Platform
Suites
Solutions
By team
Company
Book a Demo
The problem

Why finance agents need their own security

Agents act on moneyAn agent that drafts a distribution or updates wire details can move money. A person knows the consequences of getting it wrong; an agent doesn't.
Generic tools lack the contextLog analysis and generic guardrails see an action. They can't tell whether this payee, amount or account is right under the LPA and the side letters.
Too strict, and the agent is uselessBlocking every payment action keeps you safe and makes the agent pointless. Telling a valid operation from an invalid one takes domain knowledge.
Documents can carry instructionsCapital call notices, statements and emails arrive from outside. Some carry changed bank details, or hidden text written to steer an agent.
The difference

Generic checks ask if the agent is safe. We ask if the money is right.

Generic agent checks

  • Prompt injection and jailbreaks
  • Leaks of personal data and secrets
  • Tool misuse and excessive permissions
  • The same tests for every agent, from a support bot to fund accounting

Finance checks from Segmental

  • A capital call notice with changed bank details gets paid
  • A fee or NAV calculation drifts from the LPA
  • A figure is reported with no source in the documents
  • Totals don't tie out to the ledger, and the agent posts anyway
Our approach

White box, not black box

Rules you can readEvery check is a rule written in your terms and approved by your experts, not a risk score from a model you can't inspect.
Decisions you can explainWhen an action is held, you see which rule tripped and the clause or record behind it.
Knowledge that is yoursThe checks come from your own fund documents, policies and past decisions, distilled into a knowledge graph.
Where we play

Every layer above the foundation, with the loop closed

Security for finance agents isn't one control. Segmental covers the four layers above your infrastructure, and what it learns after an incident becomes a new test before the next release.

  1. 4 · Post-incidentLearn from every incidentReplay the decision trace to see whether the same input would cause the same action, trace it to its source, and turn it into new tests and a tighter rule. The next release must pass them before it ships.
  2. 3 · RespondTrip, hold, routeWhen a check trips, the payment is held and the agent drops to a narrower, rule-based path. A reviewer sees the evidence, the source and the rule that fired, then releases or rejects it. Fallbacks are designed in advance, never left to the model.
  3. 2 · RunReal-time finance checksScan inbound documents for changed bank details, spoofed senders and injected instructions. Before an action executes, check it against limits, approved payees and the LPA, require a source for every number, and make sure totals tie out to the ledger.
  4. 1 · BuildFinance evals and red-teamingEvery change to an agent runs finance evals and red-team attacks, such as a poisoned capital call notice or a spoofed payee email, as a step in your release pipeline. Results are measured in wrong postings and dollar impact, and a failure shows the field, the amount and the source line.
  5. 0 · FoundationIdentity, network, data, endpointStays with your platform vendors. Segmental is designed to work with your identity, secrets and logging tools; it is not an infrastructure provider.
What the checks know

Built from finance knowledge, not generic attack lists

The checks come from four kinds of knowledge. Together they become eval cases, red-team attacks, runtime rules and fallback playbooks.

RulesAccounting standards such as US GAAP, IFRS and Lux GAAP, ILPA guidance, and the terms of the LPA and side letters.
RecordsFund documents and statements, labeled field by field, so every answer can be checked against its source.
Failure patternsPayee-change fraud, fee leakage, misallocated capital and totals that don't reconcile.
Reviewer decisionsHow fund operations reviewers accept, fix or reject an agent's work.
With your stack

Your platforms provide the controls. Segmental provides the finance rules.

Control
Your platforms
Segmental
Bounded authority
Per-agent identity, short-lived tokens, read and write separation
Limits, approved counterparties and LPA terms for each action
Circuit breakers
Kill switches, rate limits, approval hooks
When to trip: a payee change, an unusual amount, totals that don't tie out
Observable reasoning
A tamper-evident record kept outside the agent
Evidence in finance terms: the notice, the clause, the ledger line
Safe fallbacks
Retry limits and fallback routing
A review queue, a rule-based calculation, a held payment
Examples

What the checks catch

A distribution to a new accountThe bank details differ from the investor's record and the bank's country doesn't match. The payment is held and routed for approval, with both documents attached.
A notice with changed wire instructionsAn inbound capital call notice lists new bank details. It is flagged before anyone pays, with the earlier instructions alongside.
An instruction hidden in a documentText in an inbound PDF tells the agent to send data elsewhere or change a payee. The document is flagged and the action held.
A fee outside the agreementA fee an agent calculated doesn't match the LPA and the side letters. The check trips, and the clause behind it is shown.
Questions

What security and operations teams ask us

Does this replace our security tools?

No. Identity, network, data and endpoint security stay with your platform vendors. Segmental adds the finance-specific layers above them and is designed to work with what you run. Connecting it to your identity, logging and alerting tools is part of the pilot.

What does white box mean here?

That every check is visible. Each one is a rule your experts approved, built from your own documents, and every held action shows the rule and the evidence behind it.

Can agents work without seeing sensitive data?

Yes. Segmental can redact anything you specify, such as tax IDs and wire numbers, in documents coming in and going out. Redaction is set per role, so an operations lead can see a field that an agent or an outside reviewer can't.

How do you decide an agent is safe to ship?

By what a wrong answer would cost. Each release is scored on wrong postings and dollar impact, not just a pass rate, and it doesn't ship until it passes the cases added after the last incident.

Is this available today?

It is in pilot. The core checks work today, and we extend them around each firm's own agents, rules and workflows as it joins.

How does this relate to compliance?

It uses the same knowledge. The rules that check fees, consents and payments for compliance are the ones that decide, in real time, whether an agent's action should go ahead.

Join the pilot. Bring one agent workflow that touches money.

A distribution, a payment or a change to bank details. We will show you the checks Segmental would run around it.

Talk to us about the pilot