One agent, multiple finance skills.
AgentOne is the single agent across Aurum's FinanceOS. It sits inside the product you are already working on. It knows the context, knows you, and grows by learning new skills. It holds its own identity, acts on behalf of a named person under explicit authority, and writes nothing without human approval.
How can Aurum's AgentOne help?
From question, to request, to action, directly in seconds.
The work that slows finance down is rarely accounting. It is the queue in front of it: someone has to profile the data, decide the mapping, write the rule, build the loader, then work out why a transaction did not post. That work belongs to whoever knows the system, and it waits behind everything else they own. AgentOne does it in conversation, on the screen where the work already is, and leaves a record an auditor can read.
It works where you already are
AgentOne is embedded across all our products, on the screen where the work happens. It can already see the break, the account and the file in front of you, so you spend no time describing the context.
It signs as itself, never as you
AgentOne has its own identity in the platform. It does not borrow your login and it does not hide behind a generic service account. Every action it takes records who acted, on whose behalf, and under what authority.
Reads anywhere, writes only under approval
Reads can happen wherever you work. Regulated writes happen in the platform, after a human approves them, with the evidence and the signature attached.
Finance teams globally put AgentOne to work across their operations.
.svg?2026-03-11T09%3A48%3A43.242Z&width=3840&quality=80)
One agent, not a fleet
New products don't bring new agents. The same agent learns a new skill.
The common answer to "more AI" is more agents: one per task, each with a name, a permission set and a place you have to go to find it. That model fails twice over. The user has to know which agent to ask, and every additional agent widens the surface an auditor has to test.
AgentOne is one agent with skills. When Close and FP&A arrive, they arrive as skills the same agent learns, under the same identity and the same authority model. Each skill carries its own permissions, so learning more never means being trusted with more.
Reconciliation skills
Diagnose why items are not matching
Samples the outstanding items on both sides, computes cross-side statistics and proposes a candidate matching rule, stating the size of the sample it worked from rather than implying it looked at everything.
Author matching, validation and exception rules
Creates and assigns rules with tolerances for rounding and timing, and reviews the current exception backlog by rule.
Suggest matches and categorise exceptions
Proposes matches for items the configured rules did not pair, for a person to accept, and classifies exceptions as they are raised so the backlog can be analysed by cause rather than only by age.
Answer status questions about a reconciliation
Returns an entity's summary, linked accounts, volumes and recent operations without you navigating to it.
Reconciliation skills
Diagnose why items are not matching
Samples the outstanding items on both sides, computes cross-side statistics and proposes a candidate matching rule, stating the size of the sample it worked from rather than implying it looked at everything.
Author matching, validation and exception rules
Creates and assigns rules with tolerances for rounding and timing, and reviews the current exception backlog by rule.
Suggest matches and categorise exceptions
Proposes matches for items the configured rules did not pair, for a person to accept, and classifies exceptions as they are raised so the backlog can be analysed by cause rather than only by age.
Answer status questions about a reconciliation
Returns an entity's summary, linked accounts, volumes and recent operations without you navigating to it.
Journal and accounting skills
Explain, transaction by transaction, why items did not post
Analyses the unposted transactions of an entity and attributes each one to a cause — the most common month-end question in journal processing, answered without anyone reading logs.
Build and extend GL code mappings
Creates mappings of key-value combinations to debit and credit accounts, validating every code against your chart of accounts, refusing duplicate combinations and keeping dimensions consistent.
Create journals and author posting rules
Configures which transactions a rule posts, how they group, the amount, the date, the debit and credit treatment, memos and reversals — and rejects combinations the engine cannot honour instead of letting them fail at run time.
Run a journal and explain what the run did
Triggers a run by name, reports its status, and returns the execution history, the configuration behind it and how many transactions remain unposted per entity.
Journal and accounting skills
Explain, transaction by transaction, why items did not post
Analyses the unposted transactions of an entity and attributes each one to a cause — the most common month-end question in journal processing, answered without anyone reading logs.
Build and extend GL code mappings
Creates mappings of key-value combinations to debit and credit accounts, validating every code against your chart of accounts, refusing duplicate combinations and keeping dimensions consistent.
Create journals and author posting rules
Configures which transactions a rule posts, how they group, the amount, the date, the debit and credit treatment, memos and reversals — and rejects combinations the engine cannot honour instead of letting them fail at run time.
Run a journal and explain what the run did
Triggers a run by name, reports its status, and returns the execution history, the configuration behind it and how many transactions remain unposted per entity.
