Supported API
Prefer documented endpoints with scoped service identities, stable contracts, and auditable responses.
Enterprise integration guide
How should an AI agent connect to an ERP?
The safe pattern is to preserve the ERP as the system of record, give the agent the smallest necessary data and tool surface, and separate read, recommendation, draft, and confirmed-write authority.
By Onyx Devs Lab · Published · Updated
Preserve the source of truthDo not duplicate ownership silently
Separate authorityRead, recommend, draft, confirm
Keep evidenceLog inputs, tools, rules, and outcomes
Choose the integration surface
The connector should match available interfaces and business risk.
Read replica or export
Use controlled copies for analytics when operational load or vendor access makes direct reads unsuitable.
Controlled interface automation
Treat browser or desktop mediation as a constrained last-mile adapter with explicit monitoring and takeover.
Design the authority model
A useful agent is not automatically entitled to act.
Read-only analysis
Natural-language questions can produce evidence-linked answers without creating business mutations.
Draft and recommend
The agent prepares a transaction or decision for a named operator to review.
Confirmed write
Consequential actions require validation, idempotency, permission checks, confirmation, and post-action verification.
Delivery method
Map systems of record
Identify ownership for customers, products, inventory, finance, and workflow state.
Define tool contracts
Specify inputs, outputs, permissions, failure behavior, and audit fields.
Test with real exceptions
Evaluate missing data, conflicting records, stale state, retries, and unauthorised requests.
Release by authority tier
Start read-only, then graduate specific actions only when evidence supports it.
Frequently asked questions
Does ERP integration require replacing the ERP?
No. A separate AI and data layer can add retrieval, recommendations, and workflows while the ERP remains authoritative.
What if the ERP has no usable API?
Possible alternatives include approved exports, read replicas, event feeds, or controlled interface automation. Each has different freshness, reliability, and governance trade-offs.
Should an AI agent write directly to production records?
Only narrowly defined actions should gain write authority, with server-side permission checks, validation, idempotency, explicit confirmation where needed, and outcome verification.
Start with one concrete operating problem.
We will first clarify the workflow, evidence, data boundary, and outcome standard before recommending the next step.
Contact Onyx