Kinetic + fabrication shops
Epicor Kinetic AI for Metal Fabrication: Jobs, Heat Numbers and Outside Processing in Plain English
Short answer
AI for Epicor Kinetic in a metal fabrication shop lets estimators, planners and quality leads ask questions about jobs, materials, heat numbers, outside processing and quotes in plain English and get answers grounded in live Kinetic data, with the query shown. It runs on-prem or in a private cloud, is read-only by default, and respects Kinetic security so customer drawings, pricing and material certs stay inside your boundary.
- ERP
- Epicor Kinetic
- Industries
- Metal Fabrication, Sheet Metal, Job Shops
- Written for
- Operations Manager
A fabrication shop is a pile of small decisions made against a moving background: a laser that is booked solid, a press brake with a changeover that eats the morning, a coil of 304 that arrived with the wrong mill cert, and a powder coat vendor who is running a week behind. Epicor Kinetic holds the facts behind all of it, in job heads, job materials, job operations, part lots and purchase orders, but getting a straight answer usually means a BAQ someone wrote two years ago or an export to Excel.
The questions that actually move the day are cross-cutting. Which open jobs are waiting on the same outside process? Which jobs consumed material from a heat that a customer has just flagged? Which quotes were won at a margin far below what the actual job cost turned out to be? Each one touches four or five Kinetic tables and a naming convention that only your senior planner remembers.
Kinetic's own search, Data Discovery and BAQ tooling are good at known questions. They are less good at the open question asked at 7:15 on a Tuesday by a supervisor who does not know which table holds the answer. A language layer over the same data, with the same security, fills that gap without replacing anything you already trust.
This page covers what that looks like in a fabrication environment: the Kinetic objects involved, how a private model is deployed beside your database, where it helps estimators, planners and quality staff, and where a plain BAQ or dashboard is still the better tool.
What usually gets in the way
The problems we hear most from operations manager teams running Epicor Kinetic.
Quoting from history is manual
Estimators price new work by remembering similar parts. The actual cost of past jobs sits in job material, labor and operation records, but comparing a new DXF or STEP part to history means opening old jobs one at a time.
Material traceability is slow when it matters
When a customer asks which heat number went into which shipped lot, answering means walking part lots, material issues and shipments by hand. Mill test reports often live as scanned PDFs outside the ERP.
Outside processing hides late jobs
Galvanizing, plating, anodizing and powder coat are subcontract operations tracked through purchase orders and job operations. A vendor slipping three days is visible only if someone cross-checks PO receipts against the job schedule.
Scheduling truth lives in the heads of two people
Machine groups, setup sequences, nesting batches and remnant sheets are scheduled by experience. When those two people are out, the board in Kinetic stops being a reliable guide.
Scrap and remnant data is under-used
Nesting software knows utilisation, Kinetic knows cost, and the two rarely meet. Questions like which customer's parts generate the most scrap by material grade stay unanswered.
Where AI earns its place in Epicor Kinetic
Each use case names the ERP objects it reads or writes, so your ERP team can judge the integration effort before anyone commits budget.
Historical job cost lookup for quoting
An estimator describes the part family, material and quantity and asks what similar jobs actually cost versus quoted, with the jobs listed so the estimator can verify the match.
Touches: Erp.JobHead, Erp.JobMtl, Erp.JobOper, Erp.QuoteDtl, Erp.QuoteMtl, Erp.QuoteOper
Outcome: Faster quote turnaround on repeat-pattern parts and fewer under-priced jobs, since actual cost is checked rather than recalled; the estimator still sets the price.
Heat number and lot traceability
Quality asks which shipped orders contain material from a given heat or lot and gets a traced list from material issue to job to shipment.
Touches: Erp.PartLot, Erp.LotSelectedTracking, Erp.JobMtl, Erp.ShipDtl, Erp.OrderRel
Outcome: A traceability answer that takes minutes of review instead of hours of table walking, with the supporting records listed for audit.
Outside processing watchlist
A buyer asks which jobs are blocked on subcontract operations that are past their promised return date and gets the PO, vendor and affected customer orders.
Touches: Erp.POHeader, Erp.PODetail, Erp.PORel, Erp.JobOper (subcontract), Erp.OrderRel
Outcome: Earlier visibility on late subcontract work so expediting happens before the customer calls.
Machine group load and bottleneck check
A planner asks which work centers carry the heaviest queue over the next two weeks and which jobs are driving it.
Touches: Erp.JobOper, Erp.JobHead, Erp.ResourceGroup, Erp.Resource
Outcome: A quicker read on where overtime or outsourcing will be needed, as a decision aid rather than an automatic reschedule.
Scrap and rework by customer and material
A manager asks which customers and material grades generate the most scrap cost and gets a ranked, drillable result.