Data skills
Profile the data before anything is built on it
Reports how often each field is populated, how many distinct values it holds and what typical values look like, then lists the value combinations that genuinely occur, with counts — so a proposal fits the data you have rather than an assumed schema.
Warn about configuration that cannot work
Flags combinations too granular to maintain, and rows whose key values are empty and would never match, while it is still a design conversation rather than a posting-day failure.
Create and test connections, and build loaders
Sets up SFTP, Azure Blob and REST connections, previews a source, proposes the field mapping and creates the loader. Credentials are never accepted in conversation and never appear in a reply.
Build scheduled pipelines by describing them
Creates the connections, transformations and schedule from a plain-English description of the flow, including a workflow step that extracts structured data from PDFs and scanned documents.
Data skills
Profile the data before anything is built on it
Reports how often each field is populated, how many distinct values it holds and what typical values look like, then lists the value combinations that genuinely occur, with counts — so a proposal fits the data you have rather than an assumed schema.
Warn about configuration that cannot work
Flags combinations too granular to maintain, and rows whose key values are empty and would never match, while it is still a design conversation rather than a posting-day failure.
Create and test connections, and build loaders
Sets up SFTP, Azure Blob and REST connections, previews a source, proposes the field mapping and creates the loader. Credentials are never accepted in conversation and never appear in a reply.
Build scheduled pipelines by describing them
Creates the connections, transformations and schedule from a plain-English description of the flow, including a workflow step that extracts structured data from PDFs and scanned documents.
Skills it carries everywhere
Refer to things by their business names
Name an entity, journal, account, loader or report the way your team refers to it and the agent resolves it to the underlying record. Nobody needs to know internal identifiers.
Answer from documentation, with the topic cited
Product questions are answered by retrieving passages from indexed documentation rather than from the model's recollection, and it says when a topic is not covered instead of guessing.
Pull a figure into the conversation
Returns report results, KPIs and dashboard views, scoped to one division or aggregated across every division you may see.
Escalate when it cannot resolve something
Raises a support request from inside the conversation, to a destination configured server-side that cannot be redirected.
Skills it carries everywhere
Refer to things by their business names
Name an entity, journal, account, loader or report the way your team refers to it and the agent resolves it to the underlying record. Nobody needs to know internal identifiers.
Answer from documentation, with the topic cited
Product questions are answered by retrieving passages from indexed documentation rather than from the model's recollection, and it says when a topic is not covered instead of guessing.
Pull a figure into the conversation
Returns report results, KPIs and dashboard views, scoped to one division or aggregated across every division you may see.
Escalate when it cannot resolve something
Raises a support request from inside the conversation, to a destination configured server-side that cannot be redirected.
Identity and authority
The agent signs as itself.
Never as you.
Borrowed credentials are the fastest way to make an agent unauditable. An agent using your login produces a trail that says you posted a journal at three in the morning. A generic service account produces a trail with no human in it at all. AgentOne does neither: it holds its own identity in the platform and acts on behalf of a named person — never as them.
Three questions on every action
Who executed it, on whose behalf, and under what authority. The effective permission is the intersection of the three, never the widest of them.
Never more than the person
No agent, ours or external, can see or do more than the person it is acting for. If you cannot reach an entity, neither can an agent acting on your behalf.
Permissions per skill, not per agent
Each skill carries its own permissions, so a new skill is a new scoped capability rather than a broader agent with more reach.
No write without a human
Reads can happen on any surface. Regulated writes happen in the platform, after approval, with the evidence and the signature attached to them.
Mandates for unattended work
Autonomous work runs under a mandate that names a human. Mandates are bounded, revocable and revalidated at every execution, so a change of role suspends the work instead of outliving it.
External agents are named, not hidden
A connecting agent — Claude, Copilot, your own internal automation — is recorded as the actor in its own right. It never appears behind the user's name.
Written once, never rewritten
Every action lands in a WORM audit trail that cannot be altered after the fact.
The audit trail
Four fields on every line.
Every entry in the WORM record carries the same four fields, so the question an auditor actually asks — who did this, and on whose authority — is answered by reading a single line.
Actor
Who executed it. A person, AgentOne and the skill it used, or an external agent.
Subject
The named person, or the company, the action was taken on behalf of.
Authority
What permitted it. An approval, an authenticated session, a mandate, or a signed configuration.
Surface
Where it happened. Inside a product, in AgentOne, in the background, in an external tool, or at system level.
The audit trail
Four fields on every line.
