AI agents for SMEs: what already works today and what does not
5 min read
Short answer
AI agents are systems that do not just answer but carry out several steps on their own: looking up data, operating programs, creating records or preparing emails. In small and mid-sized companies, they work well today for clearly defined tasks with fixed tools and human approval. For open, business-critical processes without oversight, they are still too unreliable.
What does this mean for your company? Book an intro call
AI agents are the big topic among software vendors. The promises go as far as digital employees replacing entire departments. Reality in small and mid-sized companies is more down to earth, but genuinely interesting: for clearly defined tasks, agents work well today.
What sets an agent apart from a chatbot?
A chatbot answers questions. An agent gets tasks done. It receives a goal and tools, such as access to customer management, the inbox or a calendar, and decides for itself which steps are needed. It looks things up, compares, creates records and reports the result.
Where do agents work well today?
- Processes with fixed tools: reading an enquiry, finding the customer in the system, checking open orders and preparing a draft reply. More in Automating a shared inbox.
- Research in your own data: “Which suppliers delivered this item to us over the last two years, and at what price?” The agent searches several sources and summarises.
- Preparation instead of decision: compiling documents for a meeting, flagging discrepancies in lists, creating tasks from minutes.
What these examples have in common: the framework is clear, the tools are limited, and a person looks at the result at the end.
Where are agents still risky?
- Long chains: every step has a small chance of error. Over many steps, it adds up.
- Write access to critical systems: bookings, payments or master data should not be changed without approval.
- External communication: emails to customers without review risk wrong commitments.
- Manipulation through content: emails or web pages can contain hidden instructions an agent might follow. Specialists call this prompt injection.
How do you deploy agents safely?
- As few rights as possible: read before write, only the systems that are really needed.
- Approval before impact: a person approves anything with external effect or financial consequences.
- Logging: every step the agent takes is recorded traceably.
- Set limits: cap the number of steps and the cost per task.
- Test with real cases: above all the difficult and unusual ones.
Agent or classic automation?
An honest question first: does the task need an agent at all? If the steps are always the same, classic automation is more reliable, cheaper and easier to check. Agents pay off where flexibility is needed because every case looks a little different. Which projects pay off in principle is covered in Which AI projects pay off?
Julius Launhardt deploys AI agents in small and mid-sized companies where they reliably help, with clear limits and approvals. More on the AI consulting page.
Frequently asked questions
What is the difference between an AI agent and automation?
Classic automation follows fixed steps defined in advance. An agent decides for itself, within a framework, which step to take next. That is more flexible but less predictable.
Can AI agents replace employees?
Today they take over individual, clearly defined tasks. In SMEs they generally do not replace entire roles. Used sensibly, they relieve skilled staff of routine work.
What is a prompt injection?
An attack in which hidden instructions to the AI are placed in third-party content, such as an email or a web page. If an agent reads that content, it could follow the instructions. Limited rights and approvals protect against this.
How do I get started with AI agents?
With a task where the agent may only read and prepare drafts. Once that works reliably for weeks, you can gradually trust it with more.