AI agents for controlled business processes | Anexia

In AI agent development, defining the boundaries starts with the tool catalogue. For each tool, inputs, permitted actions and possible consequences are described. Reading delivery information is fundamentally different from changing a delivery date. An agent's rights are therefore not simply taken over from the rights of the employee running it. Actions with particularly significant consequences require a separate review or approval step that remains effective even with unusual model responses.

For agent automation, we define a completion state and rules for termination. A process must be able to recognise when information is missing, a task is complete or no further automatic action would make sense. In addition, there are limits on retries and tool calls. In AI agent development, intermediate states are saved so that a technical error does not trigger an uncontrolled repetition of the entire business process. This is particularly relevant when external systems do not support a shared transaction. The initial production scope remains limited to clearly described process variants. Additional autonomy is extended on the basis of observed failure cases and proven controls. This way, agent development grows along specific capabilities and verifiable operational limits.

Scope of services for AI agent development

The agreed scope for AI agent development can include the following deliverables: tool catalogue, state model, approval rules, execution log and test environment for failed actions. For each key deliverable, we define what it is for, who reviews it and in what form it is handed over. This allows development, operations and procurement to assess the same service. Vague umbrella terms in the proposal are replaced by specific work packages.

Agent automation also includes agreeing the necessary interfaces with your team. We clarify access, contact persons and available testing options before the relevant implementation steps. Changes to the scope are described together with their impact on effort and sequence. This gives you a clear overview of delivered results, outstanding prerequisites and additional requests.

The key technical decision in AI agent development

A deterministic workflow remains responsible for binding process steps. The model is only given room to decide where variable content makes this necessary. For agent development, we document the relevant assumptions and explain the reasons for the chosen option. Key alternatives are considered against the same business requirements. This makes later adjustments easier if load, data volume or organisational responsibility change.

A sound decision also takes maintainability and handover into account. In AI agent development, we check which skills are available internally and what additional support is needed. Technical possibilities are thus linked to a realistic operating model. The client can see which tasks will recur after the project and which decisions remain with them.

Repeated calls can create duplicate orders or send messages several times. Idempotency and transaction-based keys are therefore part of the interfaces. In agent automation, we treat this point as a specific review question. Each risk is linked to the affected process, its possible impact and a suitable measure. Where prerequisites are still missing, the statement is explicitly left open until the necessary investigation has been carried out.

We test unauthorised tool calls, loops, timeouts, restarts and correct escalation to responsible people. For AI agent development, we record the starting situation, the tested configuration and the result. Acceptance refers to the agreed scope. Open limitations are documented visibly and assigned to an owned next step. This allows your team to distinguish between proven functionality, a deliberate decision and an assumption that has not yet been confirmed.

Project example as an illustrative scenario

The following scenario is an application example and not a customer reference. A purchasing team wants to obtain delivery information, consolidate responses and prepare changes in the merchandise management system. In AI agent development, the initial scope would be limited to read queries and change proposals. Orders, price changes or binding commitments would initially require explicit approval. This would make it possible to examine the benefit without automating the entire purchasing process straight away.

For agent development, every connected function would need a clear purpose, permitted inputs and limited permissions. The test set would include missing supplier data, contradictory responses and aborted processes. A repeated call after a timeout would be particularly relevant: it must not trigger a second order or a duplicate notification. Content from a supplier email would have to be treated as business data and must not grant any additional technical rights.

An expansion of the AI agents would only be considered once the responsible people can follow proposals and stop running processes. Completed tasks, necessary interventions, incorrect actions and total processing costs would be evaluated. The planned handover would include tool contracts, a permissions concept, approval rules and logs of representative processes. Which actions would later be possible without individual confirmation would remain a deliberate decision by the client based on this evidence.

Our experience and our engineering approach

We combine custom software development with the technical tasks of introduction and further development. For agent development, we bring the perspectives of application, integration and operations into a shared workflow. Business requirements remain linked to technical decisions. You work with a defined scope of services and traceable results.

Our experience should be verifiable for your specific project. That is why, in the technical discussion, we explain how we examine the relevant risks, which deliverables are planned and how your team can assess them. For AI agent development, we match the roles involved to the actual need. Available references are presented with their respective project and company context; confidential customer details are not used as general marketing claims.

Effort for AI agent development

Complexity arises from states, external dependencies and error handling. The number of connected tools alone is not a suitable cost estimator. We therefore calculate AI agent development on the basis of a clearly defined scope and explicitly stated prerequisites. An initial analysis step can make sense if key technical information is still missing. Work packages and acceptance criteria can then be agreed more reliably.

Agent development can be organised as a limited project, as a joint implementation with your team or with subsequent support. Rights of use, access, documentation and responsibilities are clarified in the contract. Ongoing operating hours, response times and external consumption costs are subject to separate agreements. This makes it clear which services are provided once and which tasks recur after the introduction.

What AI agent development does not cover

Autonomous decisions without agreed control limits are not a standard part of this offering. For agent automation, we state this limit in the proposal itself. Any additional services required are included with their business rationale and are not silently assumed. This allows your team to plan suitable internal skills and divide the overall work sensibly among those involved.