Many companies have set up initial AI pilots in the back office over the past few months: pre-sorting documents, preparing content for the content management system, summarising data from the ERP system.In the pilot phase, this often works convincingly.The transition to productive operation is noticeably bumpier — less often for technical reasons, more often for architectural ones.

Two current specialist articles address exactly this point.A background article on heise developer argues that AI agents can only be safely integrated into existing systems if the architecture provides for a separation of data and actions.An article on Golem.de — labelled as an advertisement by Telekom — describes the other side of the same coin: pilots convince, but new risks arise in productive operation, and success is decided by infrastructure, security and data control.

Why data and actions belong apart

An AI agent is not a pure text generator, but a system that independently executes steps based on inputs: reading files, querying data records, writing content, calling interfaces.As soon as these two roles — processing information and triggering actions — coincide in the same processing layer, every piece of information read in potentially becomes an instruction.

That is the core of the argument: an agent that reads a customer email, a PDF or a database entry cannot reliably distinguish whether text contained therein belongs to the payload data or to its instructions.If the architecture cleanly separates the two, this mixing can be limited technically instead of merely defusing it through formulations in the prompt — i.e. in the instruction to the model.

In practice, this means: actions run via defined, narrowly scoped interfaces, not via free interpretations by the model.Which operation an agent is permitted to trigger at all is defined and enforced outside the model.

The pilot conceals the actual questions

In a pilot project, the framework conditions are comfortable: few users, manageable data volumes, an initiated team that checks results anyway.Errors are noticed before they take effect.It is precisely this protective layer that falls away in regular operation.

The Golem article names the shift: with scaling, risks arise that played no role in pilot operation.Infrastructure, security and data control thus move from a secondary condition to the main role.Both sources converge on the same point but approach it from different directions: one article from the design of the agent architecture, the other from operation and safeguarding.

By way of classification, it is worth noting that the Golem text appeared as an advertisement by a provider.Nevertheless, the problem described coincides with the technical argumentation of the heise article — the agreement lies in the diagnosis, not in a specific product recommendation.

Typical weak points when connecting to CMS and ERP

Anyone coupling agents to existing systems encounters recurring questions in practice:

  • Permissions: With which identity does the agent act?Does it inherit the rights of the calling person or does it work with its own, permanently broad access?
  • Data access: Which subset of the content does the agent actually need — and how is it prevented from reading beyond that?
  • Writing operations: Which changes may it execute directly, which only generate a draft for approval?
  • Traceability: Can it be reconstructed afterwards which input led to which action?
  • Malfunction: How is an agent stopped if it delivers unexpected results during ongoing operation?

None of these questions is new.They are known from the integration of classic third-party systems.What is new is that the acting component does not work deterministically — i.e. with the same initial situation it does not necessarily do the same thing.This shifts the weight from checking the result to limiting what is possible in the first place.

A pragmatic setup

A sober approach can be derived from the requirements described:

  1. Scope the use case narrowly. An agent with three clearly defined operations is easier to secure than a universal assistant.
  2. Define actions as explicit interfaces. Each operation is described in advance, including permitted parameters and value ranges.
  3. Grant minimal permissions. The agent receives access only to the data and functions that the specific process requires.
  4. Build in approval steps where external effect arises. Published content, orders or customer communication can tolerate an intermediate step.
  5. Log. Inputs, triggered actions and results belong in an evaluable log.
  6. Plan operation from the outset. Infrastructure, monitoring and responsibilities are not added afterwards but belong in the design.

Organisation rather than pure technology

Part of the scaling problems lies outside the software.If an engaged team checks the results in the pilot project, it is unclear who does so in regular operation.Responsibilities for approvals, for maintaining the data basis and for responding to anomalies must be named before an agent goes live.

It is helpful not to regard the agent as a standalone project, but as another component in an existing system landscape — with the same requirements for operation, permission management and documentation that also apply to other integrations.

What this means for web and software projects

For projects involving TYPO3, ERP connections and internal specialist applications, the effort shifts forward: the question of which actions an agent may perform via which interface belongs in the concept phase and not in acceptance.Whoever separates data and actions from the outset can extend use cases later without having to rework the security architecture every time.Pilots remain worthwhile — but they should be built from the start on the infrastructure on which productive operation is later to run.

Sources