What an AI agent is and what it is not
The term is used loosely. Anthropic, a developer of language models, draws a useful line in its guide to agents: in workflows, the language model and tools are connected through predefined code paths. In agents, the language model directs its own process and tool use (Anthropic, December 19, 2024). The analyst firm Gartner also observes "agent washing": vendors relabel existing assistants, RPA and chatbots as agents without much agentic capability (Gartner, June 25, 2025).
The table shows who determines the steps in each type of solution:
| Solution | Who determines the steps? | Typical use |
|---|---|---|
| Chatbot | the conversation | Answering questions, for example about products or documents (with RAG) |
| RPA | a fixed sequence of clicks in user interfaces | Legacy systems without an interface |
| AI automation | the workflow in code; the language model reads or classifies at single steps | Recurring work with clear rules, such as reading and posting invoices |
| Voicebot | the goal of the phone call | Answering calls, taking the request, booking appointments |
| AI agent | the language model, within the tools and limits you set | Tasks where the next step depends on the content |
An AI phone assistant is therefore an agent for exactly one task, the call. How it works is covered in the AI voice agent guide. The agents in this article work inside your systems and documents.
What companies use AI agents for
Agents pay off where a task involves a lot of reading and the next step depends on the content. Three examples:
- Requests and quotes: The agent reads a request by email or PDF, spots missing details, pulls data from the ERP and drafts a quote.
- Tenders: It checks requirements against product documentation and certificates and shows which points are covered and which are not. A RAG system supplies the company knowledge.
- Support cases: It searches manuals and earlier tickets for solutions and proposes an answer to the team.
When an agent is too much
Not every task needs an agent. Anthropic recommends finding the simplest solution and adding complexity only when needed; agentic systems often trade latency and cost for better results (Anthropic). Gartner writes that many use cases sold as agentic do not need an agentic implementation (Gartner). Three questions help you decide:
- Does the next step change with the content? If not, a fixed workflow is enough.
- Can you live with an error, or does a person check the result before it takes effect?
- Do the systems involved have interfaces (APIs)?
Where the workflow is fixed, AI process automation is the more predictable and cheaper route.
Limits: sign-off, GDPR and the EU AI Act
Sign-off by people
An agent works on its own, but not unsupervised. According to Anthropic, agents can pause for human feedback at checkpoints or when they hit blockers, and they get stopping conditions such as a maximum number of iterations. In practice: the agent creates drafts, a person approves before anything goes out, and every step is logged.
Decisions about people: GDPR Art. 22
If you process data of people in the EU, they have the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them or similarly significantly affects them (GDPR Art. 22, with exceptions in paragraph 2). An agent that alone screens out job applicants or rejects loans would be a typical case. It may prepare, a person decides.
EU AI Act
For companies with EU business: agents that communicate with people, for example in chat or on the phone, fall under the transparency duty of Art. 50(1). People must be told they are dealing with an AI system unless that is obvious from the circumstances. For high-risk AI systems, Art. 14 requires effective oversight by natural persons (Regulation (EU) 2024/1689). Whether an agent is high-risk depends on its field of use. The classification belongs in your AI governance; the article on the EU AI Act and AI phone assistants shows an example. This guide does not cover US state or federal law.
Costs and risks
There is no price list because the effort depends on your starting point. Three factors decide most of it: the number and type of connected systems, the number of exceptions and approval steps, and the number of steps a task needs. On top come running costs for model calls, which are higher for agents than for a fixed workflow.
Gartner quantifies the risk of planning too big too early as a prediction: over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. Gartner advises pursuing agents only where they deliver clear value (Gartner, June 25, 2025). We see it the same way: start with one task whose value you can calculate, then expand.
How titanspear.ai builds AI agents
We build agents as part of AI process automation on our own automation platform. We connect your systems through interfaces (APIs), and a person approves before anything goes out. We sign a data processing agreement with you under Art. 28 GDPR, and before anything is connected, we settle in the contract whether data may be used to train third-party models. The path:
- Intro call (30 minutes by video, free): Which task costs the most time today?
- Map the process and set the target: steps, exceptions, approvals, expected benefit.
- Connect the systems: through interfaces, with access rights and a log.
- Pilot with real data: one task, one group, measured against examples from your specialists.
- Expansion and support: with a fixed contact person.
The framework for rules and accountability is in the AI governance guide, the entry point for the overall project in our AI consulting. How to build AI functions into existing systems is covered on the AI integration page.
Sources and note
All sources opened on October 11, 2026: Anthropic, Building effective agents; Gartner, press release of June 25, 2025; Regulation (EU) 2016/679; Regulation (EU) 2024/1689. This guide is not legal advice. Your legal counsel or data protection officer handles the legal assessment of your case.