Any ERPIndustry

EV battery manufacturing + private AI

AI for the ERP Behind an EV Battery Plant: Genealogy, Yield and Passport Data

Short answer

AI ERP for EV battery manufacturers puts a private language model over the ERP and the systems around it so quality, planning and compliance teams can ask questions about lot genealogy, cell grading, supplier material and scrap in plain English. It runs on the customer's own infrastructure, which matters because electrode recipes, yield curves and customer pack designs are among the most sensitive data a battery maker holds.

ERP: Any ERP
Industries: EV Batteries, Energy Storage, Automotive
Written for: VP Operations

Discuss your Any ERP environment See use cases

A battery plant is a process-manufacturing plant at the front and a discrete assembly plant at the back. Slurry mixing, coating, calendering and slitting run as continuous or batch processes with material yield and recipe control. Winding or stacking, electrolyte fill, formation and aging, and then grading turn that material into cells, which are assembled into modules and packs. Most ERPs were built for one of those worlds, and battery makers live in both.

The ERP usually owns the bill of materials, purchasing, inventory, costing, sales orders and finance. The MES, the formation and test equipment software and the quality system own the process data. The hard questions sit between them: which cathode lot went into which cells, which of those cells ended up in a pack that shipped to a given customer, and what did that lot do to formation yield. Answering that today often means an engineer and three exports.

Regulation is raising the stakes. The EU Battery Regulation (2023/1542) brings carbon footprint declarations, recycled content documentation and a digital battery passport for industrial and EV batteries. Customers apply IATF 16949 expectations, UN 38.3 transport testing and UL or IEC safety standards, and many US buyers ask for sourcing documentation tied to content rules. Each of these is a data assembly job across systems.

This page describes how private AI on top of the ERP can help battery makers with those jobs, what it should not be trusted to do, and how to deploy it so recipes, yields and customer designs never reach a public model API.

What usually gets in the way

The problems we hear most from vp operations teams running Any ERP.

Genealogy is split across systems

Material lots live in the ERP, process parameters in the MES, and formation and grading results in tester databases. A recall or customer complaint needs one chain from raw material lot to pack serial, and building it is a manual project each time.

Yield loss hides in early process steps

A coating weight drift or a slurry viscosity excursion shows up weeks later as capacity fade or self-discharge at grading. By then the ERP has already costed the scrap as a single variance with little cause attached.

Formation and aging capacity is a bottleneck nobody can see

Formation channels and aging rooms are fixed-capacity assets with long dwell times. Planning them from the ERP's generic work center capacity gives schedules that look feasible and are not.

Regulatory data is assembled by hand

Battery passport fields, recycled content shares, carbon footprint inputs and supplier due diligence records are pulled from purchasing, bills of material and supplier portals by a compliance analyst for each product family.

Supplier material variability is poorly tracked

Cathode and anode active material, separators and electrolyte vary lot to lot. Incoming inspection results rarely get linked to downstream cell performance in a way planners and supplier quality engineers can query.

Where AI earns its place in Any ERP

Each use case names the ERP objects it reads or writes, so your ERP team can judge the integration effort before anyone commits budget.

Cell-to-material genealogy queries

A quality engineer asks which cells contain a named cathode lot, and which packs and customers received those cells, and gets a traceable answer from joined ERP, MES and test data.

Touches: Lot and serial master, goods receipt, production order components, shipment lines, MES cell records

Outcome: Turns a multi-day containment investigation into a query that scopes the affected population in minutes, subject to how complete the source data is.

Formation and grading yield explainer

The assistant compares yield by cell lot, line, coating batch and supplier lot, and highlights the factor that separates good and bad groups for an engineer to confirm.

Touches: Production order confirmations, scrap postings, batch characteristics, grading results

Outcome: Moves early-process root cause work from spreadsheet joins to a reviewed shortlist of likely drivers.

Supplier incoming-quality summarizer

Reads certificates of analysis, incoming inspection records and NCRs per supplier lot and drafts a supplier scorecard narrative for the supplier quality team.

Touches: Purchase order receipts, inspection lots, vendor master, quality notifications

Outcome: Cuts the manual prep before supplier reviews and surfaces repeat issues sooner.

Battery passport data assembly

Collects declared fields from bills of material, material masters and supplier declarations into a draft record per product, flagging missing or inconsistent entries.

