Why are technical requirements important for the use of AI?

An AI can be impressive in a demo and fail in everyday use, for example if it cannot access the necessary information, if the results do not flow back into the existing programs, or if processing is too slow. Early checks save money: AI functions are often already included in the software used or available as a ready-made online service (Software as a Service, SaaS). Additional in-house technology is only worthwhile if the specific use has special requirements. Also think about operations: self-developed solutions provide flexibility but require maintenance and expertise in-house. With an external service, the provider takes over many tasks. In return, a dependency arises, for example on their requirements, prices or interfaces. For SMEs in Bavaria the following applies: Use what is already there and add only what the specific use really requires.

In 5 steps to the right technical requirements:

• Step 1: Describe the workflow in concrete terms
  Do not start with servers or the cloud. First describe the use: Which information does the AI need? Which task does it perform? How often is it used, by how many people, and does the result have to be available immediately (real time) or is “in the background” sufficient (e.g. overnight)? A small document assistant has different requirements than a camera inspection in ongoing production.

• Step 2: Clarify systems and data sources
  Where does the information come from, where do the results go? Does the AI need to be connected to document management (DMS), inventory management (ERP), customer management (CRM) or a machine? Is it enough to upload files manually, or should the exchange run automatically? Check whether the existing programs offer suitable interfaces (APIs). These are defined “connections” for data exchange. Also define who is allowed to access which data. Involve the IT managers and the people who work with the programs every day. They know special cases that are not described in any manual.

• Step 3: Match needs with what already exists
  Get an overview of what is already there: Does the software in use already have AI functions? Can an external online service be integrated cleanly? Only if that is not enough does additional technology come into play. Consider only what matters for the use: computing power, storage, network, existing programs, access rights and connections. This also includes the question of where the AI does its processing: in-house (on-premises) or in the cloud. Locally, you retain full control over data and systems, but you bear responsibility yourself for procurement, operation and security. In the cloud, you pay based on usage and can adjust capacity flexibly. For sensitive data, you should check where the provider processes it (e.g. in data centers in the EU) and which protective measures they offer. Example: Local image analysis may require a more powerful graphics card. If the processing runs in the cloud, this capacity does not have to be on-site.

• Step 4: Choose the simplest suitable solution
  Compare three approaches: 1) ready-made standard application, 2) external online service with integration, 3) customized or self-operated solution. In between are no-code and low-code platforms: with prebuilt components, processes can be automated and programs connected even without deep programming knowledge. The further away from the standard you go, the higher the effort usually is for integration, operation and maintenance. Clarify: Who maintains updates? What happens if a service or a connection changes? In-house solutions are not automatically better. External services are unsuitable if important systems cannot be integrated or if the dependency would be too great.

• Step 5: Test in the real workflow
  Before making major investments, carry out a realistic test: Can the AI access the intended sources? Does data exchange between the programs work? Are speed and reliability sufficient? Also simulate disruptions (e.g. access briefly unavailable, connection changed, online service not reachable). Also check how the solution deals with unusual or deliberately misleading inputs. Only then will it become clear whether a demo can become a reliable solution for everyday use.

Use it now: AI computing power from Germany, publicly funded for business and research

You can find more information on the AI Compass topic "Technical requirements" in our download portal.