Back to the supplier from the overview. Once it was clear that the entry was worthwhile, the next question came up, in which department to start. The answer was rarely where it looked most impressive. It was where a lot of text was prepared by hand every day.
The benefit arises area by area and at different speeds. Putting the whole company through an AI transformation is a major project. I recommend looking specifically at individual processes or departments. That way we arrive at a concrete question like „What does AI bring in quality assurance?" and can clearly scope a pilot whose success can be defined by key figures (KPIs).
We look at the areas in a manufacturing company. For each area it states what becomes possible, what you need for it and where the limit runs.
Possible levers for AI support lie where a lot of text is processed or where recurring patterns, processes and workflows appear. It is worth looking more closely at manual tasks as well as media and platform breaks. Securing knowledge, experience and know-how in digital form is a further area of application. With demographic change this gains importance, because when skilled workers leave or drop out, valuable information otherwise gets lost that should be secured and made accessible.
Engineering and technical office
What becomes possible. Checking a specification in a single pass against your standard specification and having the deviations named. Rediscovering earlier design decisions without asking the colleague who made them. Breaking standards texts down to the question of whether a particular design is permissible.
What you need. A model with a large context window, that is, with the ability to survey a lot of text at once. A complete specification will otherwise not be read in a single pass.
Where the limit lies. The system provides hints, not approvals. The check stays with you, and that should also be laid down in writing.
Work preparation
What becomes possible. Transferring customer requirements into a template for the production order. Finding similar earlier orders and using their effort as a reference point. Uncovering contradictions between drawing, parts list and customer text before they become apparent in production.
What you need. Access to your own order history via a document search. What matters is opening up your data, not the largest possible model.
Where the limit lies. Suggestions on sequence and time required are reference points, not planning. A language model does not calculate reliably.
Production and production management
What becomes possible. Condensing shift handovers from notes into a structured handover. Searching the fault reports of recent months for patterns. Generating work instructions from existing documentation in understandable language, multilingual if needed.
What you need. A small, fast model is enough. The benefit arises through the connection to existing documentation, not through model size.
Where the limit lies. Intervening in the control system is a completely different project with its own safety requirements and its own legal assessment. That should be considered separately.
Quality management
For automotive suppliers this is the area with the clearest benefit, because the documentation burden exists anyway.
What becomes possible. Making test protocols searchable across years. Grouping complaint texts by recurring fault patterns. Audit preparation, by having the matching evidence from your holdings compiled for a requirement. Also evaluating fault images if needed, if you use a model that processes text and image together.
What you need. A system that answers exclusively from the presented documents and supplies the source reference. That is not optional here, because a test record must be verifiable.
Where the limit lies. Store every generated evaluation together with input, model version and date. The same query does not necessarily produce the same answer, because technical effects in the background cause deviations. Anyone relying on being able to reproduce a result identically later will be disappointed.
Maintenance
What becomes possible. Retrieving a machine's maintenance history in seconds. Searching manufacturer documentation for a specific fault message, even across several manuals. Merging spare-part designations from different spellings.
What you need. A document search across your manuals and maintenance reports. This already works with small models and is often the fastest visible success in the whole company.
Where the limit lies. Manufacturer documentation can be protected by copyright. Internal use is mostly unproblematic, passing it on to third parties is not.
Logistics and shipping
What becomes possible. Breaking shipping regulations for a destination country down to the specific consignment. Checking delivery papers for completeness. Extracting customer-specific packaging requirements from framework contracts.
What you need. Above all, currency. Regulations change, so the document base has to be kept up to date. Having a model answer from memory is particularly risky here.
Where the limit lies. Customs and export questions need a human approval. A mistake in this area is expensive and comes to light late.
Sales
What becomes possible. Drafting a quotation from existing building blocks and comparable cases. Preparing a technical customer enquiry before it goes into engineering. Condensing conversation notes into a usable summary. Looking up in your own quotation history what you last offered a similar customer.
Why local. Your data is almost entirely personal. Contacts, notes, contact history. Local operation shifts the data-protection question from the contractual level to the technical level, because no processor and no third-country transfer arises.
Where the limit lies. In sales, an invented statement becomes a commitment. Prices, delivery times and technical properties may only come from verified sources. An automated assessment of the creditworthiness of customers is a different category and falls under the strict requirements of the EU AI Act.