Touches: Erp.LaborDtl, Erp.JobMtl, Erp.NonConf, Erp.Part, Erp.Customer
Outcome: Evidence for price adjustments or process fixes, grounded in booked scrap rather than anecdote.
Late order explanation
Customer service asks why a specific order is late and gets the chain: material availability, operation status, subcontract return, or quality hold.
Touches: Erp.OrderHed, Erp.OrderRel, Erp.JobHead, Erp.JobOper, Erp.PODetail
Outcome: Customer replies that cite the real reason, drafted from live records and reviewed by a person.
Drawing and spec retrieval with ERP context
A retrieval layer indexes drawings, mill certs, weld procedure specs and customer quality requirements so a question returns the document and the part and job it relates to.
Touches: Erp.Part, Erp.PartRev, Erp.JobAsmbl, attached documents and file shares
Outcome: Less time hunting through folders for the right revision or certificate; the source document is always cited.
Reference architecture
A private model and a retrieval layer run beside Kinetic, never inside it. Questions are translated into constrained read queries over Kinetic through REST or a read replica, executed under the asking user's rights, and returned with the generated query visible.
- 1
Kinetic data access
Read-only access through Kinetic REST (OData v2 service endpoints and BAQs) or a SQL Server read replica. Curated BAQs for sensitive joins keep column access explicit.
- 2
Semantic layer
A mapped vocabulary of fabrication terms (heat, nest, remnant, outside op, work center) to Kinetic tables and user-defined fields, maintained with your planners so the model stops guessing.
- 3
Document retrieval
Indexes of drawings, mill certs, WPS documents and customer requirements, chunked and linked to part, revision and job numbers, stored on your infrastructure.
- 4
Private inference
An open-weight model served with vLLM or Ollama on your GPUs or sovereign cloud. No prompt or result goes to a public API.
- 5
Governance and audit
Per-user permissions mirrored from Kinetic security, query logging, and human approval before any write-back such as a BPM-routed update.
Integration notes for your ERP team
- Use a dedicated Kinetic API user or per-user token with a least-privilege security group; do not reuse a manager account.
- Prefer BAQs for cross-table reads such as job cost versus quote, so joins and column access are reviewed once and reused.
- Map user-defined columns (UD fields) on JobHead, Part and PartLot in the semantic layer; fabrication shops store heat, grade and finish there often.
- Index attached documents from the file share or Kinetic attachments with the part and revision as metadata.
- Expect SQL Server read replica load to be modest; schedule heavy analytical questions off peak.
- Write-back, if used, should go through Kinetic business objects and BPM method directives, with a person approving each change.
Deployment options
On-prem next to Kinetic
Shops running Kinetic on their own SQL Server with ITAR or customer NDA constraints
A single GPU server on the plant network reaches the database over an internal link; nothing leaves the building.
Private cloud tenant
Shops on Epicor-hosted or self-managed Azure who want AI in their own subscription
A model endpoint in your own tenant with private networking to the Kinetic REST layer; you control the keys and the logs.
Hybrid with a retrieval-only start
Shops that want document search first and ERP questions later
Start with drawings and certs on a file share, then add Kinetic query access once the semantic layer is agreed.
Compliance and data control
How the architecture supports your obligations. Certification and accountability stay with your organisation; the design keeps the evidence straightforward.
ITAR and EAR for defense parts
Architecture keeps technical data on infrastructure and personnel you designate; the customer determines whether their setup meets their ITAR obligations.
ISO 9001 and AS9100 records
The assistant is read-only and cites source records, so it supports retrieval of controlled documents without altering them.
AWS D1.1 and ISO 3834 weld documentation
Weld procedure and welder qualification documents are indexed with source citations; the model never certifies compliance itself.
Customer NDAs on drawings
Private inference means drawings are not sent to a third-party AI service, which many customer NDAs forbid.
How an engagement runs
Phase 1 . 1-2 weeks
Discovery
- -Top 20 questions from estimating, planning, quality and purchasing
- -Kinetic version, hosting and security review
- -Data classification for drawings and certs
Phase 2 . 2-3 weeks
Semantic layer and security
- -Fabrication vocabulary mapped to tables and UD fields
- -Read-only service account and BAQ set
- -Permission model matching Kinetic security groups
Phase 3 . 3-4 weeks
Pilot with one team
- -Working assistant for estimating or planning
- -Accuracy test set with expected answers
- -Usage and miss log reviewed weekly
Phase 4 . Ongoing
Expand and harden
- -Added teams and document sources
- -Monitoring, model update and rollback plan
- -Runbook for your IT staff
Questions to ask any vendor, including us
A short list that separates real Epicor Kinetic AI work from a chatbot demo.
- Does every answer show the exact query or records it used?
- Does the assistant run under the asking user's Kinetic rights, or a shared account?
