LLMs & AI Agents · Industries
Applications of LLM Agents Across Industries
An LLM agent plans steps, pulls data from your systems and carries a task through to a result. Our projects show how agents are built and which tasks they take on today in automotive, manufacturing, healthcare, public affairs and government.
Aleksandra Osztynowicz
Key takeaways
- An LLM agent uses a large language model as its core and adds tools, memory, planning and rules.
- Companies do not train their own language model for this. The work lies in tools, data, integration and testing.
- Examples from our projects: planning data from SAP BW in seconds, 90% less manual intervention in order entry, 70% less routine work for a public affairs team.
- Not every process needs an agent. Sometimes rules or a single model step are enough.
What an LLM agent is
An LLM agent uses a large language model as its central control. The model understands the request, breaks it into steps, decides which tools it needs and checks the result. What sets it apart from a chatbot is that it acts: an agent reads from systems, writes drafts, opens tickets or triggers the next step in a process.
How an LLM agent is built
| Building block | Role |
|---|---|
| Core | The language model understands the request, reasons and chooses the next step. |
| Planning | Large tasks are broken into smaller steps, and the plan is adjusted when needed. |
| Tools | Defined access to systems, such as reading the ERP, querying a database or drafting an email. |
| Memory | History, context and earlier results persist across several steps. |
| Knowledge | Answers from company documents, usually connected through RAG. |
| Rules and prompts | Set the role, goal and limits, for example which actions need approval. |
In operation, logging and handover to people are added: anything the agent cannot resolve with confidence goes to the person responsible.
Applications by industry
Automotive: answers from several systems
In practice: premium car manufacturer (under NDA)
Product planners spent hours querying SAP BW and several data warehouses. An assistant now takes the question in German or English, typed or spoken, decides which systems to query and combines the results. Hours become seconds, and the data stays in the company’s own infrastructure.
Manufacturing: orders from email
In practice: Radaway
Every incoming email is classified, orders are read including attachments, products are matched against the database and the result is validated before it reaches order entry. The result: 90% less manual intervention and more than 95% accuracy in product matching.
Healthcare and events: answering questions live
In practice: German Congress of Surgery
At the 141st Congress of the German Society of Surgery, a voice assistant handled navigation and programme questions for four days. It called up programme data for hundreds of parallel sessions through function calls as it answered, on a screen kiosk and in VR.
Public affairs and consulting: analysing documents
In practice: RPP Group
For the public affairs team, five specialised agents review new policy documents, summarise the key points, structure positions and prepare content to internal standards. Routine work fell by 70%, and the experts have the final say.
Government: routing inquiries correctly
In practice: technical inspection authority
Citizens ask in everyday language, while the answers sit in technical documentation. A main agent reads the question and distributes it across 11 specialised knowledge areas, without anyone having to pick a category first.
Example: an agent in customer service
A simplified example from online retail shows how an agent works with systems. The scenario is invented and only serves as an explanation.
- Request: a customer asks in the chat where her order is.
- Plan: the agent recognises that it needs the order number, order status and tracking.
- Tools: through an interface it queries the order system, read-only and only for this customer.
- Answer: it gives the status and the tracking link.
- Handover: if the parcel is damaged or lost, it opens a ticket and hands over to the service team.
Technically, access runs through defined interfaces, such as an API gateway in front of the order system and database. That keeps it clear what the agent may and may not do.
How companies introduce LLM agents
- Define the task: a clearly scoped workflow with a measurable result, such as order intake or planning queries.
- Define tools and permissions: which systems the agent may read or write, and which actions need approval.
- Connect knowledge: prepare documents and data so the agent works from current sources.
- Prototype with real cases: test on real requests, including exceptions and missing information.
- Evaluate: measure accuracy, handover rate and run time before the agent goes live.
- Monitor operation: review logs, analyse errors and keep improving the agent.
No custom language model is trained for this. Why the step from pilot to operation often fails is covered in our article The pilot worked, production didn’t.
When an agent fits
Not every process needs an agent. Fixed workflows with clear rules often run better on rules and scripts. Many tasks only need a single model step, such as a summary. An agent pays off when a task connects several steps and systems and runs differently from case to case. When it grows larger still, several agents work together, as described in our article on multi-agent systems.
Which task should an agent take on?
We look at whether rules, a model step or an agent fits your process, and how it connects to your systems.
Request a process analysisFrequently asked questions
An AI system that uses a large language model as its core and adds planning, tools, memory and rules. This lets it work through tasks in several steps and act inside existing systems.
A chatbot answers questions. An agent also plans steps, retrieves data or triggers actions through defined tools, and hands over to people what it cannot resolve with confidence.
For example automotive for planning queries, manufacturing for order intake, healthcare and congresses for visitor questions, public affairs for document analysis and government for citizen inquiries.
Usually not. Agents build on existing language models. The work lies in tools, data connection, integration, testing and operation.
Aleksandra has been building custom AI solutions at theBlue.ai since 2021, with a focus on agentic implementations and local LLM deployments that precisely fit enterprise needs. As an AI Engineer, she helps organizations automate their processes with production-grade systems, from on-premise open-source models to RAG-based internal knowledge bases for regulated industries, and continuously expands her knowledge in the rapidly changing world of artificial intelligence.
In her articles, she shares practical experience from real enterprise AI projects and shows that deploying AI in companies doesn’t need to be complicated.