Agentic-as-a-Service

Many AI programs only respond: you ask a question, you get a text back. A so-called agent goes further. It breaks a task down into steps and works through those steps on its own, one after another. In doing so, it may also use tools, for example searching the internet, sending an email, or changing an entry in a database. Agentic-as-a-Service means: such agents are not built in-house but rented from a provider. The suffix “as a Service” is common in the tech industry and means that a company pays for something as an ongoing service, similar to a streaming subscription, instead of buying and operating it themselves.

Why companies rent agents instead of building them

Developing a working agent is demanding. It requires access to large AI models, computing power, connections to existing software, and experts to oversee the whole thing. For a mid-sized company, this is often unaffordable. A rented service lowers this entry barrier to a single monthly bill.

For providers, this model is economically attractive. Pure chat subscriptions bring in a fixed amount per user. An agent that completes a task entirely, on the other hand, can be measured against the working time it saves. This is exactly what companies like Salesforce, Microsoft, or OpenAI advertise: they are not selling computing time but completed work. Some providers therefore charge per completed transaction, for example per customer inquiry processed.

This also explains why the term appears so frequently in business news. Analysts see in it a market that could be significantly larger than the one for simple chat programs. Whether this expectation will be fulfilled remains open. So far, many agents fail in practice because they make mistakes in longer chains of tasks.

From assignment to completed task

At the core is a large language model, a program that has learned from vast amounts of text how to continue language. Around this model lies a control layer. It gives the model a goal and has it formulate a plan. It then executes the planned steps and reports back to the model what came out of them.

The tools are important. An agent can call programs that run a search for it, issue an invoice, or schedule an appointment. In addition, there is a memory in which intermediate results and previous processes are stored. Without this memory, the agent would have to start over at every step.

With the rental model, there is an additional management layer that does not exist when building in-house. It governs which customer sees which data, how many tasks may run simultaneously, and where a human must give approval. A common misconception is that the agent decides everything on its own. In serious offerings there are limits: transfers or cancellations are usually only presented by the agent for confirmation.

Agents in customer service, accounting, and code

Customer service is furthest along. An agent reads the inquiry, looks up the order in the system, checks the return deadline, and initiates the refund. Previously, three people using three programs would have done this. It works similarly in accounting, where invoices are checked and matched.

A second major field is software development. Services such as GitHub Copilot in its agent variant or Anthropic’s Claude Code take over entire work packages: finding bugs, writing changes, running tests. The developer checks the result at the end.

Anyone who reads this term in the news should distinguish it from two similar words. A chatbot only responds, an agent acts. And “Agentic-as-a-Service” refers to the business model of renting, while “agentic AI” generally refers to the underlying technology. As a private individual, one usually encounters this unnoticed, namely when no human is sitting behind the customer chat of an online shop anymore.