Service and customer support
What becomes possible. Classifying fault reports and matching them against earlier cases. Generating service reports from the technician's keywords. Speeding up spare-part clarification by making the machine history searchable.
What you need. A connection to your service history. Without it the benefit is small.
Where the limit lies. If a customer communicates directly with the system, that must be labelled.
Marketing
What becomes possible. Preparing technical content for trade publications. Preparing translations. Evaluating customer feedback. Generating draft texts that are then edited.
What you need. Here, cloud use is usually acceptable, as long as no internal figures and no customer data flow in.
Where the limit lies. Marketing is the area with the only labelling obligation already in force. The first step is therefore an inventory of website chat, image material, subtitles and voice outputs. What you do not know, you cannot label. The classification for this is set out under Legal notes.
Purchasing
What becomes possible. Setting supplier offers against each other. Searching framework contracts for a specific clause. Comparing technical specifications. Preparing price negotiations with your own history.
What additionally lands with you. Two checks that no one else takes on. The licence of a deployed model, because it applies per variant and decides whether you may later pass a system on to customers. And supply-chain security, because companies below the statutory thresholds too are increasingly obliged by larger clients to comply with security standards.
Where the limit lies. An automated assessment of the creditworthiness of business partners falls into the highest risk category. Supplier assessment by technical criteria does not.
Finance and controlling
What becomes possible. Having deviations between costing and post-costing explained. Generating report texts from figures. Searching contract documents for payment and price-escalation clauses. With volatile raw-material prices, the last point is particularly relevant in this region.
Why local. Costing data is a trade secret. The cloud question is already settled in most companies here anyway.
Where the limit lies. A language model does not calculate reliably. Use it for explanation and preparation, not for calculation.
Human resources
What becomes possible. Drafting job advertisements. Generating onboarding documents from existing documentation. Making works agreements and collective rules searchable. Preparing training materials.
Where it tips over. As soon as the system pre-sorts applications, evaluates performance or allocates tasks, the strict requirements for high-risk systems apply.
Practical recommendation. HR is not an entry project. The technical effort is small, the regulatory one high. Anyone who starts here without preparation builds a system they have to switch off again.
IT
What becomes possible. Evaluating log files. Generating documentation. Pre-structuring support requests. Drafting scripts.
What you carry. Network segmentation, logging, model versioning and the checking of outputs before they move into downstream systems.
The point that comes to light too late in projects. A language model has no rights management of its own. Whoever has access to the system sees everything the connected data source delivers. The rights must therefore be mapped in the retrieval layer, not in the model.
Reassurance on the platform. You do not need Linux to start. Windows, Linux and macOS all work. Details under Models and technology.
Management
What you get from it. Calculable costs instead of ongoing licence fees. Independence from vendor decisions. An argument towards customers who ask about data security. And knowledge that stays in the company, even when employees leave.
What you cannot delegate. Under the amended BSI Act, responsibility for cybersecurity lies expressly with management, mere delegation is not enough. Added to this is the obligation to ensure sufficient AI competence in the company.
Your actual decision. The question is not the model. It is which area you start with and which expansion stage.
What applies the same in every area
Four points apply regardless of the department, and they can hardly be retrofitted later.
- The system answers from your documents, not from memory, and outputs the source reference with it. That is the most effective measure against invented statements.
- The access rights lie in the data layer, not in the model.
- Model name, version and date are documented for every output. Without this information it cannot be proven later which system produced a result.
- The users are trained. That is a duty, not a comfort.
Who operates the system after introduction?
This question decides success in the long run. It matters more than the choice of model. Clarifying responsibilities belongs in the project plan.
Clarify before the start who maintains the document base, who answers users' questions, who decides on a model change and who checks whether the results are still correct. In a company with a single IT position this is a real capacity question and often the actual reason for a no.
If these roles are not filled, after twelve months you have a system that runs technically and that no one uses any more. That is the most common silent project abandonment, and it appears in no statistic.
Where to start?
Recommended is the area with a high volume of documents and a low risk class. In practice that is usually maintenance with the manual search, quality management with the test protocols or sales with the quotation. All three deliver quickly measurable benefit without a laborious classification standing in the way.
Not recommended as a first project are human resources, credit assessment and anything that intervenes directly in a machine control. These cases are solvable, but they are not learning projects.
Which area yields the most for you can usually be clarified in a conversation within an hour.