Seven signs your ERP data is not ready for Copilot or AI Agents

Introduction

Many organizations begin their Copilot or AI-agent plans by identifying processes to automate or deciding which capabilities to license. A more useful starting point is the condition of the ERP data those tools will rely on. An agent working with duplicate records, incomplete fields, stale integrations, or conflicting business definitions can produce an answer that sounds credible yet is incorrect. This article outlines seven technical signs that an ERP data foundation may need attention before wider AI adoption, along with practical checks teams can use to assess their readiness.

 1. Duplicate customers, vendors, products, or projects remain active

Duplicate master records create ambiguity when an agent aggregates information or takes action. The same customer might have several account IDs, while a single vendor could appear in multiple records with the same tax registration or bank account.

Why it matters: An agent could display an incomplete order history, overlook invoices, or incorrectly combine information from duplicate records.

How to check	Compare active records using tax registration numbers, email, telephone, bank account, manufacturer part number, and normalized legal name. Measure duplicates by entity and legal entity, including records introduced during migration.
The fix	Define matching rules for each master-data domain and assign ownership for merges, corrections, and deactivations. Resolve identity issues before building AI scenarios on the data.
Practical check	Produce a duplicate report for your five most important master-data entities.

 2. Critical fields are empty, inconsistent, or filled with defaults

ERP data can meet database rules yet remain unsuitable for business analysis. Transactions may save successfully even when important dimensions or classifications are blank, generic, or incorrect.

Why it matters: An agent may struggle to answer questions about budgets, warehouse performance, projects, or regional sales while still providing a complete answer.

How to check	Measure null, blank, and default values, invalid dimension combinations, and completeness across companies, sites, warehouses, and transaction types. Focus on legal entities, business units, financial dimensions, warehouses, product categories, tax groups, currencies, payment terms, projects, and regions.
The fix	Profile data against the requirements of the specific AI use case rather than assessing every ERP field. A field that is database-valid does not mean it is business-ready.
Practical check	Score field completeness for one proposed AI use case and identify the five fields with the greatest business impact.

3. ERP codes do not map cleanly to connected systems

AI scenarios often rely on data from multiple systems. Customer IDs may differ between ERP and CRM, product codes between ERP and e-commerce, and warehouse codes between ERP and WMS. One-to-many mappings, deprecated codes, unmapped new values, and inconsistent country, currency, tax, or status values add further complexity.

Why it matters: An agent may incorrectly combine information or rely on a single system, yet still appear to have a complete enterprise view.

How to check	Review unmatched and orphaned records, invalid reference values, rejected integration records, source-to-target record counts, and one-to-many or many-to-one mappings.
The fix	Maintain a controlled mapping catalogue covering source and target values, business meaning, owner, transformation rules, and exception handling.
Practical check	Trace key mappings for one cross-system process such as order-to-cash or procure-to-pay.

4. Integration failures and data latency are not measured

A connected system is not necessarily a synchronized one. Data can be delayed, rejected, duplicated, or partially transferred even when an interface appears to be running normally. That is why monitoring, alerts, and reconciliation should be designed into your Dynamics 365 integration architecture rather than added after go-live.

Why it matters: An agent answering questions about inventory, unpaid invoices, or orders may rely on stale information. If the agent can take action, outdated data can also lead to an operational error.

How to check	Review failed messages, dead-letter queues, retries, pending records, validation and mapping errors, source-to-target timestamps, reconciliation differences, and unresolved failures. Track failure rate, queue age, data lag, and processing against agreed service levels.
The fix	Define freshness requirements for each AI use case. A monthly report may tolerate daily data, while inventory allocation or cash application may require much fresher information.
Practical check	Create an integration health report covering failure count, queue age, latency, and reconciliation status.

 5. Related transactions cannot be joined reliably

Useful ERP questions usually require a chain of related records rather than a single transaction:

Sales order → Shipment → Invoice

Purchase order → Receipt → Vendor invoice → Payment

Problems arise when these links are missing, broken, duplicated, or inconsistent. An invoice may not link to its source order, a receipt may not reference its purchase order, or a payment may not link to the invoice it paid.

Why it matters: An agent may retrieve individual records without understanding the history behind them, thereby distorting answers about unpaid invoices, incomplete deliveries, open commitments, or inventory discrepancies.

How to check	Look for orphaned records, invalid references, invoices without expected source orders, receipts without purchase orders, payments not linked to invoices, duplicate document numbers, and open downstream transactions against closed documents.
The fix	Test relationships through representative end-to-end business scenarios rather than checking individual records alone.
Practical check	Trace three important processes from their starting transaction through to their financial or operational outcome.

