> Site index: [llms.txt](https://ortelian.com/llms.txt) with all pages and descriptions.

# AI agents for finance teams in complex companies.

Source: https://ortelian.com/ai-agents-for-finance/
Markdown: https://ortelian.com/ai-agents-for-finance.md

AI agents for finance teams that match invoices, prepare the close and draft variance commentary from a shared world model. Your team keeps sign-off.

AI agents for finance take the matching, checking and first explanations around the numbers, and leave approvals and sign-off with your finance team. They work from a shared world model of your business, so every figure an agent uses links back to the entry or document it came from.

They are built for finance leaders at complex companies, where the close, payables, receivables and reporting depend on facts spread across the ERP, bank files, contracts, email and spreadsheets. For other functions, see [AI agent examples](https://ortelian.com/ai-agent-examples/).

[Talk to us](https://ortelian.com/contact/)

[How we work](https://ortelian.com/how-we-work/)

Who it's for

## Finance teams that reconcile by hand.

Ortelian fits when finance work depends on context no single system holds: the ledger in the ERP, the bank statement, the purchase order, the contract amendment that changed a price, the email in which a customer agreed a credit. The people in finance know how those pieces connect. Most of that knowledge lives in their heads and in spreadsheets they keep themselves.

Finance is no longer late to AI. In a Gartner survey of 121 finance leaders in June 2024, 58% of finance functions were using AI, up 21 points from 2023 ([Gartner, September 2024](https://www.gartner.com/en/newsroom/press-releases/2024-09-11-gartner-survey-shows-58-percent-of-finance-functions-use-ai-in-2024)). Gartner predicted at the time that 90% of finance functions would deploy at least one AI-enabled technology solution by 2026, with fewer than 10% seeing headcount reductions ([Gartner, September 2024](https://www.gartner.com/en/newsroom/press-releases/2024-09-12-gartner-predicts-that-90-percent-of-finance-functions-will-deploy-at-least-one-ai-enabled-tech-solution-by-2026)). The open questions are which work to start with and how to keep control.

If your team can show what a correct reconciliation, a clean variance explanation or a well-handled invoice exception looks like on real examples, you have a standard worth building to.

Where agents start

## What can AI agents do in finance?

These are the finance jobs we look at first. The audit decides which one comes first for you. They run in the tools your team already uses. Access to the platform comes with the engagement; there is no self-serve product and no public API.

- Invoice and purchase order matching

  An agent checks each supplier invoice against the purchase order, the goods receipt and the contract terms. Invoices that agree on supplier, amount and quantity move on to your normal approval flow. Exceptions go to the person who handles them, with the documents side by side and the difference marked.
- Month-end close prep

  Before the close, an agent gathers balances, open items and the bank, ERP and subledger transactions that should agree. It matches what it can and explains the differences it understands. The rest is flagged for review with the evidence attached. Your team reviews and posts.
- Variance commentary

  For the monthly management report, an agent builds the figures from the same sources each time, notes what moved against budget or the last period and drafts the first explanation. Each number links to the entries behind it. The controller edits and signs off.
- Collections prep

  An agent pulls overdue invoices together with what the customer said in email, open disputes and the agreed payment terms, then drafts the next step for each account. Anything sent to a customer waits for a person.

Controls

## What finance needs before an agent touches the numbers.

Finance works under stricter rules than most functions, and the agents follow them. These are the defaults. The exact limits are agreed in writing before anything goes live. The mechanisms are set out in [approvals and control for AI agents](https://ortelian.com/notes/approvals-and-control-for-ai-agents/) and [AI agent governance](https://ortelian.com/ai-agent-governance/).

- Read first, write only where agreed

  Agents read the ledger, bank files and documents. They write back only to the fields and systems you agree, such as a match status or a draft comment.
- No payments or postings by default

  Releasing a payment, posting a journal entry or sending anything to a customer, supplier or bank waits for a person. The agent prepares and a person approves.
- Separation of duties kept

  The agent that prepared an exception never approves it. Approval rights follow the roles you already have.
- A trail an auditor can follow

  Every run records which facts and documents the agent used and what it proposed. Any figure or flagged exception can be traced back to the entry, invoice or email behind it.

Why it works

## Agents need a map of the business, not another window.

Finance work breaks at the joins between systems: the invoice in the payables inbox, the order in the ERP, the price agreed in a contract amendment, the customer that appears under two names. An agent that only has tools rebuilds those joins on every run and loses them when the run ends. Ortelian builds a [world model for knowledge work](https://ortelian.com/notes/what-is-a-world-model-for-knowledge-work/): your systems of record joined with outside sources, with every fact keeping its source. Agents [work from that map first and use tools second](https://ortelian.com/notes/map-first-tools-second/), and each agent gets only the tools its job needs. See the [platform](https://ortelian.com/platform/).

