AI agents for ERP: why process histories matter to model builders

What are AI agents for ERP?

An AI agent for ERP is software that does a task end to end inside or alongside an ERP: reconcile a vendor invoice to a purchase order, chase a missing approval, flag a mismatched receipt, or propose a journal entry for review. It differs from a chat assistant that answers questions about the system.

The shift from answering to acting is why the data question changes. A model that explains what a three-way match is needs textbooks. An agent that resolves one in a particular company needs examples of how it was actually resolved.

What data does an ERP agent need?

It needs sequences, not snapshots. A balance tells you where a company ended up; the history tells you how it got there and who decided.

Most of this is invisible outside the company. Documentation describes the intended process; the ledger shows the real one.

Why is this data scarce?

Public text is finite, and the part describing real internal work is thinner still. Researchers at Epoch AI have estimated the effective stock of public human-generated text and projected that, if trends continue, language models will fully use it between 2026 and 2032. It is a forecast with wide uncertainty, and it concerns public text overall, not ERP records specifically, but it helps explain why developers look for permissioned sources.

ERP histories are scarce for three further reasons:

1. They are private by default and never published.
2. Companies trim history during upgrades, so older years disappear.
3. Rights are layered: some content belongs to the company, some to its customers or vendors.

How do rights and privacy limit what can be licensed?

Only records the company created and has the right to license are in scope. Customer and vendor identities, personal data and anything under a restrictive contract need handling rules agreed before work begins.

For consultants, the confidentiality question is practical: what you may say about a client's system without breaching your engagement terms. The guide on ERP implementation SOW confidentiality covers it, and the question on IT providers sharing client data addresses the provider side.

How is licensing different from giving an agent live access?

Live access through connectors lets an agent read current data inside a session. Licensing provides a defined dataset for training or evaluation under agreed terms. The comparison of MCP data access and data licensing sets them side by side, and AI agents in IT service management shows the same logic in another domain.

What does it mean for an ERP consultant?

You see the system landscape, the upgrade calendar and the archive decisions before most advisers do. That is also when a company can still choose to preserve history.

An introduction fits companies with 50+ full-time employees at peak (contractors excluded), several years of documented operations, rights to license the material and an authorized sponsor. Freshness can matter to buyers too; see why data freshness matters. The referral page for ERP consultants covers the role.

Limits and open questions

• Demand for any particular dataset depends on what buyers are building at the time, and no outcome is promised.
• Sources on agent training data are still developing, and published estimates change as new research arrives.
• Many companies have thin or deleted ERP history, which rules them out.
• Rights can be unclear when records mix customer content with the company's own.

What does an agent training or evaluation set look like in practice?

Builders usually want two kinds of material. Training material shows many examples of a task done correctly and incorrectly. Evaluation material is held back to test whether an agent handles a case it has not seen.

The practical consequence is that breadth and age matter. A company whose ERP is linked to CRM, support and email records gives a buyer the surrounding context for each decision, which a ledger alone cannot.

What should a consultant ask before raising it?

1. How many years of transactions exist, and are they all in one system or spread across a legacy and a current one?
2. Does an owner, CEO or CFO exist who can authorize a review?
3. Is the data mostly the company's own, or does it carry customer and vendor obligations?
4. Will history be preserved through the next upgrade?

None of these questions needs you to view a record. They are the same questions you ask when scoping a migration.

Which ERP footprints tend to be strongest?

Not every ERP is equally useful. The strongest footprints share a few traits, and a consultant can spot them from the project file alone.

• A single instance that ran for many years, rather than several short-lived ones.
• Approval workflows that were actually configured and used, not bypassed by email.
• Custom fields and rules that reflect how the business really works.
• Connected systems nearby, such as CRM, procurement portals and shared drives.
• An owner who treats the history as a company asset and has kept it.

Companies with thin configuration, frequent re-implementations or mostly paper-based approvals usually offer less. Say so plainly if that is the case; it saves everyone time and protects your credibility with the owner.

How rewards work

Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company. The reward is paid only after the buyer pays and SourceX receives its fee, and no reward is guaranteed. Partners never export, upload or describe confidential records. See the program terms.

Next step

If a client has years of ERP history and an owner open to a conversation, register as a partner and introduce them. You can also map candidates with the network opportunity finder or check who qualifies.