6. The same KPI has different definitions across the organization

AI cannot provide a consistent answer when different teams use different definitions of the same metric. Revenue, gross margin, available inventory, overdue invoices, active customers, and on-time delivery can all yield conflicting answers when their underlying rules differ.

Why it matters: An agent can return a technically grounded answer that conflicts with either the CFO’s report or the operations dashboard.

How to check	Compare definitions across ERP reports, Power BI models, data warehouses, SQL procedures, Excel reports, custom APIs, and Copilot or agent instructions. Check differences in dates, currencies, legal entities, cancellations, intercompany activity, and actual versus forecast values.
The fix	Create a semantic catalogue for the initial AI use case. Each important KPI should have an agreed definition, formula, source, filters, owner, refresh frequency, and exception rules.
Practical check	Compare the definitions of your five most important AI-related KPIs across every reporting layer.

7. AI-used data cannot be traced or governed

Before an agent uses business data or takes action, an organization should know the information’s source, how it was transformed, who can access it, and what the agent is permitted to do.

Microsoft’s current Dynamics 365 guidance reinforces this principle. Finance and operations security continues to govern authentication, authorization, business logic, and data access when agents interact with ERP data. Our guide to implementing Copilot in Dynamics 365 Finance and Operations explains how to build audit-ready governance around these controls.

Why it matters: Without traceability and appropriate access controls, organizations may struggle to explain an AI-generated result, investigate an incorrect recommendation, or demonstrate that sensitive information was handled properly.

How to check	For important data, identify the source system and field, transformation logic, last refresh, owner, security rule, and whether the value is calculated, estimated, or manually changed. Confirm whether the agent can read, recommend, create, update, or approve.
The fix	Treat lineage, security, auditability, and approval thresholds as part of the AI implementation scope. Confirm agent activity logging and appropriate controls for high-risk actions.
Practical check	Perform a field-level traceability test and access review for one proposed agent before connecting it to live ERP data.

How LevelShift helps organizations prepare for ERP AI

LevelShift helps organizations assess ERP data readiness before committing to broader Copilot or AI-agent adoption. The work can begin with one priority process, such as procure-to-pay, order-to-cash, inventory management, or financial reporting, rather than attempting to assess the entire ERP environment at once. A pallet management provider, for example, moved to fully automated vendor invoicing on Dynamics 365 Finance and Operations after replacing siloed systems and manual steps.

The assessment can examine:

• Master and transactional data quality
• Transaction relationships and business rules
• Integration mappings and data freshness
• KPI and reporting definitions
• Security roles and access boundaries
• Audit and traceability requirements

For organizations using Dynamics 365 Finance and Operations or Business Central, this may include reviewing data entities, Dataverse, third-party integrations, Power Platform components, reporting models, and agent permissions. The result is a prioritized roadmap that shows what can proceed, what requires correction, and where a controlled pilot can deliver value.

Conclusion

ERP data does not need to be perfect before an organization begins exploring Copilot or AI agents. However, it must be sufficiently complete, consistent, current, connected, secure, and traceable for the intended use case. A reliable data foundation makes AI easier to scale.

That makes AI readiness a practical exercise rather than a reason to delay adoption. Organizations can start with a single business process, identify the data and governance issues that matter to it, and address those gaps before expanding the scope. Finance is often a good first candidate, and our Dynamics 365 Finance and Operations automation guide shows where Copilot adds value once the data is ready.

As those foundations improve, the organization can introduce AI into additional processes, supported by clearer expectations, stronger controls, and greater confidence in the information underpinning its outputs.

The objective is not to create an ERP environment that is perfect for AI. It is to create one that is reliable enough for AI to be useful, measurable, and trusted.

FAQs

1. Does ERP data need to be perfect before using Copilot?

No. It needs to be sufficiently reliable for the specific use case, with known limitations and appropriate controls in place.

2. What is the first data-quality issue to check?

Start with duplicate master records and the critical fields required by the proposed AI use case.

3. How does stale integration data affect AI agents?

An agent may produce an outdated answer or take an action based on information that no longer reflects the current business state.

4. Can Copilot use data that a user cannot access?

Microsoft states that Copilot adheres to existing data permissions and uses only data available to the user.

5. Where should an organization start its AI-readiness assessment?

Choose one high-value business process, define the information the agent needs, and assess its data, integrations, relationships, definitions, and security before expanding further.