Skip to content

Platform

The agentic platform under the finance work.

Connect your systems live. Turn your procedures into skills. Choose the models our agents use. Control every action and trace every run.

Book a working sessionSee security and deployment
Illustration: one piece of finance work passing through the platform. A bank reconciliation of 1010 Operating account reads NetSuite, Plaid, Ramp, Snowflake and Gmail live, follows the Bank reconciliation skill, and is worked by the Close Agent on claude-sonnet-5, a model its access group allows. It prepares journal entry JE-4471 for $12,480.00, which waits for Dana Ruiz, Controller, to approve, and the run is traced: three systems, six tool calls, 38,410 tokens and $0.04.

01Connected systems

Works in the systems your finance team already runs.

Agents query finance systems when the work needs them. Reads happen at the source. Writes are limited to the actions you allow, and ERP writes need a person's approval.

Illustration: the Close Agent at the centre of the systems it works in. It reads live from NetSuite, Plaid, Ramp, Snowflake and Gmail, and acts in NetSuite, Salesforce and Slack: the journal entry, the invoice attachment and the Slack update are approved, and the CRM update waits for approval.

Works across your finance stack.

ERP, banking, spend, data, CRM, communication and document systems.

Explore all integrations
  • NetSuite
  • SAP S/4HANA
  • Plaid
  • Ramp
  • Snowflake
  • Salesforce
  • Gmail
  • Outlook
  • Slack
  • SharePoint

02Skills

Your processes become skills. Every agent follows them.

A skill is a finance procedure written for an agent: when it applies, what to do, which systems it works in and what must be checked before the work is complete.

We ship finance skills out of the box. Your team can use them as they are, clone them or write your own.

Illustration: the skills library, with Bank reconciliation opened beside it. It is assigned to the Close Agent and used to reconcile cash accounts during close. It works in NetSuite, Plaid and Gmail, reading each, in five steps from pulling ledger activity to preparing the proposed adjustment, and the work is complete only once the ending balance ties, every unmatched item has evidence and the proposed journal entry balances.
  • Assigned, not assumedAn agent uses only the skills assigned to it, so each one follows only the methods you assign to it.
  • Loaded when neededOnly a skill's name and purpose sit in the agent's context. The full procedure loads when a task calls for it.
  • Your call who shapes themCreating and editing skills is a permission of its own, so you decide who changes how agents work.

Built-in, by area

Close and accounting

  • Bank reconciliation
  • Trial balance review
  • Journal review

AP and spend

  • Invoice check
  • AP aging review
  • Expense review

Analysis

  • Ledger questions
  • P&L variance analysis

Documents

  • Excel
  • PowerPoint
  • Word
  • PDF

03AI Gateway

Any model. Your keys. Your budget.

Every model call passes through the AI Gateway in your deployment. Choose the providers and models agents may use, run on your own keys and set spend limits by access group.

Illustration: the AI Gateway. Six model providers and your own endpoint, over the model policy of the Finance Production access group, which the Close Agent belongs to: Claude Sonnet and GPT allowed, every other model blocked, and $186 of a $1,000 monthly limit spent.
  • Bring your own keysRun on your own provider accounts and keys, switched on for your organisation during setup. Keys are encrypted at rest.
  • Switch modelsEach agent runs on the model you pick for it. Change it when a better one ships, within what its group allows.
  • Daily and monthly spend limitsGive each access group a daily and a monthly limit in dollars. Once one is reached, its agents stop running.

Keep inference in your own cloud.

When your deployment runs in your AWS account, point the gateway at a model there too, and the calls stay in your cloud as well.

  • Amazon Bedrock

    Use models through Bedrock in your own AWS account, under the access you already manage there.

  • A model you host

    Connect any OpenAI-compatible server you run, and agents call it like any other provider.

See how AI is governed

04Roles and permissions

The right authority follows every action.

Platform permissions decide what someone can configure. Finance roles decide who may prepare, review and approve the work.

Illustration: a finance team's roles. Dana Ruiz, Controller, approves; Alex Chen, a senior accountant, reviews; Mike Keller, an accountant, prepares. The person who prepared or requested a journal entry can't approve it, so approving JE-4480 is blocked for Mike Keller, who prepared it.
  • Custom rolesBuild roles from individual permissions. An admin can't grant a permission they don't hold themselves.
  • Finance rolesMark who is a Controller and who has approval rights, apart from what their platform role lets them configure.
  • Separation of dutiesWhoever prepared or requested a journal entry can't approve it, and a preparer can't review their own task.
  • Single sign-onSign in through your identity provider over SAML or OIDC. People join by invitation.
  • Service accountsGive automation API keys on the same roles as people. Acting as another user is refused unless a trust policy allows it.
See how access is secured

05Traceability

People change things. Agents do work. You can see both.

Every material change and every agent run leaves a record. See who changed what, how an agent completed the work and which approvals sat behind an action.

Activity log

Governance record

Shows who changed what.

Illustration: the activity log, newest first: journal entry JE-4471 approved by Dana Ruiz and posted to NetSuite, a change to Alex Chen's finance roles, a model provider enabled and a sign-in. No one can edit or remove an entry.
  • Every kind of changeSign-ins, members and roles, service accounts, providers, integrations, agents, skills and writes to your ERP.
  • No edits, no deletionsThe product has no way to edit or remove an entry. It only ever adds to the log.
  • Kept for a year, security for twoEntries are kept 365 days and security events 730, set for your deployment rather than by an organisation's admins.

Agent trace

Execution record

Shows how the agent got there.

Illustration: the trace of one agent run, completed in 24 seconds with 38,410 tokens for $0.04. The model call that ran it spans the run, and the tool calls inside it follow: the skill it loaded, the workspace it provisioned, a NetSuite query that failed and was made again, Plaid, Gmail and the code it ran, each with its time.
  • Every step, its own spanModel calls, tool calls, sandbox work, sub-agents and context compaction each get a span with its timing and any error.
  • Tokens and cost per stepInput, output, cached and reasoning tokens, and the cost in dollars, on every step.
  • Kept in your deploymentTraces are written to your deployment's own database and kept 30 days, their payloads 7. Nothing goes to an outside tracing service.

See our agents work through a finance process that matters to your team.

Bring us a process your team still works manually. We’ll show you how our agents would run it.

Book a working session