Introduction

Rillion AI automatically suggests account coding, cost centers, VAT codes, and approval workflows for your invoices. It learns from your invoice history and the content on each invoice, getting smarter over time. This guide covers what to do — and what to avoid — to help the AI perform at its best for your organisation.

1. How the AI works

Rillion's AI leverage two things when processing each invoice:

• Current invoice content -  supplier name, amounts, references, line descriptions, and attached PDF or XML data. Max 50 pages.
• Your historical invoice data -  how similar invoices were previously coded and routed in your organisation.

The AI balances these two signals ongoing. For most invoices, recent consistent history is a strong predictor. For invoices with clear references or structured data, invoice content takes priority.

2. Getting started

Clean up your setup before activating

Before going live with AI, make sure your Rillion environment is in good shape:

• Remove or archive old, inactive roles, users, accounts, VAT codes, dimensions. AI will not suggest inactive data. 
• Decide if your reference matching is still something you want to work with and maintain, if not, ask support to deactivate reference matching from AI. 
• Ensure your chart of accounts and objects (dimensions) are up to date. Inactive accounts and objects should be deactivated (have a valid to date), not just unused.

Set realistic expectations early on

AI accuracy improves over time. In the first weeks, you should expect:

• Lower accuracy on coding for new or infrequent suppliers - the AI has little history to work from.
• Better accuracy on recurring invoices from established suppliers.
• Workflow suggestions that may need refinement, especially if your flows are complex or involve many roles.

3. Daily use - how to handle invoices

Always correct the workflow

If the AI suggests the wrong approval flow, do not approve the invoice and add a manual note elsewhere - the AI will not learn from that.

Use 'Unknown' or 'Investigation' when a flow is wrong

If an approver receives an invoice they should not have received, they should click Unknown or Investigation to send it back — not just approve it. This teaches the AI that the flow was incorrect.

Understand the confidence indicators

The AI shows a confidence level (colour-coded) for each posting line and workflow prediction:

• Green (high confidence) — AI is confident based on strong historical patterns or clear invoice data.
• Yellow (medium confidence) — AI has made a suggestion but is less certain. Review before approving.
• Red (low confidence) — AI found little to go on. Manual review is recommended.

Important: confidence levels are a guide, not a guarantee. Even green suggestions can occasionally be wrong, especially early in the rollout or after major changes to your setup.

Use the AI explanation

Each AI suggestion includes an explanation of why it was made - which invoice data , rules or historical patterns led to the suggestion. Use this to:

• Verify that the reasoning makes sense before approving.
• Understand why the AI made an unexpected suggestion.
• Identify patterns worth correcting (e.g. if the AI is misreading a reference field).

Rematching invoices

If you update the reference on an invoice and want a fresh AI suggestion for the flow, you can rematch the invoice. 

Note: rematching applies primarily to flow suggestions -  coding suggestions are based on the invoice as originally received.

4. Improving AI accuracy over time

Keep your roles and flows up to date

AI can only suggest roles and flow proposals (flow proposals are not default configuration) that exist and are active in your system. When people change roles, go on leave, or leave the organisation:

• Deactivate roles promptly rather than leaving them active but unused.
• Update flow proposals to reflect current responsibilities (if you want to continue using reference matching and flow proposals)
• Avoid approving invoices through temporary workarounds - this teaches the AI incorrect patterns. Use rule based flow auxiliary instead.

Manage complex invoices

For invoices with many coding lines (e.g. large utility or consolidated supplier invoices), consider setting a maximum number of coding lines in your AI settings. This prevents the AI from over-splitting lines and creating extra manual work.

Improve your invoice references

The AI reads structured data on invoices — including reference fields. If invoices contain clear, structured references (e.g. cost center codes, employee numbers, project codes), the AI can use these as strong signals.

• Encourage suppliers to include consistent references in their invoices.
• Work with your AP team to ensure buyer references in Rillion are clean and current.
• Avoid free-text or inconsistent references - they can confuse the AI.

Balance between history and invoice content

The AI is designed to prefer clear invoice data (e.g. a cost center clearly printed on the invoice) over historical patterns. However:

• If history has been consistent and correct for a supplier, it is a reliable signal.
• If your organization recently changed its coding structure or workflows, old history may mislead the AI during the transition period.  

5. VAT coding

VAT code prediction follows historical pattern, AI predicts VAT from historical posting patterns for that supplier and account combination. If no historical pattern found, AI leverage the VAT found on the invoice. 

Common VAT issues to be aware of:

• Partially deductible VAT (e.g. real estate companies) - configure your settings correctly and consider a custom prompt if the AI consistently gets this wrong.
• VAT-exempt companies - ensure the VAT-exempt setting is configured at company level.
• Mixed VAT rates on the same invoice — the AI reads line-level data but may need help with complex VAT structures.

6. Custom prompts

For specific business rules that the AI cannot learn from history alone, Rillion can add custom prompts to your configuration. Examples

• Read a specific reference format and map it to an object or cost centre.
• Always use the account specified on the invoice lines. 
• Specific instructions on how to predict VAT codes
• Specific instructions on allocations

7. Monitoring performance with Analytics

Rillion provides analytics to help you track AI matching performance over time. 

Go to “Invoice summary”  Choose “Matching type” = AI matched

Use it to:

• See overall coding accuracy (Successful accounting) and flow accuracy (successful flow) percentages for the chosen time period.
• Identify specific suppliers or companies where accuracy is lower than expected.
• Track improvement over time after making changes to your setup or prompts.