The platform is hosted in the EU. Your team keeps working in the ERP and tools it already uses. Ortelian hosts and operates the platform as part of the engagement.

*Diagram: A code box holding a file tree. Everything the work needs is inside it.*

An example

## An invoice that does not match its quote.

This is an illustration, not a client case. A manufacturer quotes a customer a price with a volume discount. The order comes in at a different quantity, production ships in two deliveries and the invoice goes out at list price. Today someone in finance finds out weeks later, when the customer disputes the invoice.

With the quote, the order, the deliveries and the invoice joined in one world model, an agent checks each invoice before it goes out against what was quoted and delivered. It flags the mismatch with the documents and the agreed terms, and drafts the correction. A person approves it. The same joins later serve margin reporting and collections, so the second job costs less to build than the first. More on that sequence in [how to build a world model for your company](https://ortelian.com/notes/how-to-build-a-world-model-for-your-company/).

How an engagement runs

## Audit, redesign, build, then stay.

A full engagement starts with a two-week audit on site that maps how the finance work actually runs, finds the constraint and [redesigns the process](https://ortelian.com/notes/redesign-the-work-before-you-automate-it/) with your team around what AI can now do. Before anything is built, we [agree the standard on real examples](https://ortelian.com/notes/agree-the-standard-before-you-build/): accuracy, cost, time and human effort. Then we build, test and deploy in two-week cycles. By the end of the first cycle a first part of your world model is live and usable by your team, including from your own AI tools such as [ChatGPT Enterprise](https://ortelian.com/ortelian-vs-chatgpt-enterprise/) or [Microsoft Copilot](https://ortelian.com/ortelian-vs-microsoft-copilot/) through an MCP connection set up inside the engagement, with something already automated. Each further cycle extends it. We stay to host the platform and keep improving the work. Full detail on [how we work](https://ortelian.com/how-we-work/) and [how long AI implementation takes](https://ortelian.com/how-long-ai-implementation-takes/).

Not every step should become an agent. If a step has a fixed input and a fixed output, such as a recurring journal entry, it stays software or rules-based automation; see [AI agents vs RPA](https://ortelian.com/notes/ai-agents-vs-rpa/). Steps that need judgement within agreed limits go to agents. Approvals, payments and changes of direction stay with your people. To size the case first, see [how to measure AI agent ROI](https://ortelian.com/notes/how-to-measure-ai-agent-roi/).

*Diagram: An audit of about two weeks that maps the work and sets the direction, then delivery cycles of two weeks each that build, test and deploy, and keep going. The platform and world model run underneath and carry into every cycle.*

Questions

## Questions finance leaders ask.

**Does this replace our ERP or accounting software?**

No. The ERP, the accounting system and your spreadsheets keep doing their jobs and feed the world model. Agents work alongside them and write back only where you have agreed they may.

**Can an agent post entries or release payments?**

Not by default. Agents gather, match, check and draft. Postings, payment releases and anything sent to a customer, supplier or bank wait for a person. Where you agree an agent may write back, it does so within written limits and every action is recorded.

**Which finance work fits agents best?**

Work that needs judgement across several sources: matching invoices to orders and receipts, explaining reconciliation differences, drafting variance commentary and preparing collections. Steps with a fixed rule usually stay software. The audit decides where to start.

**How do auditors see what an agent did?**

Every fact in the world model keeps its source, and every agent run records the facts and documents it used and what it proposed. A reviewer can follow any figure or flagged exception back to the ledger entry, invoice or email behind it.

**What data does it need?**

Agreed access to the systems the finance work depends on, such as the ERP or accounting system, bank statements, the payables inbox, contracts and the spreadsheets the team keeps, plus any outside sources the work relies on. We agree access, model providers and data handling before we start. The platform is hosted in the EU.

**How long until the first agent is in use?**

The audit takes about two weeks on site. Delivery then runs in two-week cycles. By the end of the first cycle a first part of your world model is live and usable by your team, with something already automated. Each further cycle extends it.

---

## Let's talk about your close.

Bring the finance work that takes the most chasing and where it gets stuck. We'll work out whether an audit is worth it.

[Talk to us](https://ortelian.com/contact/)

[Or email Laurens](mailto:laurens@ortelian.com)