Short answer: Agentic AI in ERP and accounting means software agents that complete whole finance tasks, such as reading supplier invoices and matching them to purchase orders, chasing collections, or flagging exceptions in the ledger, rather than just answering questions. In 2026 these agents moved from demos to shipping products at major software vendors, but they still work on a “draft and human review” model and depend entirely on the quality of the data in your system.
What is agentic AI in accounting?
The difference between three generations of AI in finance systems:
| Generation | What it does | Example |
|---|---|---|
| Traditional automation | Fixed rules executed as written | Automatic monthly accrual entry |
| Copilot | Answers and suggests on request | “Summarise Q3 expenses for me” |
| Agent | Plans, executes multiple steps, uses tools, escalates when uncertain | Reads supplier email, extracts the invoice, matches it to the PO and prepares a draft for approval |
An agent does not wait to be asked. It is given an outcome (“process incoming supplier invoices”) and decides the steps itself, within limits set by the system.
What the major vendors have actually shipped
Oracle: agents in the ERP and in the database
In October 2025, Oracle announced four finance agents embedded in Fusion Cloud ERP and EPM:
- Payables Agent: processes supplier invoices from email, portals, e-invoicing and PDFs, matches them to POs and receipts, creates accounting entries and runs tax and fraud checks before routing for approval.
- Ledger Agent: shifts accountants from pulling reports to continuous monitoring through natural-language requests, and can create adjustment journals.
- Planning Agent: variance analysis, forecasting and what-if scenarios.
- Payments Agent: evaluates early-payment and financing options and tracks payment execution with banks.
In April 2026 Oracle introduced Fusion Agentic Applications: teams of agents working toward defined outcomes, including a Collectors Workspace aimed at reducing DSO, a Claims Settlement Workspace and a Cost Accounting Close Workspace. Crucially, they operate inside the existing Fusion security framework, permissions and approval hierarchies, and surface “exceptions, tradeoffs, and decisions where human judgment materially changes the outcome.”
At the database layer, Select AI Agent in Oracle Autonomous AI Database lets developers build agents that run inside the database itself, using tools such as natural-language-to-SQL, RAG and custom PL/SQL procedures, and inheriting the database’s security and auditing controls. Oracle AI World takes place in Las Vegas from 26 to 29 October 2026, including an Oracle and OpenAI hackathon built on Oracle AI Database.
Microsoft Business Central: a payables agent with clear limits
Microsoft’s official documentation for the Business Central Payables Agent is a good picture of what a realistic agent looks like today:
- It monitors a dedicated mailbox, reads PDF attachments, matches vendors and purchase orders, and creates drafts only. It does not post invoices.
- It runs under its own agent user identity, so its actions are attributed in the audit trail.
- It stops and asks for help when uncertain, and any new vendors it creates stay blocked until a user reviews them.
- It has operating limits: up to 100 emails and 500 invoices per day, with PDFs of no more than 5MB and 10 pages.
- It relies on Azure OpenAI; where that service is not available locally, an administrator must allow data movement across geographies.
OpenAI: agents that use the computer
At DevDay on 29 September 2026, OpenAI announced an Agents API with computer use and multi-agent capabilities, plus Bedrock Managed Agents built with Amazon to run agents entirely within AWS. Agents will therefore not be limited to the big global ERP suites; developers can build them on top of any system through APIs.
Accounting tasks best suited to agents today
| Task | What the agent can do | Human role |
|---|---|---|
| Supplier invoices | Extract data, match to PO and receipt, prepare draft | Review, post, approve |
| Collections | Prioritise accounts, draft reminders, track promises to pay | Negotiation with key accounts, credit decisions |
| Period close | Flag exceptions and suggest next steps | Approve adjustments, apply professional judgment |
| Analysis and reporting | Natural-language questions, variance analysis | Interpret and decide |
| Expenses | Turn receipts into expense reports | Set policy and approve exceptions |
What this means for Saudi companies
E-invoicing is a hidden advantage
Saudi businesses compliant with ZATCA e-invoicing Phase 2 already hold structured invoices in a standard format. That makes much of the “read the invoice” work easier and more accurate than in markets still relying on PDFs and images.
