Last verified 2026-07-12

NAV break & reconciliation-exception triage

Can AI investigate NAV breaks and reconciliation exceptions? Yes as an assistant — it clusters, ranks and hypothesises causes so a human works a shortlist with evidence attached. No as a controller: a model that both explains and clears its own exceptions removes the independent control the reconciliation exists to provide.

The pain	Reconciliation throws exceptions daily — a position, price, cash movement or trade that does not tie between the administrator, the custodian and the manager's book. Each break is a small investigation: what is it, is it timing or a real error, whose side is wrong, what is the evidence, can it be cleared before the NAV strikes. The volume is dominated by benign timing differences, but you cannot ignore them, because the one that matters is hiding in the pile.
What AI does today	Triage. Cluster the breaks, label the obvious timing differences, hypothesise a cause for the rest ("FX rate mismatch on these three", "trade booked a day late one side"), pull the supporting records together and suggest the next step — so a human works a ranked shortlist with evidence attached instead of an undifferentiated queue. The boundary: the recon engine matches deterministically and rules auto-clear the known patterns; AI only earns its place on the unpatterned residue.
Proof it's real	Operator assessment. Reconciliation vendors are adding break-explanation features and fund administrators describe AI anomaly-flagging in their operations, but we found no named, measured production deployment of AI break-triage at a fund or administrator. The capability is plausible and the tooling is appearing; the deployment evidence is not there yet. See To verify.
What it can't do	It cannot clear a material break unsupervised — the human investigator is the control, and stays. The whole point of reconciliation is independence; a model that both explains and clears its own exceptions removes the control it exists to provide. And a confident wrong hypothesis is worse than none — it anchors the investigator toward the wrong cause.
The real alternatives	Your administrator's middle office — for most managers the breaks are already worked by the admin as part of NAV production; the real question is their tooling, not yours. A reconciliation platform (deterministic match rules + auto-clear of known timing patterns) — mature, and clears most volume before anything intelligent is needed; wins almost always as the first buy. An in-house triage view on the recon extract — middle-office teams already script their own exception views on the recon output; an AI triage layer on top is the natural extension where group policy allows internal builds. Excel tick-and-tie — survives on small books and single-custodian setups. AI triage is an add-on to the platform route, not a rival to it — if you don't have deterministic recon in place, that's the gap, not AI.
What you need in place	A working recon engine first (AI over un-matched raw feeds is triage over noise); access to both sides of every reconciliation; and clear ownership — fund accounting or the middle office owns clearing, with the escalation path unchanged. If the admin runs your recs, this is a question for their roadmap before it is a project of yours. And recon tooling is often a group-level platform decision — the desk's lane is exception-handling quality, not tool procurement.
Effort & cost	Weekend script — feed a day's exception file to an LLM, ask it to group and rank by likely materiality; free, and a fair test of whether triage helps your book. Off-the-shelf — recon platforms with break-explanation features arriving in their subscription tiers (five figures a year, typically). Real project — a triage layer on your recon engine with evidence-gathering into the investigator queue; five-to-six figures. If your break volume is low or already outsourced, don't build — ask your admin what their tooling does instead.
What to watch	Watch for anchoring — does the suggested cause make investigators stop looking? Measure whether real breaks are found faster, not whether the queue looks tidier. A tidy queue that buried one true break has failed.

Questions operators ask

Can AI clear reconciliation breaks automatically?

Known, benign timing patterns are already auto-cleared by deterministic rules — no AI needed. What AI adds is triage of the unpatterned residue: clustering, ranking and hypothesising for a human to resolve. A material break cleared by the same system that explained it is a control failure, not efficiency.

Do we need AI if we already have a reconciliation platform?

Maybe not. The platform clears most volume; AI's value is proportional to your unexplained-exception tail and the hours it eats. Measure that tail first — if it's small, the platform plus rules is the finished answer.

To verify

Named AI break-triage deployment — no fund or administrator publicly naming AI exception-triage in production with measured results as of 2026-07-12; grade upgrades when one appears.

Related: the Fund AI desk and the other operational use-cases in the library.

Changelog

2026-07-12 — published.