Solution · Operations & supply chain
AI agents for operations and supply chain: less retyping, fewer surprises.
Orders arrive as emails and PDFs, delivery notes as scans, supplier confirmations as free text. Someone retypes all of it into the ERP and chases what is missing. AI agents read those documents, create validated transactions, follow up with suppliers and put the exceptions in front of the right person, with audit logs and four-eyes checks wherever money moves.
Where agents pay off in operations
Operations runs on unstructured input: emails, PDFs, phone calls and spreadsheets that all have to become clean transactions in the ERP.
The pattern is extract, validate, act. An agent reads an incoming document, matches it against master data (customers, articles, prices, suppliers, open purchase orders), creates a proposal in the ERP and routes whatever does not fit to a person. Unlike OCR templates or RPA scripts, agents cope with variation in layout, language and completeness, and they can ask the sender for what is missing.
Why now: document-understanding models handle real-world scans and messy emails at a usable level, ERP vendors from SAP and Microsoft Dynamics to Mittelstand systems expose interfaces, and EDI covers only the large trading partners. The long tail of small customers and suppliers still sends email, and the people who used to retype it are hard to hire.
We start with one high-volume, measurable document flow, order intake or delivery notes, design ERP writes as proposals before they become postings, and let people handle the exceptions. When agents beat scripts is explained in our article agents vs RPA vs chatbots.
- Typical entry pointOrder intake from email and PDF; delivery-note processing
- SystemsERP (SAP, Microsoft Dynamics, others), WMS, supplier portals, email, EDI, DMS
- Human checkpointExceptions, price and quantity deviations, anything with financial impact
- First pilotTypically 6–8 weeks
Use cases
Six workflows from the inbox to the shop floor. Each names the systems involved and where a person decides.
Order intake into the ERP
The agent reads customer orders arriving as email text, PDF, spreadsheet or EDI, matches customer, article numbers, prices and delivery dates against master data and contracts, and creates the order proposal in the ERP. Complete, consistent orders are released under rules you define; every deviation goes to a clerk with the original document and a specific flag.
Intelligent document processing
Delivery notes, customs documents, certificates of origin and analysis, packing lists: the agent extracts the data, validates it against the purchase order or shipment, files the document in the DMS with the right metadata and flags missing or expired certificates. Disputes and unclear cases stay with people.
Supplier communication and confirmation follow-ups
After a purchase order goes out, the agent tracks the confirmation, reads it when it arrives, compares dates, quantities and prices with the order and chases missing confirmations with emails from approved templates. Deviations are escalated to the buyer with a proposed response. The agent renegotiates nothing.
Exception handling
Late deliveries, quantity mismatches, damaged goods: the agent gathers the facts (order, confirmation, tracking, goods receipt, photos), proposes options such as partial delivery, substitution or customer notification, and drafts the communication. The planner or customer service decides, with a complete case file in front of them.
Demand and inventory signals for planners
Each morning the agent summarises open orders, forecast deviations, stock-outs, supplier lead-time changes and blocked shipments into a short briefing and answers planners' questions in plain language over ERP and WMS data. It places no orders itself; it makes the planner's first hour useful.
Shop-floor knowledge, SOPs and reporting
A knowledge assistant answers questions from machine manuals, maintenance history and SOPs with sources. The agent drafts SOP updates and quality documentation from change records and prepares the weekly operations report. People approve every document that becomes binding.
Worked example
An order-intake agent: from unstructured emails to a validated ERP order
A typical scenario: a manufacturer or wholesaler receives several hundred orders a week at orders@, as email text, PDF attachments and the occasional spreadsheet, and three people retype them. This is how an agent takes over the routine while the team keeps the decisions.
- Detection and extraction. An email arrives. The agent decides whether it is an order, an enquiry or a complaint, then extracts the header (customer, order number, requested date, delivery address) and the line items from body and attachments.
- Customer identification. It matches sender, company name and customer number against ERP master data. Ambiguous matches, such as a new contact at a known customer or a subsidiary, go to a person.
- Line matching. Customer article numbers are mapped to your own through mapping tables and past orders. Unknown articles produce a suggestion with a confidence score and a decision for the clerk, never a silent guess.
- Validation. Prices are checked against price lists and contracts, quantities against packaging units and minimums, requested dates against availability and lead times, and the credit limit is verified. Each check returns pass or flag.
- Clean orders. If every check passes, the order is created in the ERP. During the pilot it is a draft the clerk releases with one click; later, orders from defined customers are released automatically.
- Exceptions (human gate). Anything flagged becomes a task in the clerk's worklist with the original email, the extracted data, the specific issue and a proposed action, for example “price deviates by four percent, contract price applies, confirm?”