Every entry in the WORM record carries the same four fields, so the question an auditor actually asks — who did this, and on whose authority — is answered by reading a single line.
Actor
Who executed it. A person, AgentOne and the skill it used, or an external agent.
Subject
The named person, or the company, the action was taken on behalf of.
Authority
What permitted it. An approval, an authenticated session, a mandate, or a signed configuration.
Surface
Where it happened. Inside a product, in AgentOne, in the background, in an external tool, or at system level.
Product Suite
Aurum’s FinanceOS automates the backoffice of the CFO
Automated Reconciliation Software
Where the reconciliation skills work. Matching, exception management and sign-off run on the same platform and the same audit trail, which is why the agent can move from a break to the rule that caused it without leaving the product.
Journal Entry Automation
Where the journal skills work. Posting rules, GL mappings and journals are what the agent builds and explains, so "why did this run post less than expected?" is answered transaction by transaction from the same configuration.
Financial Data Automation
Where the data skills work. Connections, transformations and schedules built from a plain-English description of the flow, with document extraction available as a workflow step rather than a separate integration.
Safeguarding Control Centre
The agent is embedded here too, under exactly the same identity and authority model — so there is no separate governance for the most sensitive position you hold.
How are AI agents used in accounting?
Automation executes set rules whereas AI agents handle the exceptions. In bank reconciliation that is the transaction that did not match, the remittance advice that arrived as a PDF, the payment service provider fee deduction that changes format each month. An agent reads what a finance professional would, forms a view on what happened within the context it is given access to, and explains its reasoning, ready for review by a human. It also enables reports and dashboards to be generated on request, so teams can access the information they need quickly and easily. Teams apply this across reconciliation, accounts payable matching, intercompany reconciliations and month-end close preparation.
Agents earn their place where work is too varied to encode, but too high volume to be completed effectively by a person. Exception management is the clearest case: the agent reads the surrounding data, proposes a cause such as a timing difference or a deducted fee, and shows how it reached it. Unstructured inputs are another, since remittance advice, bank statements and counterparty emails carry the information that dictates a match but often arrive in no fixed shape. Then there are questions that cross systems, where an answer needs reconciliation, safeguarding and data from across systems together. AI agents can also complete one-off analysis on request.
Automation and AI contribute different things but work best when they do so alongside each other. Automation gives you volume and consistency on the work you have already defined, which is what lifts auto-match rates and shortens the reconciliation cycle. AI extends coverage to the exceptions, which is where finance teams’ hours are lost. Agents resolve exceptions with visible reasoning, since an agent can explain why it reached a conclusion rather than returning a result to be reverse-engineered by the team. Capability also compounds, because an agent that learns a new skill applies it across everything it already touches. Teams using AI and automation together experience fewer days needed to close and an increased proportion of items matched without intervention.
Most finance software adds a separate assistant to each module, which leaves the user deciding which one to ask, teaching one essential context, while the others fall behind. Instead, Aurum runs one agent across the platform. As a result, new products arrive as new skills for the same agent rather than as separate assistants. Context across each element of the team’s financial automation makes the agent more capable. Finance teams shape it through each request, since the skills that get built are the ones their reconciliation, safeguarding and close work calls for.
AI agents in finance and reconciliation workflows are capable of both explaining outcomes and building out tools that enhance visibility. On explanation, you can ask why a particular match failed and get the reasoning behind it, in plain English, without tracing the logic yourself. On construction, you describe the process you want in a single sentence, and the AI agent assembles the data workflow behind it, including the recurring schedule. Because AgentOne runs across Aurum's products rather than within one, a single question can also draw on reconciliation data, ledger data and source feeds together. Through the Aurum MCP, that platform data is also reachable from the AI tools your team already uses.
Safety in the use of AI agents when handling sensitive financial data depends on accountability controls and audit capabilities. Safe AI agents in finance are built so that responsibility stays with named people. They should act on instruction but not assume identity, so any action taken by the agent is never recorded as one taken by you. Nothing should be written without a person approving it. With AgentOne, every action is captured in a log that cannot be altered afterwards, showing which steps a human took, which the agent took and who authorised them. An auditor can reconstruct any decision from that record.
Get a demo
One agent that learns. Authority that never dilutes.
See AgentOne where your team already works — in the break, in the journal run, in the safeguarding position — and see the line it leaves behind in the record.
One agent, every skill
Its own identity, never yours
No write without approval
Immutable WORM trail
See the product in action
Trusted by hundreds of companies including
