AI automation in medium-sized businesses: defining a viable pilot.

Start with a clearly scoped task

For a medium-sized company in Germany, the first question is not which AI tool is particularly well-known. What is crucial is which specific task the team completes repeatedly and what could be identified as better processing. A pilot should be able to integrate into ongoing operations.

A suitable project could, for example, be the structured preparation of incoming inquiries. This is an illustrative example: which fields are needed and whether AI even makes sense must be checked in the actual process.

Distinguish between rules, AI and business decisions

An existing CRM value can often be processed according to a clear rule. Summarizing an unstructured conversation, on the other hand, can be a task for AI. A binding price decision or the approval of an offer requires a defined area of ​​​​professional responsibility.

Therefore, before implementation, describe what the system is allowed to prepare, what will be processed automatically and when a human will check. A vague order such as improving sales with AI is neither a reliable specification nor a sensible acceptance condition.

Account for existing systems and data flows

Note the CRM, email, platform, and project tools involved. What data is already available? How are contacts assigned? Where are accesses or interfaces missing? The quality of an AI result does not resolve conflicting customer master data.

Involve the people responsible for IT, data protection and information security at an early stage. The purpose of use, permissions, data transfer, retention and provider conditions must match your requirements.

Compile a small, meaningful test set

The pilot needs normal cases and deliberate exceptions. Check for complete and incomplete requests, duplicate contacts, conflicting information, and technical outages. Appropriate test data must be allowed to be used for this purpose.

• Record the expected result for each case before the test.

• Visibly mark errors, uncertainties and missing information.

• Check whether a second processing creates duplicate actions.

• Determine responsibility for handling an exception.

• Evaluate the output with the employees who will later work with it.

Agree on acceptance criteria and operations before implementation

Record the conditions under which the pilot is taken over, adjusted or terminated. Suitable criteria can be completeness, comprehensible allocation and processing effort. A sheer number of automatic executions says little about whether the process is useful for the team.

Also document who checks errors, who is allowed to change rules, and how a process can be stopped. An established process needs support. New functions and further integrations are additional development and should remain visible in scope.

Collaboration with Systemlink from Switzerland

Systemlink works from Wädenswil in the canton of Zurich with companies in Switzerland and Germany. A project can be narrowed down in an initial online discussion. We look at processes, systems and data together and then agree on a concrete next step.

This collaboration does not require your team to relocate or use a new platform. Whether existing applications are expanded or individual parts are rebuilt depends on the specific project.

Related to this topic: AI automation · Ops automation · Automate business processes: where do you start?