An ERP is the system of record for finance, HR and — in higher education — the student. It is also, almost by definition, the system most heavily depended upon, most carefully change-controlled and least tolerant of unexpected load. Putting AI on top of it is valuable precisely because that is where the data is, and risky for the same reason.
Do not query the transactional system
The first rule is that AI workloads should not run against the ERP directly. Extraction into a governed analytics platform — incremental, with lineage back to the source — isolates the ERP from analytical load and gives the AI system a foundation with quality gates and access control designed for it.
Write-back is a different category of risk
Reading ERP data through a governed platform is well understood. Writing to the ERP — posting a journal, updating a record, changing a status — is where agentic systems become genuinely consequential. Our practice is that every write goes through the ERP's supported interfaces (Integration Broker or Component Interfaces in PeopleSoft, for instance), under a service account with least privilege, with an approval gate on any action class that has not yet proven its accuracy, and with a complete log.
Customisation is the hidden complexity
Most ERP estates are heavily customised, and the customisations encode business rules nobody has documented. An AI system that ignores them produces outputs the business rejects. Understanding the customisation footprint is part of the data assessment, not an afterthought.
Done this way, AI on ERP data is some of the highest-value work available: reconciliation, intake automation, case preparation, policy assistance. Done carelessly, it is how an organisation discovers that its finance system has a support contract with conditions.
Also worth reading
Why most AI pilots never reach production — and what the successful ones do differently
The gap between a working demo and a production system is rarely the model. It is integration, entitlements, evaluation and the question of who owns it on a Tuesday afternoon in eighteen months.
PerspectiveData residency for Canadian public sector AI: what is actually required
A practical read on where regulated data can and cannot go, which deployment patterns satisfy provincial requirements, and how to document it for a privacy impact assessment.
TechnicalEvaluating AI systems: replacing spot checks with a measured baseline
How to build a maintained test set, choose scoring rubrics, calibrate an LLM-as-judge against human raters, and wire the result into your deployment pipeline.
PerspectiveThe governance work that speeds AI delivery up
Risk tiering, pre-approved control sets and platform-enforced policy. Why organisations with real AI governance ship faster than those without it.
TechnicalLakehouse or warehouse: choosing on workload rather than on fashion
An honest comparison for organisations with mixed BI and AI ambitions, including the cost profiles that rarely appear in vendor material.
Start with three weeks and a straight answer
The AI Readiness Assessment is fixed in scope, fixed in price and produces four deliverables you own — whether or not you continue with us.