Touches: BOM, material master, vendor declarations, product family master

Outcome: Shifts compliance staff from collecting data to reviewing gaps, with the final declaration still owned by a person.

Pack build shortage triage

Planners ask why a pack order cannot start and get the shortage chain: cells held in aging, modules awaiting a BMS board, or an open supplier delay.

Touches: Planned and production orders, MRP exceptions, inventory by status, purchase order schedule lines

Outcome: Shortens the daily shortage meeting by arriving with causes instead of lists.

Engineering change impact on cell and pack BOMs

For a proposed change to a separator or tab material, lists open orders, stock, supplier lots and customer approvals affected, and drafts the change notice for review.

Touches: ECO records, multi-level BOM, open production orders, inventory, customer approval documents

Outcome: Gives change boards a fuller impact view before the decision, not after.

Dangerous goods shipping document checks

Checks outbound shipment records for lithium battery classification, state of charge declarations and test summary references before documents are released.

Touches: Delivery documents, material master transport data, test report attachments

Outcome: Catches missing or inconsistent shipping data early; the dangerous goods signatory still approves.

Reference architecture

The AI layer joins data the ERP already holds with read access to the MES, quality and test systems, then answers questions with the queries it ran visible to the user. Models are open-weight and run on the plant's own GPUs.

1
Connectors
Read-only ERP access through APIs or reporting views, plus MES and tester database connections, each using a scoped service account that respects existing roles.
2
Genealogy semantic layer
A mapping of lot, batch, cell, module and pack identifiers across systems so a question about a material lot resolves to the right downstream serials.
3
Model serving
Open-weight models served with vLLM or Ollama on customer hardware or a sovereign cloud, sized for engineering query volume rather than consumer scale.
4
Retrieval and agents
Retrieval over certificates, specifications, NCRs and work instructions, with agents that draft records and notices but never post transactions without approval.
5
Governance and audit
Every question, generated query and answer is logged with the user identity, and recipe or design data can be excluded from retrieval by role.

Integration notes for your ERP team

Lot and serial relationships are read from the ERP's batch or lot tables, then joined to MES cell identifiers using a maintained mapping rather than free-text matching.
Tester and formation databases are accessed through read replicas or exports so production test systems are not loaded by analytical questions.
Certificates of analysis and inspection PDFs are indexed with their lot number as metadata, so retrieval filters on lot before the model reads anything.
Supplier declarations for passport fields are stored as structured records where possible, with the source document linked for audit.
ECO and BOM reads use the ERP's effectivity dates, so impact queries reflect what is active, not just the latest revision.
Write-back is limited to drafts such as change notices and supplier corrective action requests that a named approver releases.

Deployment options

Air-gapped on-prem

Cell makers whose electrode recipes and process windows are core intellectual property.

Model, retrieval index and connectors run inside the plant network with no outbound path, and updates are brought in through a controlled process.

Private or sovereign cloud

Pack integrators and newer plants without a large data center footprint.

A dedicated tenant under the customer's own cloud agreement, with regional residency matching the plant's customers and regulators.

Hybrid by site

Groups with a pilot line in one country and volume plants in others.

Per-site retrieval and models keep process data local, while a governance layer standardizes questions and reporting across sites.

Compliance and data control

How the architecture supports your obligations. Certification and accountability stay with your organisation; the design keeps the evidence straightforward.

EU Battery Regulation 2023/1542

The system assembles declared data and supplier evidence into drafts and flags gaps; the manufacturer remains responsible for the declaration and passport content.

IATF 16949 and customer-specific requirements

Traceability and containment queries support the documented process; control plans and PFMEA remain controlled documents in the quality system.

UN 38.3, UL 1973, UL 2580, IEC 62660

Test report references are linked to product records so shipment and certification checks can be queried; testing itself stays with accredited labs.

Export control and customer NDAs

On-prem deployment lets the customer keep controlled technical data and customer pack designs off third-party AI services.

Sourcing and content rules (US and EU)

Bill of materials and supplier origin records are queried to support the customer's own due diligence files.

Where Netray fits

ERPray

Planners, quality engineers and compliance staff can ask grounded questions over ERP data with the query shown, role-aware and read-only by default.

DataRay

Genealogy and yield questions need tester databases, file shares and certificates alongside the ERP, which is the mixed-source case DataRay is built for.

Custom build

Passport assembly agents and plant-specific genealogy mappings usually need a bespoke build around the plant's own identifiers.

