AI agents for asset managers and fund back offices

Quick answer

Asset managers and fund back offices use Nori to run the data work that fills every morning: keeping the security master current, building fund look-through tables, reconciling custodian files, and checking the inputs to NAV. Agents run on a schedule in isolated sandboxes inside your own cloud, read the files and feeds you already license, and report in Slack. They draft, match and explain. A person in operations approves anything that writes to a system of record.

The data problems a fund back office deals with every day

Back-office work is mostly data work with a deadline. The same four problems show up at firms of every size.

The security master drifts

Identifiers change after corporate actions. Vendors and custodians update on different days. A bond's maturity date differs between two feeds. Each mismatch is small, and each one breaks a position match or a report downstream until someone finds it. The fixing is done by hand, usually by the people who know the instruments best.

Look-through is a quarterly scramble

Fund holdings arrive as PDFs and spreadsheets in a different layout per administrator, weeks after quarter end. Names are abbreviated and identifiers are missing. Analysts retype and match for days to produce one exposure table, and the mapping knowledge lives in a spreadsheet one person maintains.

Reconciliation breaks pile up

Positions, cash and transactions from each custodian must match the books. Most breaks have routine causes: a trade settling a day late, a price source difference, a corporate action booked on one side only. Each still needs someone to look, find the cause, and write the note. On a busy day the queue outruns the team.

NAV inputs are checked under time pressure

Before a NAV is struck, someone checks for stale prices, missing prices, large moves, FX rates and accruals. The checks are well defined but run against the clock, and the evidence that they were done is often a tick in a checklist.

Controls and data location limit the tools you can use

Holdings, positions and client data cannot be pasted into a chat tool. Any automation has to run where the data is allowed to be, under your identity provider, with a record of what it did and who approved it.

How Nori helps asset managers and fund operations teams

Nori gives the operations team agents that work the way an analyst does: read the files, apply the written procedure, show the evidence, and ask when the procedure does not cover a case. The whole team shares them through Slack, schedules and one set of instructions.

Security master upkeep on a business-day schedule

An agent loads vendor reference files and custodian files each morning, validates and cross-matches CUSIP, ISIN, SEDOL, FIGI and ticker, applies your survivorship rules, and writes field-level lineage. Conflicts go to an exceptions queue with the evidence and a proposed resolution. It works on feeds you already license. See how to build an agentic security master for schemas and a sample exception.

Fund look-through that rebuilds itself

An agent parses each holdings report, checks that parsed weights sum to the printed total, maps holdings to the master, and multiplies weights through every fund layer. The output is an exposure table by issuer, sector and country, with unmapped holdings and funds without reports shown as their own rows. See the fund look-through worked example.

Custodian reconciliation with the cause attached

After the matching engine or SQL job has run, an agent takes the breaks and investigates each one against the trade blotter, the corporate actions calendar and the prior day's breaks. It groups them by likely cause and drafts the break note. A sample from a morning summary, with illustrative data:

The agent does not post adjustments. An analyst approves or corrects each proposed action, and the decision is recorded with the break.

NAV data checks before the cut-off

A scheduled agent runs your pre-NAV checklist as queries: prices unchanged for more than a set number of days, holdings with no price, moves beyond a threshold with no matching market or corporate action, FX rates missing or stale, accruals that changed sign. It posts the exceptions with the rows behind them. The checks and thresholds are yours, stored in the repository and changed only by reviewed pull request. The agent checks inputs; it does not calculate or approve the NAV.

Deployment inside your own cloud

For Enterprise customers Nori deploys the full stack, control plane included, into your AWS VPC or Azure VNet, or on-prem. Sessions use your identity provider through SAML, SCIM and OIDC, your LLM gateway and your compute. Files, queries and prompts stay inside your network. Hosted customers get a dedicated, single-tenant control plane and data store per organization.

What a morning looks like

The schedule below is one operations team's set of Nori automations. Times are in the team's local time zone.

06:00  Mon-Fri   secmaster-daily      load reference and custodian files, match, queue exceptions
06:30  Mon-Fri   custodian-recon      investigate breaks from the 06:15 matching run
07:15  Mon-Fri   nav-input-checks     stale, missing and outlier prices; FX; accruals
07:00  Monday    lookthrough-refresh  parse new holdings reports, rebuild affected portfolios
webhook          late-file            rerun one custodian when its file lands after cut-off

By the time the team is at their desks, the channel has three short messages. Each lists what was checked, what passed, and the handful of items that need a decision. An analyst replies in the thread:

j.okafor  08:02
approve EXC-2026-1111-001. The split is confirmed on the issuer notice.

Nori  08:03
Done. CUSIP 66611T109 end-dated 2026-11-09, 66611T208 added from
2026-11-10 on SEC-0048213. Lineage records you as decided_by.
Rerunning custodian_1 position match for that security: now matched
(4,150 = 4,150). Break closed.

Nothing here replaces the team's judgment. The agents remove the collecting, matching and note-writing, so analysts spend their time on the exceptions.

Controls: who approves what

• Read broadly, write narrowly. Agent roles read landing files, the warehouse and reference tables. They write to staging tables and exception queues. Writes to a system of record happen only after a named person approves.
• Rules are files. Survivorship rules, tolerances, checklists and conventions live in your repository. Changing one is a pull request with required reviewers from operations.
• Every run leaves a record. Run history keeps the trigger, the prompt, the commands, the queries and the output of each session. Lineage tables record the source file and the person or rule behind each value.
• Scoped identities. Integrations are scoped to an organization, a person or a specific agent, so the reconciliation agent cannot use the credentials of the secmaster agent.
• Your models. Sessions run on your own Claude or ChatGPT subscription, API keys or LLM gateway. Nori never charges for tokens.

Nori does not give investment, legal or compliance advice, and the agents do not make investment decisions. Your operations, risk and compliance staff stay responsible for the records and for how any output is used. The underlying platform pattern is described in agentic data operations, and the role view is on Nori for finance teams.

Where to start

Start with the job that has the clearest procedure and the most repetition. For most back offices that is security master exceptions or break investigation for one custodian.

1. Write the procedure the team follows today into the team skillset, including when to escalate.
2. Give the agent read access to yesterday's files and run it by hand from Slack next to the analyst doing the same work.
3. Compare results for a week or two. Correct it in the thread; corrections are saved to shared memory.
4. Put it on a schedule, keep the approval step, and add the next custodian or the next check.

To see it on your own file layouts, book a demo.

Frequently asked questions