- Customer communication. The order confirmation is generated from the ERP order, not from the email, and sent after release. Questions to the customer are drafted for the clerk to approve.
- Audit trail. Every step is logged: source document, extraction, checks, who released what and when. Quality management and finance can trace any order back to its origin.
- Weekly review. Touchless rate, exception reasons and errors found downstream are reviewed with the team, and mapping tables and rules are improved.
Guardrails and risks in operations
ERP write access by design
The agent uses a dedicated technical user with the minimum rights for the transaction in question. It creates proposals or drafts before it is allowed to post directly, never changes master data, and every permission is reviewed with IT and the ERP owner.
Four-eyes checks on financial impact
Price deviations, credit notes, order values above a threshold, supplier changes and anything affecting an invoice require a second person. The thresholds are configurable and logged, so the control is visible to auditors.
Audit logs and traceability
Source documents, extracted data, checks and decisions are stored with timestamps and user identities. This supports quality management, ISO audits and customs reviews and is a precondition for extending autonomy later.
Master data quality
Agents expose bad master data quickly: duplicate customers, outdated prices, missing packaging units. We plan the clean-up alongside the pilot rather than letting the agent paper over it. See our data foundations service.
Resilience and fallback
When a model or interface is down, orders must still flow. Every workflow has a manual fallback, monitoring with alerts and a defined owner, and supplier and customer communication uses approved templates so tone and commitments stay under control.
How we start
Operations assessment
We map your document flows, volumes, formats and systems, measure time per order, error rates and backlog, and select the first flow with a clear baseline. IT security and works council questions are raised at this stage.
Design and integration plan
Extraction schema, validation rules, exception categories, the ERP integration approach and roles are designed with your order desk, IT and, where relevant, quality management.
Pilot
Built on your real documents. Shadow mode first, then draft mode with clerks releasing orders. We measure touchless rate, accuracy and time per order against the baseline.
Scale and operate
More customers, document types, plants and flows such as supplier follow-ups and exception handling, with managed AI operations keeping rules, mappings and interfaces current.
Frequently asked questions
We already have EDI and OCR. Why would we need an agent?
EDI covers your large partners, OCR templates cover stable layouts. The agent is for everything else: the long tail of customers who email free text, the supplier who changes his PDF layout, the order with a missing article number. It complements EDI and OCR rather than replacing them, and it can ask the sender for what is missing, which neither can.
Will the agent write directly into SAP or Dynamics?
Eventually, in a limited way, if you decide so. We start with read access and proposals, then drafts that a clerk releases, and only then direct creation of orders for defined customers and value ranges. Every write uses a technical user with minimal rights and is logged. Master data is never changed by the agent.
How accurate is the extraction?
It depends on your documents, which is why the pilot measures it on your real data rather than on a vendor's demo set. Confidence thresholds route uncertain fields to a person, so accuracy is a question of how much lands in the exception queue, not of silent errors. We report field-level accuracy and touchless rate every week of the pilot.
What about data protection for supplier and customer documents?
Most of the content is business data, but contact names, signatures and phone numbers are personal data under the GDPR, and some documents are confidential. We define what the agent may read and store, keep processing in the EU where required and agree processor terms with model providers. Your data-protection officer reviews the design.
What happens to our order-processing team?
Agents remove tasks, not usually roles. Retyping disappears; exception handling, customer contact, master data quality and supplier coordination remain and become the core of the job. Many teams use the capacity to absorb growth or shorten lead times. We plan the change with the team lead and train the team on the new tools.
Does this work with an older or smaller ERP?
Usually yes. Where there is no API we work through interfaces, file exports, email or a controlled front-end automation, and we say clearly when a system needs groundwork first. The readiness assessment answers this question for your specific landscape.
Related pages
Agentic workflow automation
Multi-step business processes automated end to end by LLM-powered agents across ERP, CRM, ticketing and email, with human approval steps built in.
Manufacturing & engineering
Agents for quoting, order intake, service knowledge, spare parts and quality documentation, built for ERP and PLM landscapes, OT/IT separation and export-controlled drawings.
Logistics & supply chain
Order intake from emails and PDFs, customs documents, exception handling and freight audits: agents built for 24/7 operations and thin margins.
AI agents for procurement
Request intake, RFQ drafting, quote comparison, supplier onboarding checks and spend analysis, with buyers approving every award and every order.
Let's find the first workflow worth automating.
A 30-minute intro call, no slides and no obligation. We listen, ask about your processes, and tell you honestly where AI agents would pay off and where they would not.