menu_bookGuide

AI agents for capital calls and NAV reporting

What an AI agent actually does at each step of a capital call and a NAV cycle, where a human must approve before anything leaves the building, and what evidence lands in the audit trail on every run.

The problem these processes create

Capital calls and NAV production are the two processes in fund operations where the cost of an error is highest and the time available to catch it is shortest. Both are deadline-driven, both draw on data spread across a ledger, a custodian, a pricing feed and a commitment schedule, and both are still run in most firms as a sequence of spreadsheets, emails and manual checks held together by the experience of a handful of people.

Two things follow from that. First, review capacity becomes the constraint: someone senior has to look at everything, because there is no reliable way to know in advance which cases are exceptions. Second, audit evidence is reconstructed after the fact - pulled together from mailboxes, file shares and memory when a regulator, auditor or LP asks how a figure was reached and who approved it.

AI agents change the first problem only if they do not make the second one worse. An agent that drafts a call notice quickly but leaves no record of the data it used, the calculation it performed and the person who approved it has moved the work, not the risk. The rest of this guide is specific about both halves: what the agent does, and what is captured while it does it.

Capital calls, step by step

A capital call in Next Matter runs as a single orchestration across the commitment schedule, the ledger, the investor records and the outbound channel. The agent handles calculation, drafting and reconciliation; the accountable person approves at two fixed control points.

Calculate the call against the commitment schedule

The agent reads the drawdown requirement and pulls each LP's commitment, prior drawdowns, remaining undrawn amount and any side-letter terms from the source of record. It computes the per-LP call amount pro rata, applies excuse and exclusion provisions, and produces a call schedule with the calculation inputs recorded alongside each line rather than only the result.

AI agent

Reconcile and flag before anything is drafted

Every line is checked against the ledger and the investor register: commitment totals that do not tie out, an LP whose bank details changed since the last call, a transfer registered mid-period, a rounding difference above tolerance, a stale side letter. Lines that reconcile continue; lines that do not become exceptions and stop, with the mismatch, both source values and the LP's history attached.

AI agent

Route exceptions to the right reviewer with context

Each exception goes to a named reviewer as a task containing the specific discrepancy, the underlying records from both systems, and the agent's proposed resolution. The reviewer accepts, amends or rejects. Nothing is silently corrected, and nothing sits in a queue that no one owns - unresolved exceptions escalate on the schedule you define, well before the notice deadline.

Human decision

Draft the notices

The agent generates each LP's call notice from the approved schedule and the fund's template: amount, due date, purpose, wire instructions, remaining undrawn commitment, and the language required for that vehicle and jurisdiction. Drafting is per LP, so investor-specific terms are applied rather than averaged away, and every notice carries a link back to the figures it was built from.

AI agent

Maker-checker approval before issuance

This is the hard gate. No notice is issued until a second person - never the person or agent that prepared it - reviews the schedule and the drafted notices and approves them. The approver sees the calculation basis, the exceptions raised and how each was resolved. Approval is recorded against a named identity with a timestamp; refusal returns the batch with comments.

Human approval - required

Issue, track and chase

Approved notices go out through the investor portal or email channel you already use. The agent tracks delivery, logs LP queries against the relevant investor record, and monitors the funding deadline, raising a task for each LP that has not funded as the date approaches rather than after it has passed.

AI agent

Match receipts, then approve the booking

Incoming payments are matched against expected amounts from the bank feed. Exact matches are proposed for booking; partial payments, FX differences and unidentified receipts become exceptions with the candidate matches attached. Funds are recorded as received in the ledger only after a person approves the match - the agent never books cash on its own judgement.

Human approval - required

What the audit trail holds afterwards

For each call, the record contains the commitment data the agent read and when it read it, the calculation performed per LP, every exception raised with its resolution and resolver, the exact notice content sent to each investor, the approver identity and timestamp at both gates, the delivery record, and the receipt matching with its approval. It is produced as the process runs, not assembled afterwards, and it is the same record whether the call went perfectly or needed four manual interventions.

Short version of the same process: how to automate capital call processing and notices.

NAV reporting, step by step

A NAV cycle has the same shape - gather, reconcile, review, approve, publish - but the inputs are wider and the review is where almost all the value sits. The agent's job is to make the reviewer's queue short and each item in it self-explanatory.

Pull inputs from every source system

On the cycle schedule, the agent collects trial balances and positions from the ledger, holdings and cash from the custodian, prices from the pricing feeds, FX rates, and manual inputs such as valuations for unlisted holdings, accruals and fee calculations. It records the source, timestamp and version of every input, so a figure can always be traced to the data behind it.

AI agent

Reconcile positions, cash and prices

Ledger against custodian, position by position and account by account. Cash balances against the bank feed. Prices against the primary source and a secondary where policy requires it. Everything that ties out within tolerance passes without human attention; everything else becomes a break.

AI agent

Surface breaks with the context needed to clear them

A break arrives as a task, not a line on a spreadsheet: both sides of the comparison, the size and direction of the difference, the same instrument's history over prior cycles, likely causes ranked by what the pattern resembles - unsettled trade, missed corporate action, stale price, FX timing - and the supporting documents. The reviewer resolves it, or rejects the agent's proposal and states why.