How an engagement runs

Phase 1 . 2-4 weeks

Discovery

-Map of ERP, MES and tester identifiers across lot, cell, module and pack
-Ranked use cases by engineering and compliance pain
-Data sensitivity classification for recipes and designs
-Deployment recommendation and GPU sizing

Phase 2 . 6-10 weeks

Pilot

-Genealogy query working across two or three systems
-Yield explainer on one product line
-Accuracy review against known investigations
-Role and audit controls validated

Phase 3 . 4-8 weeks

Production

-Production deployment on customer infrastructure
-Passport draft assembly for one product family
-Runbook for adding connectors and use cases
-Training for quality and planning teams

Phase 4 . Ongoing

Scale

-Extension to further lines and plants
-Supplier scorecard automation
-Model refresh as open-weight options improve
-Quarterly accuracy and usage review

Questions to ask any vendor, including us

A short list that separates real Any ERP AI work from a chatbot demo.

Which systems does the assistant read, and how does it join lot, cell and pack identifiers across them?
Can electrode recipes and process windows be excluded from retrieval by role?
Where does the model run, and does any prompt or document leave our network?
How are wrong answers detected, and can we see the exact query behind each answer?
What does the system draft versus post, and who approves each write-back?
How will it handle our passport and customer data fields as the regulation's delegated acts are finalized?
What GPU capacity do we need for our user count, and what does it cost to operate?

Frequently asked questions

Can AI replace the MES for battery traceability?

No. The MES records process data at the machine and step level, and that record is the evidence. AI sits on top, joining ERP and MES data so people can query it. If the underlying genealogy is incomplete, the AI will say so rather than fill gaps, and fixing capture at the source is a separate project.

Does it help with the EU battery passport?

It helps assemble and check data. It can pull declared fields from bills of material, material masters and supplier documents into a draft and flag what is missing. The manufacturer still owns and signs the declaration, and the required fields depend on battery category and the final delegated acts.

Is our electrode recipe safe?

Recipe data is only exposed if you connect it. With on-prem deployment the model and index sit inside your network, and role-based filters can keep recipe tables out of retrieval entirely. You should still classify that data and test the exclusions before go-live.

Which ERPs does this work with?

It is ERP-agnostic. Battery makers run SAP, Oracle, Infor, Microsoft Dynamics and others, and the connectors read lot, batch, BOM and order data from whichever system is in place. Check the specific ERP pages on this site for connector details.

Will it predict cell yield?

It explains yield differences between groups using data you already have and highlights candidate drivers. A true predictive model, trained on process data, is a different build and needs enough history. We would scope that separately after the explainer proves the data is usable.

How long does a pilot take?

A genealogy and yield pilot on one line typically takes 6 to 10 weeks, depending on how clean the identifier mapping is across ERP, MES and tester data. The mapping work is usually the longest part, not the AI itself.

Can we use the ERP vendor's AI instead?

For ERP-only questions, vendor AI may be enough, and you should evaluate it. It generally cannot join tester databases, certificates and MES data on your premises, which is where battery questions get hard. Many plants use both for different jobs.

Related guides

EMS / PCBA + on-prem AI

AI for EMS and PCBA Manufacturers Running SyteLine, Epicor, or NetSuite

AI for EMS and PCBA manufacturers on SyteLine, Epicor, or NetSuite: automate BOM scrubbing, AVL checks, and obsolescence alerts, grounded in your own ERP data on-prem.

Faster paperwork, same standard

AI for NCR, CAPA, and 8D Drafting in AS9100 and ISO 9001 Quality Systems

AI drafts NCR, CAPA, and 8D reports from ERP quality data for engineer review, cutting drafting time under AS9100 and ISO 9001 without changing sign-off.

Agents + approval gates

AI 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.

Know what an ECO actually touches

AI for Engineering Change Impact on BOMs, Routings, and Open Orders

AI reads a proposed ECO against live BOM, routing, and open order data to show engineering exactly what a change touches before it gets released.

RAG + SQL + permissions

A 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.

Digital thread + AI

AI Across PLM and ERP: Teamcenter, Windchill, and Your ERP

AI that reads Teamcenter or Windchill alongside your ERP to trace engineering change impact, BOM sync gaps, and where-used questions, on-prem.

Plan it with numbers

Free Tool

Talk it through with an engineer who knows Any ERP

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.

Start the conversation See ERPray

All AI for ERP guides