How to add AI agents to ProShop ERP
Short answer
AI agents for ProShop ERP are multi-step workers that call the ProShop API as tools: read work orders, parts, routings and quality records, reason over them, and propose changes. Keep them read-only first, route every write through a human approval step, and log each tool call.
ProShop is built for job shops and aerospace or medical machine shops that want one system for quoting, work orders, routing, tooling, inspection and quality. Because so much of the shop lives in one database, ProShop is a good place to point an agent: the context an agent needs to answer why is this job late is mostly already in the same system.
The pattern that works is not a chatbot. It is a worker with a defined job, a short list of tools, and a stop point where a person approves. Examples are a late-job triage agent that reads open work orders and operation status each morning, an NCR drafting agent that assembles the facts, and a quote-prep agent that finds similar past parts and their actual hours.
ProShop is mostly vendor-hosted, so the agent runs beside it rather than inside it. The agent runtime, the model and the audit log are what you control, and that is where design choices about privacy, cost and failure handling are made.
What teams use it for
Late job triage
Each morning an agent reads open work orders, current operation, due dates and work center load, then posts a ranked list of jobs at risk with the likely blocker (material, tooling, inspection hold, outside process).
NCR first draft
When an inspector flags a nonconformance, the agent pulls the part, work order, operation, inspection data and prior NCRs on the same part, and drafts the description for the quality manager to edit.
Quote preparation
For an incoming RFQ, the agent searches past parts with similar material, features and quantities, summarises actual versus estimated hours, and proposes a starting estimate for the estimator to correct.
Purchasing follow-up
The agent finds POs for material and outside processing that are overdue against the work order need date, drafts supplier chasers, and queues them for a buyer to send.
Certification package check
Before shipping, the agent checks that the required certs, inspection records and traveler signoffs exist for the lot, and lists what is missing rather than letting the gap surface at the customer.
Tooling and fixture readiness
For jobs starting this week, the agent compares required tools and fixtures on the routing against what is recorded as available and flags shortfalls to the programmer.
ProShop ERP integration surface
- ProShop exposes an API, documented by the vendor, that is GraphQL based in current releases and authenticated with a token obtained for an API user. Confirm the exact endpoint, schema and token flow for your version and plan with ProShop support.
- Core objects an agent will touch are parts, work orders and their operations, routings, purchase orders, customers, quotes, inspection records and NCRs. Field names vary by version and by custom fields your shop added, so introspect the schema rather than hard coding.
- Create a dedicated API user with a restricted role so the agent only sees what its job needs. The agent should never run as an administrator.
- For bulk analysis such as historical hours, built-in export and report options can feed a separate store, which keeps heavy reasoning off the live API.
- ProShop is generally delivered as a hosted service, so there is no direct database access in the usual deployment. Plan on the API and exports as the integration surface.
How it works
- 1
Pick one narrow job
Choose a single recurring task with a clear definition of done, such as late job triage. Write down inputs, the output format and who approves.
- 2
Create a restricted API user
Provision an API user in ProShop with a read-focused role and store its credentials in a secret manager, not in agent code.
- 3
Build a thin tool layer
Wrap a handful of GraphQL queries as typed tools: get_work_order, list_open_operations, get_part, get_po_status. Return trimmed, labelled fields instead of raw payloads.
- 4
Add the model and agent loop
Use a model with reliable tool calling, either an API model under your data agreement or an open-weight model served privately. Cap the number of tool calls per run.
- 5
Insert an approval gate
Anything that changes data, such as a note, a status or a PO comment, becomes a proposal shown to a user. A person approves, then a separate narrow write tool executes.
- 6
Log and replay
Store each run: the question, tool calls, returned data and final output. This is how you debug wrong answers and prove what the agent saw.
- 7
Measure and widen
Compare agent output with what staff would have done for a few weeks. Only then add a second job or grant a write permission.
Example flows
Morning late job brief
Trigger: Scheduled at 6:00 each weekday
- Agent lists open work orders due within 10 days.
- For each, it reads the current operation, quantity complete and any hold status.
- It checks linked material and outside process POs against need dates.
- It ranks jobs by slack and writes a short reason per job.
Result: The production manager gets a one-page brief in email or chat before the first meeting, with sources listed for each claim.
NCR drafting assist
Trigger: An inspector creates a nonconformance on a work order
- Agent fetches the work order, part revision and operation where it was found.
- It searches earlier NCRs on the same part and similar features.
- It drafts a factual description and lists open questions for the inspector.
Result: The quality manager starts from a structured draft and a list of related history instead of a blank form.
Keeping data on-prem
Because ProShop is typically hosted by the vendor, the question of on-prem applies to everything around it. Run the agent runtime, the vector or cache store, the audit log and, if needed, the model inside your own network or a private cloud, and let only the narrow API calls to ProShop leave. For ITAR or CUI work, keep prompts and retrieved shop data off public model endpoints and use an open-weight model on your own GPU. Where a customer has a self-hosted or special deployment of ProShop, the same pattern holds with the API reached over your network. Check your contract and flow-down clauses to see which data may pass to any external model service.
Security checklist
- Dedicated API user with the least role needed; no shared credentials.
- Read-only by default; writes only through approved, narrow tools.
- Full audit of prompts, tool calls and outputs, stored where you control retention.
- Strip customer drawings and export-controlled data from prompts unless the model runs in your boundary.
- Rotate tokens and monitor for unusual call volume.
Ways to get it done
Build in-house with a workflow tool and an LLM
Effort: A few weeks for a first read-only agent; ongoing prompt and tool upkeep.
Shops with a developer or technically strong ops person and one clear first use case.
ProShop's own features and exports plus scheduled reports
Effort: Days; no agent required.
When a fixed report answers the question and no reasoning over text is needed.
Partner build with Netray
Effort: Typically weeks to a pilot, scoped per use case.
Regulated shops that want a private model, approvals and audit designed in from the start.
Pitfalls
- !Letting the agent write directly to work orders or POs before you trust its accuracy.
- !Passing raw GraphQL results to the model; trim and label fields or answers drift.
- !Hard coding field names that change with custom fields or a version upgrade.
- !No definition of done, so the agent loops or produces long vague summaries.
- !Skipping logs, which makes wrong answers impossible to diagnose.
Frequently asked questions
Does ProShop have an API an agent can use?
ProShop provides an API, which in current releases is GraphQL based with token authentication. Exact endpoints, schema and access depend on your version and plan, so confirm with ProShop support and introspect the schema. Agents work through that API as a set of defined tools rather than touching a database directly.
Can an agent update work orders automatically?
It can technically, but you should not start there. Begin read-only, then allow proposals that a person approves, and only later permit narrow low-risk writes such as adding a note. Anything affecting quantity, cost or quality status should keep a human in the loop.
Which model should the agent use?
Choose by data sensitivity. For non-controlled work, a hosted frontier model with a business data agreement is simple. For ITAR, CUI or customer-restricted drawings, an open-weight model served on your own hardware keeps data inside your boundary. Test tool-calling reliability on your own tasks before deciding.
Is this overkill for a small shop?
Often a scheduled ProShop report or a simple automation is enough. Agents earn their place when the task needs reading several records and free text, such as late job triage or NCR history. Start with one job and drop it if it does not save real time.
How do we know the agent is right?
Require it to cite the records it used, keep full run logs, and compare its output with staff decisions for a few weeks. Track corrections. If it cannot show the work order, operation or PO behind a claim, treat the claim as unverified.
Want AI agents working on your ProShop ERP?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment