If your office answers the same customer questions by hand every day (specs, the certificate for a batch, the ship date, stock) and vendors keep pitching you chatbots and "AI agents", start by sorting your last 50 inquiries by where the answer lives: a document, the ERP or an engineer's head. That tells you whether you need an assistant that cites the datasheet, a live read from the ERP, a customer portal or just better routing to the right person. Then test one kind of question for two weeks on your own data, with drafts a person approves, and ask every vendor to run that test before you sign.
This guide is for the owner, sales director or customer service lead at a manufacturer. It gives you the map, the test and the questions that should stay with your people.
Three places where the answers live
What you need depends on where your people find the answer. At a manufacturer there are three places:
- A document. A datasheet, a declaration of conformity, a certificate, installation instructions, warranty terms. The answer exists; someone has to find it.
- The ERP. Order status, ship date, stock level, batch number. The answer changes daily and is only true in the system.
- A person. Will this material hold up in unusual operating conditions, can the lead time be shortened, what discount for a larger order. The answer needs a decision from an engineer, a planner or a salesperson.
Each place needs something different. A tool that searches documents well does not know the order status. An ERP integration cannot say whether a seal survives 180°C. Neither should grant discounts on its own.
The question-to-source map: sort your last 50 inquiries
Take your last 50 customer inquiries: emails, phone calls logged by customer service, messages from distributors. For each one, write down five things:
- Kind of question, e.g. "certificate for a batch", "ship date", "replacement part".
- Where the answer lives: document, ERP or a person (by name).
- What fits: a draft reply citing the document, a read from the ERP, self-service in a portal, or routing to the right person only.
- Who approves before the reply reaches the customer.
- What you do not automate, even if it would be technically possible.
Here is what the map could look like at a company making plastic components (an illustration; the numbers are made up):
- "Please send the declaration of conformity for batch 2311" → document (declaration archive linked to the batch in the ERP) → draft email with the declaration attached and the batch number → approved by customer service → do not automate sending while the batch has an open nonconformance.
- "When will order 4471 ship?" → ERP → read of the status and planned ship date → approved by customer service until the test shows the ERP data is current → never give a date the planner has not confirmed.
- "Will this profile handle outdoor use all year round?" → a person (the process engineer) → the system only collects the details and passes them to the engineer with the product datasheet → approved by the engineer → do not automate the answer.
- "What price for 5,000 units?" → a person (sales) → routing to the salesperson with the customer's order history → approved by sales → do not automate discounts.
After 50 rows you will see the proportions. If most questions are about documents, start with an assistant that searches your documents. If most are about order status, start with the ERP, and possibly not with AI at all. If most need an engineer, the biggest gain is good routing and a complete intake, so the engineer does not have to chase missing details.
Questions answered from documents: an assistant that cites its source
Datasheets, certificates and manuals are good ground for AI, on one condition: every answer points to the document and the place it came from. Your customer service person sees a draft reply with a link to the right datasheet page and checks it against the source instead of hunting for the document.
Before you start, check three things:
- Which version of a document is current. If three versions of a datasheet sit on the shared drive, the assistant will find all three. Settle where the current version lives first.
- The link to a batch or product. A batch certificate only makes sense if you know which batch it belongs to. If that link is in the ERP, the assistant needs an ERP read as well.
- What the assistant does when there is no source. It should decline and hand the case to a person, not guess. Count those refusals separately in the test.
Questions answered from the ERP: a read from the system, sometimes a portal instead of AI
Order status and ship dates are records in the ERP. A language model can help understand the email ("where is the order we placed last week?") and word the reply. The information itself has to come straight from the ERP, at the moment of the question.
Ask your ERP partner what the system exposes to other applications. SAP Business One, for example, has the Service Layer, a REST (OData) interface another application can use to read business data. Other systems have their own interfaces or allow exports. The answer shapes the scope: reading through an API is a different project from a daily export to a spreadsheet.
Sometimes a customer portal is the better answer: logged-in customers see their own orders and dates. If most questions are "where is my order" from regular buyers, a portal may take them off customer service better than a chat would. AI makes sense where questions arrive by email, in every possible form, and someone has to understand them first.
Questions you do not automate
Some questions stay with people, whatever the tool promises:
- Discounts and off-list prices. That is a commercial decision.
- Complaints. The customer wants to know someone owns the case. A system can collect photos, the batch number and the description, but a person decides.
- Dates the planner has not confirmed. Better a later reply than a date that does not come from the production plan.
- Safety questions and unusual applications. The answer stays with the engineer.
Here AI helps differently: it recognizes the kind of case, collects what is missing and passes it to the right person complete.
A draft for your team or a chat for your customers
The decision that matters most is who reads the answer first. A draft your employee checks, corrects and sends is a different system from a chat that answers the customer by itself.
If you sell in the EU, there is a legal difference too. Article 50 of the AI Act applies from 2 August 2026: AI systems that interact directly with people must tell them from the first interaction that they are dealing with AI, unless that is obvious. According to the European Commission's guidance, that duty does not cover a person acting as the intermediary with the customer, or systems working only in the background. Designing that notice is the system provider's job, so ask your vendor how they handle it; machine-readable marking of AI-generated content sits with the provider too. Plan a customer-facing chat together with the AI notice, and settle the details with your lawyer.
It is safer to start with a draft for the team: the customer still gets the reply from a person, and you see the quality of the drafts before anything goes out.
A two-week test on your own data: what to measure and when to stop
Pick one kind of question from your map: the most frequent one in your 50 rows that has its answer in a document or the ERP. For two weeks the system drafts replies to real inquiries while your team answers as usual and compares. Customers see nothing new.
At the end, count four things:
- Drafts sent unchanged. How many replies your team could have sent straight away.
- Drafts corrected. How many needed changes, and which: a wrong fact, a missing attachment, the tone.
- Questions handed to a person. How many cases the system rightly passed to an engineer or salesperson.
- Refusals for lack of a source. How often the system found no document or data. Each one points to a gap in your documentation worth closing with or without AI.
Stop if drafts with a wrong fact keep appearing, if your team corrects nearly every reply, or if the ERP data turns out to be out of date. Then fix the sources first and run the test again.
Before the test, agree who scores the drafts and by what rules. Without that, two weeks leave you with an impression instead of numbers.
How we do it
In an AI discovery sprint we take one of your processes: two weeks, your own data, measured benefits and a recommendation to build or stop. The first stage is fixed-price. At the end you get results you can check; here is how to accept them.
For one manufacturer we built a production management system connected to the company's SAP, with 15+ processes in one system.
Send us a few lines about your most common customer questions and the name of your ERP. A few sentences, and you'll know how we'd tackle it.