AI agents in the enterprise: where they actually make sense

Over the past few months, every other article about AI opens with the words “revolutionary AI agent”. The real picture is a little more sober - agents are a powerful tool for specific types of tasks, not a cure-all. Let’s look at where they genuinely work today.

What is an “AI agent”

Put simply: a language model (LLM) that:

1. receives a task,
2. has access to tools (search, a database, an API, e-mail, …),
3. independently decides which tools to use and in what order,
4. delivers a result.

The difference from “plain” ChatGPT is precisely that an agent actively does something, rather than just replying with text.

Where it works today

1. Processing incoming e-mails and tickets

• Classification by type (complaint / order / enquiry / spam).
• Data extraction (order number, contact, request).
• Answering common questions (price, availability, opening hours).
• Escalating complex cases to a person, including a summary.

Typical benefit: 40-70 % of tickets handled without human intervention, the rest resolved faster.

2. Intelligent search over company documents (RAG)

An employee asks in natural language - “what is the holiday approval process for a department head?” - and the agent searches guidelines, contracts and manuals and returns an answer with links to the source.

Typical benefit: a drastic reduction in search time, especially in legal and HR departments.

3. Pre-processing documents (invoices, contracts, delivery notes)

OCR + LLM + validation against the ERP system. The accountant receives the document pre-filled and only confirms or corrects it.

Typical benefit: 60-80 % time saved when processing a single invoice.

4. Sales and after-sales

• Summarising calls and meetings, writing them into the CRM.
• Generating follow-up e-mails.
• Analysing customer feedback and classifying topics.

Where it doesn’t work yet - or is expensive

• Fully autonomous, high-risk decisions. An agent that sends out a €100,000 invoice on its own is an idea that can wait a while longer.
• Complex multi-step processes without human oversight. The more steps, the greater the chance the agent “gets lost”.
• Tasks done once every six months. The return on investing in an agent will not be recovered.

A practical ROI estimate

Try answering three questions:

1. How many hours a month does the task take?
2. What is the error rate of the current process?
3. How often does the task repeat (daily / weekly / seasonally)?

An agent usually makes sense when:

• the task takes more than 20 hours a month,
• it is repeatable,
• it has clear success criteria (we can measure whether it was done well).

Security and data

An AI agent that works with company data should never send sensitive content to public APIs without control. Solutions:

• Private instances (Azure OpenAI, AWS Bedrock).
• On-premise models (Llama, Mistral) for particularly sensitive cases.
• Masking personal data before sending it to the model.

---

At RBH Technology we build AI projects incrementally - we start with a pilot delivering measurable value in 4-8 weeks and, based on the result, decide where to go next. Write to us if you have a concrete problem to consider.