- Where do drawings and mill certs get processed, and is any of it sent to a third-party model?
- How are user-defined fields and shop-specific terms handled?
- What happens if the model is wrong on a traceability question?
- Can we run it without internet access?
- What does it cost to run per year in GPU and support terms?
- Who maintains the semantic layer when we change Kinetic?
Frequently asked questions
Does this replace Epicor Kinetic reports and BAQs?
No. BAQs and dashboards remain the right tool for known, repeated questions. The language layer handles the open questions nobody built a report for, and it can call your existing BAQs rather than bypass them. Many shops find the assistant also shows which new BAQs are worth building.
Can it help with quoting?
It can retrieve how similar past jobs actually cost versus what was quoted, and list those jobs for the estimator to verify. It does not set prices or approve quotes. Pricing judgement, customer relationship and capacity context stay with the estimator.
How does it handle heat number traceability?
It traverses part lots, material issues, job material and shipments in Kinetic and lists the records it used. It is a fast lookup, not a certification. Any answer used for a customer or audit response should be checked against the underlying records.
Is our drawing data sent to a public AI service?
No. The model runs on your hardware or in your own private cloud tenant, and retrieval indexes are stored there too. This matters for customer NDAs and export-controlled drawings. We recommend verifying the network path during pilot.
Do we need Epicor Cloud or can this work on-prem Kinetic?
Both work. Self-hosted Kinetic offers direct read-replica options, while Epicor-hosted cloud usually means going through the REST API and BAQs. The REST route is slower for heavy analytics but keeps you within what Epicor supports for your hosting arrangement.
What hardware do we need?
A pilot for a small team often runs on one workstation-class or server GPU with a mid-sized open-weight model. Sizing depends on concurrent users and document volume, so we size it during discovery using your question mix rather than guessing up front.
Can it write back to Kinetic?
It can, but read-only is the default and we suggest staying there for months. If you add write-back, changes go through Kinetic business objects and BPM rules with a person approving each one, and every action is logged.
Related guides
AI for Epicor Kinetic, Beyond What Prism Covers
Add AI to Epicor Kinetic beyond Prism: private LLM over BAQs, BPM data, and REST v2, on-prem or private cloud, with honest guidance on when Prism already covers you.
Kinetic Cloud + private AIPrivate AI for Epicor Kinetic Cloud, Outside the Multi-Tenant Boundary
Ground a private LLM on Epicor Kinetic Cloud via BAQs, REST v2, and Data Fabric so sensitive manufacturing data never enters a multi-tenant AI feature.
Agents + approval gatesAI Agents for ERP, Running On-Prem
A practical guide to on-prem AI agents for ERP: what they can safely automate, where human approval belongs, and how to design the guardrails.
RAG + SQL + permissionsA Private LLM Grounded on Your ERP Data
How a private LLM answers questions on your ERP data: RAG plus text-to-SQL, role-based permissions inherited from the ERP, and where each fits.
Kinetic + injection moldingEpicor Kinetic AI for Injection Molders: Resin, Tooling, Scrap and Cycle Time Questions Answered
Private AI on Epicor Kinetic for injection molders: query resin lots, tooling, scrap and cycle-time variance in plain English, with the query shown.
Prophet 21 + private AIAI for Epicor Prophet 21, Built for Distribution Operations
Prophet 21 AI for distributors: on-prem or private-cloud LLM grounded in P21 order, inventory, and purchasing data, for order follow-up, forecasting, and pricing questions.
Plan it with numbers
On-Prem LLM Total Cost of Ownership Calculator
Model the full multi-year cost of running LLMs on your own hardware, including GPU capex, power, cooling, support contracts, and operations staffing.
Free ToolERP AI Maturity Assessment
Benchmark how deeply AI and automation are embedded in your ERP operations, from data foundations to autonomous agents, across four maturity levels.
Free ToolERP Integration Complexity Calculator
Estimate the hours, cost, and elapsed time of your ERP integration workstream from interface count, integration patterns, and middleware maturity.
GuideLLM Integration Patterns for Enterprise ERP
Integrate large language models with enterprise ERP systems using RAG, fine-tuning, function calling, and prompt engineering patterns for production deployment.
GuideOn-Prem LLM Deployment Architecture: Reference Guide
Reference architecture for on-prem LLM deployment: inference servers, GPU sizing, RAG pipelines, and security zones for regulated manufacturers.
GuideA Data Governance Framework for Private AI
A practical data governance framework for private AI: classification, access controls, retention, and audit patterns that satisfy CMMC 2.0 and ITAR.
Talk it through with an engineer who knows Epicor Kinetic
Bring one real question your team cannot answer from the ERP today. We will map the data path, the model, and where it runs, and tell you honestly if AI is the wrong tool for it.