Knowledge · agent

Creating an AI agent: what you really need (and what you don't)

Create an AI agent without a developer team: the five steps from task through data and instructions to approvals – and when a platform makes sense.

To create an AI agent today, you don't need a developer team – but you do need four things that no model in the world can replace: a clearly defined task, access to the relevant data, precise instructions, and defined limits with approvals. The technology is now the smallest hurdle; most agent projects fail because of unclear assignments and missing context, not because of the AI.

Step 1: Define the task

"An agent that helps me" is not an assignment. Tasks that work well are those that are recurring and clearly defined: checking incoming documents for deadlines, creating weekly summaries, answering and logging calls. Phrase the task so that you could also hand it over to a new employee – including what is not part of it.

Step 2: Make data accessible

An agent without context guesses. It needs access to the knowledge relevant to its task: contacts, tasks, documents, appointments, goals. If this data is scattered across separate tools, that's the first thing to fix – even before the agent itself. The article AI agent with company data explores why this is so crucial.

Step 3: Formulate instructions

Instructions are your agent's job interview: don't just tell it *what* to do, but *under which conditions*. A practical example: "Only create deadlines if the date, obligation, and source are clearly documented – uncertain cases should only be mentioned, not created." Such conditions distinguish a usable agent from one that produces plausible-sounding nonsense.

Step 4: Define limits and approvals

The most important architectural decision: what may the agent do independently, and where does a human decide? A proven principle is to route any external impact through approvals – emails, calls, changes to master data only go out once you have reviewed the proposal. This way the agent can boldly do the groundwork without anything important going out unchecked.

Step 5: Test and refine

Start with a small scope, review the results, refine the instructions. An agent doesn't become good when it's created, but when it's trained in – like an employee during a probation period. Plan a few weeks for this and note which corrections you make repeatedly: that's exactly where the next instructions come from.

Build it yourself or use a platform?

Building your own (frameworks, APIs, your own infrastructure) makes sense for very specific requirements – but it means ongoing responsibility for operation, security, and data connection. The faster route for most companies is a platform where the agent, data, and approvals already work together.

With webRichtung agent, you don't create an agent from scratch but configure your personal AI employee: it knows the context of your company through the platform data, you give it personal instructions, you activate automations one by one – with adjustable settings like a minimum confidence level – and you decide on follow-up questions before anything with external impact happens. The details are shown in the documentation on automations & approvals.

Markdown · Text