AI agent development for Austrian companies

Agent, chatbot or automation

The words get mixed up, so a short map helps. A chatbot answers questions in a conversation. An automation follows a fixed path of steps. An agent decides which steps to take, in what order, using tools, and can handle cases that a fixed path cannot. That freedom is useful and also the source of risk, so agents need clear limits.

We often suggest starting with a simpler tool. If a fixed workflow does the job, it is cheaper and easier to check. Our guide on agents compared with chatbots and the page on AI automation in Austria show where each fits.

Tasks Austrian firms give to agents

Each row follows the same pattern: the agent prepares, the person decides. That pattern keeps quality high and trust intact.

Guardrails we build in

• Narrow permissions. The agent gets read access first and write access only where needed, with separate keys per tool.
• Approval gates. Sending email, changing prices, posting publicly and moving money always need a person's click.
• Logging. Every step, tool call and answer is stored so mistakes can be traced.
• Limits. Caps on cost per run, number of steps and number of messages sent per day.
• Testing. A fixed set of real cases is rerun before each change to catch regressions.

Our AI agent security checklist covers prompt injection, tool abuse and data leakage in more detail.

Data protection and the EU AI Act

If an agent reads personal data, GDPR applies, and the Datenschutzbehörde supervises compliance in Austria. We limit the data the agent sees, choose providers with suitable data processing terms and record the purpose. EU data residency options exist with some providers, and we describe what is available for your case before you decide.

The EU AI Act adds duties that depend on the use case and are being phased in. Chat-style agents that talk to the public generally need to make clear that the user is dealing with an AI. Agents used for decisions about people, such as hiring, can fall into higher-risk categories. This is a moving area, so check the current position with your lawyer.

Where an agent touches employee data or monitors staff work, involve your Betriebsrat early, because works agreements can apply.

Models, hosting and running costs

An agent needs a model, a place to run and a way to be watched. We pick the model per task: a small, cheap model for sorting and extraction, a stronger one for planning and drafting. Splitting the work this way often cuts running costs a lot without hurting quality, and it avoids tying you to one provider.

Hosting depends on your data. Low-risk agents can run on a managed platform. Agents that read customer or employee records may run on a server you choose in the EU, with keys stored in a secrets manager and access limited to named people.

• Monthly model usage is estimated from your volumes before the build, with a cap you set.
• Alerts fire when cost, error rate or run time moves outside the normal range.
• A pause switch stops the agent at once, without a developer.
• Prompts and tool definitions live in version control, so any change can be reviewed and rolled back.

How a build runs

1. Pick one taskWe choose a task that is frequent, valuable and easy to check, and write down what a good result looks like.
2. Build the first agentTools, prompts and limits, tested on a sample of your past cases.
3. Run in shadow modeThe agent prepares its output but a person does the work, and we compare the two for a couple of weeks.
4. Go live with gatesApproval steps stay on, and we widen the agent's freedom only as the results justify it.

We work from Pune with no Austrian office. A 10:00 call in Vienna is 13:30 or 14:30 for us. The AI agent development page has the standard scope.

Cost

Projects start at $1,500 and extra work is $30 an hour. Model usage is billed by the provider to you, and we estimate it from your volumes. The AI agent cost guide shows what drives price: the number of tools, the testing effort and the level of autonomy.

Ongoing care is optional. Some clients take over the agent after handover. Others keep us for a few hours a month to review logs, update prompts when their process changes and test new model versions before switching.