B2B orders arrive as emails, spreadsheets, PDFs and photographed confirmations, from customers who each have their own product codes, assortment and lead times. The job is to turn the request into a correct order without bypassing the rules that keep orders correct. That makes it a workflow, not an industry: manufacturers, wholesalers and distributors share it.
Three steps are easily blurred: extraction (what did the customer ask for), validation (does it match their entitlements and your rules) and release (who commits it to the ERP). AI can help with the first, your systems do the second, and a person owns the third until measured results justify more.
The steps, what each needs, and what stays with people
Typewise connects to your systems through lookups that read data and actions that change it, with approval by a person on by default. Whether a workflow can run end to end depends on what your systems expose, so each step below says what it needs.
Extraction
Read the request and the attachments the AI can open, translate the customer's own codes into your item numbers using a customer-specific mapping you provide, and ask the customer for missing details instead of starting an email round trip.
What it needs. The customer-to-item mapping in a source the AI can read, with an owner; extraction accuracy measured on your real documents and formats before any output is trusted. Which file types and layouts read reliably is confirmed on your own documents.
What stays with people. The mapping is the authority; until it has a home and an owner, extraction is not trusted.
Validation
Check the extracted order against deterministic rules: assortment entitlement, minimum quantities, lead-time-based delivery dates, stock and allocation.
What it needs. The rules in a system that can be queried; where a validated spreadsheet already enforces them, it is kept as the rule engine.
What stays with people. Rules stay rules. A spreadsheet or ERP check that already works is kept, not replaced by a language model; stock in the ERP is not permission to promise a date.
Release
Prepare the order in the format your ERP already imports and present it to the responsible person with the checks and exceptions visible.
What it needs. An existing import path (file or API) into the ERP that a person can release, confirmed for your ERP and version; direct database writes are avoided.
What stays with people. A person releases until the measured error rate justifies more.
Quote completion
For configurable products, guide the customer through the attributes a quote needs so the person pricing it does not re-key or chase.
What it needs. The attribute set and its dependencies in a structured form; a destination the pricing person works in.
What stays with people. Pricing and the commercial decision stay with people; the design guides choices, it does not expose the whole attribute table.
Document copies and status
Answer invoice, delivery-note and shipment-status requests from verified trade customers, using documents an integration returns and the tracking numbers linked to the order.
What it needs. Documents and tracking numbers linked to the order record; a verification of the requester against the customer record, not the sender address alone.
What stays with people. Sender identity is verified against the customer record, not inherited from an email address alone.
What your order team keeps
- Verify identity in your systems, not by what a customer says in the conversation: account data and account changes need a verified customer.
- Configure each action separately, with approval by a person on by default, and keep the limit enforced in the connected system as well as in the instructions.
- Handoffs carry the conversation history and a summary. Decide who approves, and who owns the next step outside service hours.
- Keep deterministic business rules in the system of record, and let the AI read their results rather than change them.
- Sensitive business data has its own confidentiality obligations beyond personal data; what may be processed where is agreed before the pilot.
Hosting, access control and certifications are described in the trust center, and the way agents work with approvals and audit trails on how it works.
How to roll it out
- Fix the document that feeds the ERP. Where customer-specific order forms are maintained by hand and go stale, correcting that document and its update cadence removes errors before any integration is attempted.
- Keep the spreadsheet where it is better. For hundred-line orders, a validated template may remain the best interface. AI then handles extraction and validation around it rather than replacing it with chat.
- Test freshness against the real question. An order may be followed by a question ten minutes later; a daily export may serve invoice questions and fail that one. Set the refresh requirement per request type.
- Validate assumptions with the people placing orders. Time per order type and seasonal peaks come from the operators, not from a guess; the first business case should be checked with them.
Common questions
Does the AI place orders in our ERP?
It can prepare the order in the format your ERP already imports, with validation results visible, for a person to release, as an action you configure with approval on by default. The import path and the extraction accuracy are confirmed for your ERP and your documents first. Checks such as quantities, assortment and lead times stay in your systems.
Our customers order by spreadsheet. Do we need to change that?
Not if it works. A validated template may be the best interface for long orders. AI can work around it, with extraction and checks, and the template stays.
Where does the customer-to-SKU mapping come from?
Sometimes from a local file the service team maintains rather than from the ERP. Wherever it lives, it needs a home the AI can read and a named owner. Check it first.