Leave the ERP Alone. Teach AI What “Cost” Means.
How we helped a finance team cut a five-day monthly close to 20 minutes without changing SAP, and why business definitions may be the missing layer in enterprise AI.
Every month, the finance team at a major Korean fashion platform went through the same routine.
Export data from SAP. Reconcile the numbers in Excel. Adjust costs across brands and business units. Check the results. Repeat until the managerial P&L matched how the company actually ran its business.
More than 70% of the process was manual. Four people spent roughly five days closing the books for management reporting.
The obvious solution was to change SAP. Rebuilding the process inside the ERP was estimated to take about ten times as long as our eventual implementation, at five to ten times the cost.
So we left SAP alone.
We built an AI agent that works alongside SAP and the company’s existing data platform. It reads the data, applies the finance team’s business rules, and prepares a traceable managerial P&L for human approval.
The team began using it on October 26, 2026. It passed acceptance testing on November 3, two weeks ahead of schedule.
A process that took four people five days now takes one person about 20 minutes.
The Rules Lived in Excel
A managerial P&L is often calculated differently from a company’s statutory financial statements.
Finance teams make adjustments to reflect how management wants to see the business. Costs are allocated across brands. Returns may be recognized differently for internal reporting. Some accounts are grouped into management SG&A even when the general ledger organizes them another way.
These are business decisions made by management. In many companies, they live in Excel formulas, recurring manual adjustments, and the experience of the people preparing the reports.
Before an agent can automate those decisions, someone has to make the rules explicit.
That’s where we started. Working with the finance team, we documented the adjustments one by one: which source data each rule depends on, how the calculation is applied, and what the final output needs to look like.
Once those definitions were agreed, the agent could apply them the same way every reporting period.
Two design decisions made the difference.
1. Every number has an explanation.
Each figure is linked to its business rule and its source records. If a reviewer questions a cost allocation, they can trace the result back to the data and logic behind it.
Automation still needs verification. Good automation makes verification fast. An agent that saves hours of preparation but creates hours of checking has only moved the work around.
2. The agent prepares. A person approves.
The system handles data processing, calculations, and report preparation. The finance team reviews and approves the final close.
Authority over the financial records stays with the finance team. The agent makes the existing process faster, more consistent, and easier to audit.
Why SAP and Snowflake Are Talking About Business Ontologies
Our project reflects a broader shift in enterprise AI.
In fall 2026, Snowflake introduced a preview of Business Ontology, designed to give AI agents consistent, governed business definitions. SAP also discussed its Ontology Manager for Business Data Cloud at TechEd.
The term sounds abstract. The problem is familiar to almost every enterprise.
Ask three departments to calculate gross margin and you may get three answers. One team includes certain logistics costs. Another excludes them. A third uses a management reporting definition that matches neither.
All three may be working from valid data. The disagreement isn’t about the data. It’s about what the data means.
A business ontology defines business concepts, how they relate to each other, and the rules for interpreting them. For AI agents, those definitions are essential. Without them, an agent can retrieve the right records and still calculate the wrong number. With them, the agent works from the organization’s own meaning of revenue, cost, margin, or active customer.
Should You Wait for Your Platform Vendor?
We agree with the direction SAP and Snowflake are taking. The question is whether companies need to wait for it.
In our experience, they don’t. Most companies already have systems that work well enough as systems of record. The operational pain sits in the gaps between them. Data gets exported. Rules get applied by hand. Documents get checked against spreadsheets. Someone reconciles the results before anything is approved.
Replacing the underlying systems is expensive and disruptive. A more practical starting point is a layer built around what already exists. That layer can:
- Read data and documents from existing enterprise systems.
- Maintain approved business definitions and calculation rules.
- Connect each result to its supporting evidence.
- Apply the same logic consistently across workflows.
- Record changes when business rules evolve.
- Keep people responsible for final decisions and approvals.
The ERP remains the system of record. The agent handles the work around it.
When finance changes how a cost should be allocated, the change is made in the rule layer, with a record of what changed and why. Nobody has to open an ERP customization project.
The Same Problem Shows Up in Trade
Finance is one of many functions where business logic gets trapped between systems, documents, and spreadsheets.
Consider a mid-sized US importer running on QuickBooks and Excel.
A purchase order specifies one quantity. The commercial invoice lists another. The packing list uses different units, and the bill of lading reflects what actually shipped. Someone has to compare the documents, decide which differences matter, and act.
The work is reading the documents together and knowing which gaps matter.
That’s the approach we take across NURIE AI’s products.
TraderSage: AI agents for trade documentation and reconciliation
TraderSage reviews trade documents, compares quantities and terms, and flags discrepancies that need attention. It works alongside the accounting and operational tools a company already uses.
RegSage: AI agents for product regulatory compliance
RegSage applies the same approach to regulatory work. Each requirement is connected to supporting documents and traced back to its regulatory source. A reviewer can see what the agent concluded, and which requirement and evidence informed that conclusion.
VaultSage: The document intelligence layer underneath
VaultSage reads documents, structures their contents, and connects related information, so AI agents work from a complete view of the underlying records.
Different industries, different workflows, one principle: AI becomes useful once it knows the rules behind the work.
Start With the Report Everyone Dreads
If you’re considering AI agents for finance or operations, start with one process.
A month-end close that takes several days. A reconciliation full of spreadsheet adjustments. A document review where the same checks happen every week.
Look closely at what people do by hand. Which decisions repeat? Which rules are buried in spreadsheets? Which numbers or documents have to be checked against each other?
Write those rules down. Connect them to the source data. Make the results traceable. Keep a person responsible for approval.
Our finance project began with a team spending five days every month on one report. It ended with a process that takes about 20 minutes. The improvement came from making the business logic explicit enough for an agent to use.
That’s how we build: one real operational problem at a time, next to the people who understand it best.
If your close, reconciliations, or trade documents still run on downloads, spreadsheets, and manual checks, write to us.
Contact NURIE AI: partnership@nurie.ai
❓Frequently Asked Questions
Can AI automate month-end closing without replacing SAP?
Yes. AI agents can automate many of the data preparation, reconciliation, adjustment, and reporting tasks in a month-end close without replacing SAP. They read data from existing systems, apply approved business rules, and prepare reports for human review. SAP remains the system of record, and the finance team keeps final approval.
What is a business ontology in enterprise AI?
A business ontology is a structured representation of business concepts, definitions, relationships, and rules. It tells AI agents how an organization defines terms such as revenue, cost, gross margin, and active customer, so agents interpret data consistently and produce results aligned with the company’s reporting standards.
Why do AI agents need business definitions?
The same data can be interpreted differently across departments and workflows. Finance and operations, for example, may define margin differently. Approved business rules let agents apply the correct calculation and explain how each result was produced.
How is an AI agent different from traditional ERP automation?
Traditional ERP automation runs inside configured system workflows. An AI agent can work across ERP data, spreadsheets, documents, and other systems while applying business rules in a separate layer. This helps automate cross-system processes without extensive ERP customization, although integration and data governance are still required.
Can AI-generated financial reports be audited?
Yes, when the system is designed for traceability. Each figure should be connected to its source records, calculation rules, and any relevant changes, so reviewers can verify how results were produced before approving the final report. Auditability depends on the controls and evidence the implementation retains.
Can AI agents work with QuickBooks and Excel?
Yes. AI agents can use data exported from or integrated with tools such as QuickBooks and Excel. For importers, they can also review invoices, packing lists, and bills of lading to find inconsistencies. Exact capabilities depend on available integrations, document formats, and the company’s workflow.