on day eighty-five.
3. Days 31-50 — Build it read-only and let it be wrong in privateThe agent investigates and writes findings. It takes no actions, sends no messages and touches no instruction. Run it over the closed historical cases from the first fortnight and over live breaks in shadow, with the output going to a queue nobody acts on. This is where the reference data problems surface — mismatched SSIs, stale account maps, a place-of-settlement field three teams populate differently. Treat those as the deliverable, not the setback.
4. Days 51-70 — Grade it against the human baseline and set the authority lineCompare the agent's conclusion with the analyst's conclusion and with what the CSD record eventually showed. Report agreement rate, the cases where it was right and the human was wrong, and the shape of its failures. Then draw the authority boundary explicitly and write it down: which classes of break it may draft a correction for, which it must escalate untouched, and what it is never permitted to originate. Anything irreversible or time-barred sits on the escalate side.
5. Days 71-90 — Give it drafting authority on one reversible action and document the controlLet it draft the counterparty chaser or the amended instruction for a single break class, with a named releaser. Instrument everything: time to first finding, time to close, rate of released drafts amended by the human, and penalties avoided. In the same fortnight, write the control documentation — validation approach, monitoring, re-validation on model change, third-party register entry, and the tested manual fallback. Producing this now is cheaper than producing it under an audit request in month nine.

What we read

The documents behind this note. Each entry says what it is, what it found, and why it should change what you do — then the link to the original.