AI Agents Explained: What They Can Do, What They Cost and How to Build One

What Are AI Agents?

AI agents are software systems that take a goal, independently plan a path to achieve it, and carry that plan out in several steps — including access to external tools such as email, calendars, CRM or databases. The difference from a classic AI assistant lies in the word “carry out”: an assistant drafts a suggestion, an AI agent gets the job done. It reads the request, checks the available data, makes decisions, calls systems, and verifies whether the result is correct.

The term predates ChatGPT. In computer science, the term software agent has been used since the 1990s for a program that perceives its environment and acts within it towards a goal. What’s new is that large language models (LLMs) give these agents a flexible brain: they understand natural language, reason over unstructured information, and decide on the fly which tool makes sense next. The combination of LLM, tool access and memory is what “AI agents” means in 2026.

How Does an AI Agent Work?

An AI agent works in a loop of perceiving, planning, acting and checking. Technically, it consists of four components that work together:

• ▸Language model (LLM) as the reasoning core: models such as Claude, GPT or Gemini break the goal down into sub-steps and decide what to do. The model itself can’t “do” anything — it produces text that gets interpreted as an instruction.
• ▸Tools and APIs: through defined interfaces, the agent accesses external systems — sending emails, creating records in the CRM, querying a database, searching a web shop, or executing code. Each tool is a capability the agent calls on when needed.
• ▸Memory: short-term memory holds the context of the current task; long-term memory (often a RAG system, see below) supplies company knowledge, policies and past results.
• ▸Orchestration and control: a control layer limits which actions are permitted, when a human must approve, and when the agent stops. Without this layer, an agent is a security risk.

An everyday example: an agent is given the task “Prepare a quote for the enquiry from Müller.” It reads the email, identifies the requested items, looks up the prices in the price list, checks the CRM for an existing framework agreement, drafts the quote using the template, and submits it for approval. That’s five steps across four systems — a chatbot could manage the first one at best.

AI Agents vs Chatbots vs AI Assistants

Marketing happily blurs these terms. But the differences matter in practice, because they determine benefit, cost and risk:

What Types of AI Agents Are There?

AI research classifies agents by their decision logic — from the simple reflex agent that reacts to rules, to the learning agent that improves its behaviour from feedback. For businesses, a classification by task and architecture is more useful:

• ▸Task agents: specialised for a clearly defined process — checking invoices, coordinating appointments, classifying support tickets. The most common and most sensible entry point for SMEs.
• ▸Conversational agents: hold dialogues by chat or phone and are allowed to act on them, for example booking appointments or changing order status. This also includes the AI phone assistant.
• ▸Research and analysis agents: search the web, documents and databases, and condense the results into reports — for market monitoring or tender analysis, for example.
• ▸Coding agents: write, test and fix software largely on their own. Tools such as Claude Code, OpenAI Codex or GitHub Copilot Agent have noticeably changed software development in 2025/26.
• ▸Multi-agent systems: several specialised agents work together under an orchestrator — one researches, one writes, one reviews. Powerful, but complex to operate and debug.
• ▸Browser and computer-use agents: operate websites and desktop applications like a human when no API exists. Useful for legacy systems, but slow and error-prone.

Use Cases: Where AI Agents Make Sense for SMEs

The rule of thumb from our projects: a process suits an agent if it is recurring, has clear rules, spans several systems, and its result is easy to check. Concrete examples:

• ▸Lead qualification: reading incoming enquiries, researching the company, scoring it against criteria, creating it in the CRM and notifying the right salesperson — more in our article on AI in sales.
• ▸Quote preparation: evaluating the enquiry, pulling line items from price lists, drafting the quote from the template, and submitting it for approval.
• ▸Incoming invoices: reading invoices, matching them against purchase orders, flagging discrepancies, and generating a booking suggestion.
• ▸Customer-service back office: classifying tickets, answering standard cases from the knowledge base, and handing special cases to staff with prior research attached.
• ▸Onboarding and HR administration: setting up accounts, sending documents, coordinating appointments, ticking off checklists.
• ▸Content operations: generating product descriptions from master data, maintaining them in several languages, and feeding them into the shop system — always with a final human check.

Less suitable are processes with a lot of discretionary judgement, rare one-off cases, and anything where a mistake is expensive or irreversible — payment approvals, contractual commitments, HR decisions. Here a human belongs in the loop, as we argue in our article on human-in-the-loop.

Which AI Agents Are Available on the Market?