Human decision

Apply valuations, accruals and fees

The agent applies the fund's valuation policy to unlisted or hard-to-value holdings, calculates management and performance fee accruals against the terms held for that vehicle, and prepares the NAV per share or per unit. Where policy requires judgement, the agent prepares the calculation and marks it for review rather than deciding.

AI agent

Analytical review before sign-off

The NAV is compared against the prior cycle and against expectation: movement by asset class, contribution by position, fee lines, unusual entries and anything outside the fund's normal variance band. The agent presents the comparison with explanations drawn from the activity it processed; the reviewer confirms each material movement is understood.

Human decision

Maker-checker sign-off, then publication

Finalisation requires approval from an accountable person other than the preparer, with the full picture in front of them: inputs and their sources, breaks raised and how each was cleared, valuation judgements, fee calculations and the analytical review. Only after that approval is the NAV finalised and released to the investor pack, the portal and downstream reporting. An override - a manual price, a forced match, a tolerance waiver - is permitted, but it is recorded with its author, its reason and its effect on the final figure.

Human approval - required

What the audit trail holds afterwards

Each cycle leaves a complete record: every input with its source and version, every reconciliation performed and its outcome, every break with its context, resolution and resolver, every valuation judgement and fee calculation, every override with its justification, the analytical review, and the sign-off identity and timestamp. When the auditor asks how a NAV figure was reached, the answer is a query, not a project.

Short version of the same process: how to speed up NAV production and oversight.

The governance model across both

The two processes share one operating model. It is worth being precise about it, because the difference between an AI agent that can be trusted in a regulated process and one that cannot is entirely in these three properties.

The agent drafts and flags - people approve what matters

The division of labour is fixed in the process definition, not left to the model's discretion.

Activity Handled by the AI agent Always requires a human
Data gathering and reconciliation Yes - reads sources, compares, records versions No, unless a break is found
Calculation from defined rules Yes - call amounts, accruals, fees, NAV per unit Review of the result before it is used
Exception detection and triage Yes - raises, contextualises, proposes a resolution Accepting or rejecting each resolution
Drafting notices and reporting packs Yes - from approved figures and fund templates Approval before anything is sent
Valuation judgement under policy Prepares the calculation and the evidence The judgement itself
Issuing a call notice No Maker-checker approval
Booking cash as received Proposes the match Approval of the match
Finalising and publishing a NAV No Maker-checker sign-off
Overriding a control or tolerance No Named approver, with reason recorded

Maker-checker is enforced by the process, not by policy

Four-eyes approval is a property of the step, not a note in a procedure document. The preparer cannot be the approver, the approval cannot be skipped when the deadline is tight, and the process will not advance to issuance or publication without it. Where a firm needs more than two eyes on a threshold - a call above a certain size, a NAV movement beyond a variance band - that is a second gate in the same mechanism.

The audit trail is default behaviour

There is no logging setting to enable per workflow. Every automated action and every human decision is timestamped and attributed as it happens, because the record is a by-product of how the process runs rather than a feature layered on top of it. That includes the runs that went wrong, the exceptions someone overrode, and the approvals given at 23:40 on a reporting deadline.

Agents work inside the systems you already run

None of this requires a new system of record. The ledger stays the ledger, the custodian stays the custodian, the KYC provider and the investor portal stay where they are. Next Matter orchestrates across them through connectors and a typed API, which is why these processes go live in days rather than after a migration. See integrations and governance and audit for the mechanics.

Evaluating AI agents for these processes

If you are assessing any vendor - including this one - these are the questions that separate a governed process from an automation demo.

  • Where exactly is the human gate? Ask for the specific steps that cannot proceed without approval, and whether the system enforces that the approver differs from the preparer. If the answer is "it's configurable", ask what happens when someone configures it away.
  • What does an exception look like to a reviewer? A queue of alerts is not exception handling. Ask to see one break as the reviewer sees it: both sides of the comparison, the history, the proposed resolution, the supporting documents.
  • Is the audit trail produced or assembled? Ask to export the full record of a completed run - inputs, versions, decisions, approvals, overrides - without anyone preparing it first.
  • Can it show its inputs? Every figure the agent produces should be traceable to the source records and the calculation applied, at the version read at the time.
  • What happens to overrides? They should be possible, attributed, reasoned and visible in the same record as everything else - not an untracked side channel.
  • Who changes the process? If a regulator, an LP or a new fund structure changes a requirement, find out whether your operations team can amend the workflow that week or whether it enters a development queue.
  • Does it require a migration? An orchestration layer should govern the ledger and custodian you already use. A platform that first needs to become your system of record is a different, much longer project.

Next Matter is built to answer all seven the same way for every process, because governance, approvals and evidence are properties of the engine rather than of an individual workflow. More on how AI sits inside governed processes: AI and automation and AI orchestration.

Running in production today

These patterns are not hypothetical. Next Matter runs regulated client and fund operations at Ocorian, where 300+ fund specialists work in the platform daily, at Trade Republic across high-volume client operations, at b2venture across a portfolio of around 800M EUR AUM, and at Swan in embedded finance operations. The platform is SOC 2 Type II and ISO 27001 certified, with SSO/SCIM and data residency you control.

See it on your own capital call or NAV cycle

Bring a real process and we will walk through how it runs with AI agents, maker-checker approvals and a complete audit trail.

Book a demo