Arabic is not a detail
Many global agents are validated in specific languages. Microsoft’s own documentation notes that its agent can run in other languages but accuracy may vary. Test any agent on real Arabic invoices from your suppliers before relying on it.
Where the data is processed
Some agents process data in cloud regions outside the Kingdom. Review this against the Personal Data Protection Law and your sector’s policies.
Data first
Microsoft’s documentation also notes that agent suggestions improve with historical data, and companies with little history get fewer suggestions. In short: a smart agent on messy data produces messy results.
A four-stage roadmap for a mid-sized Saudi business
- Consolidate: bring accounting, purchasing, inventory and sales into one ERP with a unified chart of accounts and clear cost centres.
- Comply: make sure you are fully e-invoicing compliant; it is your cleanest data source.
- Permissions: review who can post, approve and pay. Any agent will inherit this model.
- Limited pilot: start with one agent on one task, such as supplier invoices, in draft-only mode, then measure time saved and error rates.
Common mistakes when adopting agents
- Starting with the hardest task: many pilot agents on year-end close or strategic analysis first. Start with a repetitive, measurable task instead.
- Granting admin rights: running an agent under an all-powerful user removes every safeguard. Give it its own restricted user.
- Measuring excitement, not results: define clear metrics before the pilot, such as time per invoice and the share of drafts needing correction.
- Neglecting master data: duplicate supplier records or a messy chart of accounts make agent suggestions wrong, however smart the agent.
- Ignoring the audit trail: if you cannot tell what the agent did and when, you cannot defend your books to an external auditor or ZATCA.
Where Riv-ERP fits
Riv-ERP is a Saudi-made ERP built on Oracle Database and Oracle APEX, covering accounting, cost centres, sales, purchasing, multi-warehouse inventory, POS, payroll, manufacturing, cheques and sales reps. It is compliant with ZATCA e-invoicing Phases 1 and 2 and offers an exclusive integration with Salla and Zid stores. It already includes AI features: smart reports and analytics, demand and inventory forecasting, an in-system chat assistant, and automated journal entries and reconciliations, all on a foundation of unified data, permissions and approval workflows. If you have an AI tool or agent you want to connect to your data, the Riv-ERP team can assess an API integration, and also builds custom solutions in .NET and C#.
Request a Riv-ERP quote and get your data agent-ready
Sources
- Oracle: AI agents help finance leaders accelerate insights
- Oracle: Fusion Agentic Applications for finance and supply chain
- Microsoft Learn: Payables Agent FAQ
- Oracle Docs: About Select AI Agent
Related Articles
Frequently Asked Questions
What does agentic AI mean in accounting?
Software agents given an outcome, such as “process incoming supplier invoices”, that plan and execute multiple steps using system tools and escalate unclear cases to a person, instead of only answering questions.
Do agents post entries and invoices on their own?
In most current products, no. Business Central’s Payables Agent creates drafts only and does not post invoices, and Oracle’s agents operate within existing approval hierarchies and escalate exceptions.
Which accounting task should we pilot first?
Supplier invoice processing. It is repetitive, rule-based and reviewable, and you can easily measure time saved and error rates.
Do global agents handle Arabic invoices well?
Not necessarily. Some are validated only in specific languages and accuracy can vary elsewhere, so test on a real sample of your Arabic supplier invoices first.
How does e-invoicing help with agent adoption?
ZATCA Phase 2 invoices are structured in a standard format, so agents can read and match them more accurately than paper or scanned invoices.
When is Oracle AI World 2026?
From 26 to 29 October 2026 in Las Vegas, including a joint Oracle and OpenAI hackathon combining Codex with Oracle AI Database.