In 2026 the market splits into three tiers that differ in flexibility, effort and price:

There’s no blanket answer to which AI agent is “the best” — the real question is which one fits your process and your system landscape. If you live in Microsoft 365, check Copilot agents first. If you want to automate processes across many tools, n8n gives you more flexibility, as our comparison n8n vs Zapier shows. If you need to map a business-critical process with high demands on control and data protection, custom development is hard to avoid.

How Much Does an AI Agent Cost?

Costs break down into three blocks, and most budgets make the mistake of only looking at the first:

• ▸Build: hours for standard software, a few days for low-code platforms, two to twelve weeks for custom agents. The biggest cost item isn’t coding — it’s process analysis, data cleanup and testing.
• ▸Ongoing model costs: agents consume far more tokens than chatbots because they take several steps and tool calls per task. A realistic range is €0.05 to €2 per completed task, depending on model and complexity. At 1,000 tasks a month, that’s €50 to €2,000.
• ▸Operation and maintenance: monitoring, adjustments when processes change, model updates. Budget 10–20% of the build cost per year.

You’ll find a detailed calculation with ROI examples in our article What Does AI Automation Cost?. The key figure isn’t the price — it’s the cost per completed task compared with manual handling, and the error rate: an agent whose results all need to be double-checked saves nothing.

Building an AI Agent: Your First Agent in 6 Steps

Building an AI agent is technically easier in 2026 than ever before. The difficulty lies in the scoping. This is how we approach it in projects:

• ▸1. Choose a process, not an agent: start with a task that measurably costs time today and has clear rules. “Match invoices against purchase orders” is a good starting point; “automate procurement” is not.
• ▸2. Document the process: what inputs, what systems, what decision rules, what exceptions? What a human can’t describe, an agent can’t reliably execute.
• ▸3. Define tools and data access: which systems may the agent read, which may it write to? Apply the principle of least privilege — every write permission is a risk. Company knowledge is connected via a RAG system, not copied into the prompt.
• ▸4. Set guardrails: which actions require approval? When does the agent stop and hand over? Which inputs are off-limits (health data, for example)? These rules belong in the orchestration layer, not just in the prompt.
• ▸5. Test with real cases: run through at least 50 real historical cases, compare the results against manual processing, and measure the error rate. Only go live once the error rate is stable — and keep an approval step in place at first.
• ▸6. Measure and expand: track processing time, error rate, and cost per task. Only once the first agent runs stably does the second process follow — not before.

Advantages and Challenges

The advantages are obvious: agents work round the clock, scale without hiring, apply rules consistently, and complete routine tasks in seconds rather than hours. The challenges are discussed less openly:

• ▸Reliability: language models are probabilistic. An agent that acts correctly 97% of the time makes 30 mistakes per 1,000 cases — in live systems. Whether that’s acceptable depends on the process.
• ▸Security: agents with system access are an attack target. Prompt injection — manipulated content in emails or documents that tricks the agent into unwanted actions — is the central security problem of 2026.
• ▸Cost control: an infinite loop, or an agent that makes 40 tool calls per task, blows through budgets fast. Limits per task are mandatory.
• ▸Organisational: Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 — due to rising costs, unclear business value, or inadequate risk control. All three are scoping problems, not technology problems.

The same Gartner analysis expects that by 2028 at least 15% of day-to-day work decisions will be made autonomously by agents, and that a third of enterprise software will include agentic features. The technology is coming — the question is whether your first project ends up among the 60% that deliver value.

Are AI Agents GDPR-Compliant?

An AI agent is only as GDPR-compliant as its architecture. Three factors decide this. First, where the language model runs — API providers with a data processing agreement (DPA) and EU data processing are now standard; for especially sensitive data, EU-hosted or local models are an option. Second, what data the agent even gets to see — data minimisation applies to tool access too. Third, whether decisions with legal effect on individuals are being made: Art. 22 GDPR prohibits fully automated individual decisions with significant effects — credit refusals or rejecting job applicants, for instance — without human involvement.

On top of this comes the EU AI Act: agents that interact with people have been subject to the AI labelling obligation since August 2026. Agents in high-risk areas such as recruitment or credit decisions face additional obligations. For the typical back-office agent at an SME, it comes down to GDPR, transparency, and clean documentation.

Our Verdict

AI agents mark the shift from “AI that writes text” to “AI that gets work done.” For SMEs, they are ready in 2026 — not as universal digital employees, but as specialised executors of clearly described processes. Start small, take guardrails seriously, and measure cost per task, and you’ll build a genuine productivity advantage. Anyone who wants “an AI agent” without knowing which problem it should solve ends up in Gartner’s 40%.