Infor SyteLine3 min readNetray Engineering Team

Choosing a SyteLine AI Assistant: The Questions That Separate Real from Demo

AI assistants for ERP are having a moment, and manufacturers running SyteLine are being pitched hard. Some of what is on offer is genuinely useful; some is a thin wrapper around a chatbot that will embarrass itself against a real database. The good news is that a short list of pointed questions separates the two quickly. Here is the checklist we would use, whether or not the answer ends up being our product.

Accuracy Questions: How Does It Know My Schema?

Ask how the assistant knows your tables and columns, specifically your customizations and user-defined fields, and what happens when the model generates SQL referencing something that does not exist. Weak answers involve prompting harder. Strong answers involve a compiled schema index refreshed from the live database and a validation layer that structurally rejects phantom references before execution. Then ask the definitional question: when I ask for open orders, whose definition is used, and can my team read and change it? Look for semantic views or an equivalent single source of definitions.

  • Demand a live demo against a realistic SyteLine schema, not a toy database.
  • Ask what happens when generated SQL references a nonexistent column.
  • Ask where the definition of open orders lives and who can inspect it.
  • Ask whether the dashboard number and the chat number can ever disagree.

Trust Questions: Evidence and Writes

Two questions expose a system's trust model. First: for any answer, can I see the exact SQL that ran and the tables it read? Evidence-first systems show this by default; black boxes ask to be believed. Second: what happens when a user asks the assistant to change data? The only acceptable answer involves a governed flow, previewed changes, explicit confirmation, a full audit trail, and undo, with writes disabled by default until you decide otherwise. An assistant that writes to your ERP as a conversational side effect is a liability, not a feature.

Deployment Questions: Where Does My Data Go?

Finally, the architecture question: does ERP data or anything derived from it leave your network? For manufacturers under ITAR, CMMC, or customer flow-downs, cloud-only assistants often fail security review outright. Ask whether the system deploys fully on-premise, whether local models are a first-class option, and whether air-gapped operation is supported. SyteRay's answers, for the record: fully on-prem, local models supported, air-gap supported, evidence shown with every answer, writes governed and off by default. We built it to pass our own checklist.

Frequently Asked Questions

What should a SyteLine AI assistant demo prove?

Three things: correct answers over a realistic schema including customizations, visible SQL evidence for every answer, and graceful failure, wrong or ambiguous questions should produce clarifying dialogue or validation errors, never a confidently invented number.

Should an ERP AI assistant be able to change data?

Only through governance: previewed change sets, explicit per-change confirmation, a permanent audit trail, and undo. Write capability should ship disabled by default and be enabled deliberately per deployment. Anything looser is unacceptable for a production ERP.

Do on-prem AI assistants sacrifice quality versus cloud ones?

For structured ERP work, far less than assumed. The hard part of ERP question-answering is schema knowledge and validation, not raw model brilliance, and those come from the grounding layer, which works the same on-premise. Local open-weight models handle grounded SQL generation well.

Key Takeaways

  • 1Accuracy Questions: How Does It Know My Schema?: Ask how the assistant knows your tables and columns, specifically your customizations and user-defined fields, and what happens when the model generates SQL referencing something that does not exist. Weak answers involve prompting harder.
  • 2Trust Questions: Evidence and Writes: Two questions expose a system's trust model. First: for any answer, can I see the exact SQL that ran and the tables it read? Evidence-first systems show this by default; black boxes ask to be believed.
  • 3Deployment Questions: Where Does My Data Go?: Finally, the architecture question: does ERP data or anything derived from it leave your network? For manufacturers under ITAR, CMMC, or customer flow-downs, cloud-only assistants often fail security review outright. Ask whether the system deploys fully on-premise, whether local models are a first-class option, and whether air-gapped operation is supported.

Put SyteRay through this checklist against your own data. Book an evaluation.