How to Integrate AI in Corporate and IT Industry
Artificial Intelligence (AI) solutions are usually first developed under laboratory conditions. In order to subsequently operate a solution productively in corporate IT, organizational and technical questions must be answered:
- Do the results of an AI solution have the quality and reliability that are required for acceptance in the company?
- Is data of sufficient quality and quantity available?
- Does the deployment pay off from a business perspective?
- How can the AI system be integrated into existing application architecture?
- How can it be monitored?
In order to operate an AI system sustainably, it must be continuously updated and readjusted. Ongoing new data can be used to improve the model through continuous learning. All of this requires processes that go far beyond the initial creation of an AI model.
In this article, the authors shed light on the different dimensions that have to be considered when operating AI applications in the corporate environment and show different approaches and alternative actions.
Requirements for using AI
The perception of AI and the progress made there is essentially shaped by numerous media reports of new, sensational results from the research area. However, implementations that result from this generally have an experimental, “proof-of-concept” character and cannot be easily converted into productive systems.
However, developing an AI application for productive use involves much more than simply solving a problem using AI methods. To do this, the requirements of the environment must be considered holistically from the start. These usually go beyond the actual technical problem and have a direct influence on the selection of the AI process to be used, if not on the decision for or against an AI system. Further requirements usually result from the non-functional requirements or the general conditions of IT operation. The following questions should be checked:
- What are the requirements for the accuracy of the AI system?
- Under what conditions can the application take place at all?
- To what extent is it necessary to explain and understand the results?
- How should the application be embedded in existing systems?
- What response times are required? Are requests synchronous or asynchronous?
- How is the quality of the results of the AI system monitored? Is there a feedback mechanism?
- In what form should the AI system learn? Offline? Incremental?
- How does the application interact with other systems?
- What restrictions are there regarding the technology stack?
- What hardware resources are available?
Although many of these criteria stem from the requirements analysis for classic IT systems, they are also critical success factors for the development and use of sustainable AI systems.
Only when the requirements for a solution have been comprehensively recorded does it make sense to select and test suitable AI processes in a systematic preliminary study. As a result, such a study should evaluate the tested methods not only in terms of the usual AI metrics, but also along the recorded requirements.
Quality = acceptance
Misunderstandings, incorrect outputs or answers from AI applications cause smiles or irritations in the B2C environment. In the worst case, the application is uninstalled and no longer used. This option is not available to people who have to use AI applications in a professional context. Decisions of the AI systems therefore have a very high level of accuracy, ie to avoid “false positives” and “false negatives”. While the former signal a finding, even though there is none, “False Negatives” overlooks real findings. At least both cause acceptance problems and – depending on the area of application of the AI algorithm – can have serious effects.
In operational practice, this means additional, manual control by people. How often this is required is influenced by the accuracy and robustness of the AI systems used. With thousands of decisions or evaluations by the AI system, however, just a few per mile of “false positives” lead to a considerable amount of follow-up checks, which the processor can easily classify as unnecessary work. However, since all known AI methods also do not recognize a certain rate of correct hits, damage events will occur in operational use that the AI should actually prevent. Not only has the selected algorithm had a hit rate of the detection algorithm,
If the hits mainly consist of “false positives” and the processor is therefore unnecessarily activated, acceptance, use and benefit of the AI system decrease very quickly. If there are effects of events that are incorrectly not recognized, this impression is reinforced. To put it positively: If the AI system used succeeds in offering the human beneficiary a noticeable relief from his work, this promotes acceptance and motivates further use.
Linked to the acceptance problem is the question of what requirements are placed on the explain ability of the AI results. Because although there are many research activities and first results and processes in this area regarding neural networks, the lack of traceability in neural networks can be an exclusion criterion. Examples of this include legal or regulatory requirements regarding the traceability of decisions as well as requirements for non-discriminatory decision-making. Other methods that are based on statistical methods or rule-based methods offer great advantages in this regard. In general, it will be necessary in the business environment for legal reasons that the results of an AI system can be understood by the human viewer.
In addition to the psychological and motivational effect of the assignment, the business cost considerations must also be taken into account. While the AI-supported relief of the work process pays for the business case created in advance, the falsely reported or overlooked effects cause processing effort. In addition to working hours, this often also requires the use of additional resources (material, transport, etc.). That is why stable and robust systems are essential for everyday use.
Data for the AI
The use of systems that support artificial intelligence is usually based on extensive data collection. The algorithms are trained, developed and evaluated on this.
Accordingly, the first question when using AI systems in business use is the type of data on the basis of which statements must be made. Technical data of sensors on a machine that try to detect signs of fatigue can be classified significantly differently than systems for facial recognition in security areas. In addition to the relevant requirements such as the GDPR, labor laws, company and other agreements may also need to be observed.
Well-intentioned, but poorly made or poorly communicated AI missions create negative publicity and excitement in the media and can easily cause problems that are greater than the actual benefits.
In addition to the type, the quality and quantity of the data available also play an important role. While there is an almost inexhaustible pool of test data and qualifications for the search for cats on private photos, there are situations in the industrial context with little or only limited usable data. The reasons for this can be (too) special questions, technical or organizational problems of the survey or regulatory restrictions.
As a result, certain AI methods cannot show their strengths because there is not enough data available for training. It is therefore advisable to carefully consider which AI method or combination of the same, with the available set of data, can achieve the best results.
One of the big challenges that AI research is currently working on is being able to process data even if it is of poor quality. Well-known examples of this are autonomous driving, which must recognize obstacles or people even when the lighting conditions are poor, it is raining, fog or dust is in the air, or other effects provide a suboptimal source data. The same applies to face and person recognition, such as is used in security areas. IoT sensors that transmit data about the operating status of a machine or system must also do this robustly and reliably, even under adverse conditions such as heat, cold and mechanical stress.
To make matters worse, the above-mentioned challenges are often compounded by the fact that the data collected is classified as a trade secret. This prevents publication in a larger data pool, from which several companies could make optimizations.
The exact opposite challenge arises in situations in which extremely large volumes have to be transported and processed, such as in the area of autonomous driving. If the acquired sensor data is transferred there after a test drive to further optimize the AI algorithm, almost all hardware and network components reach their technical limits. Close coordination and cooperation with dedicated experts in these areas is then essential.
Integration of AI systems in the IT landscape
As a rule, an AI process is used as a sub-module of a larger specialist application. An AI process often supplements an existing system with intelligent suggestions or by replacing manual process steps with automatisms. Therefore, instead of considering the AI process as an integral part of a system, there are many reasons to organize and outsource the AI calculations as independent services:
- AI processes may require dedicated special hardware and must run in a separate operating environment.
- AI methods can be bundled and possibly used by multiple systems.
- The core functionality of AI is agnostic with regard to professionalism and can be worked on by an independent development team.
You go one step further if you consider AI calculations as abstract services that can be addressed via a uniform interface. Together with a generic approach that enables AI services to be created dynamically based on a library of AI processes, this result in further advantages:
- The parameterization of AI processes can be adjusted dynamically.
- AI models can be exchanged without changing the calling application.
- Different AI processes can be provided in parallel and, if necessary, several processes can be combined.
At this point, one should not forget the need for data preprocessing, which is necessary both when creating AI models and when evaluating. This includes the acquisition of data, the filtering, the cleansing and finally the extraction of characteristics. These functions must also be considered as part of an AI architecture, since they may also be required for preprocessing requests to the AI system.
Another module of such AI service architecture can be the training module if it is necessary that the system learns during operation and AI models are to be continuously adapted. To do this, the necessary data preprocessing steps must be taken into account again. The learning process can be triggered either by direct user feedback or by the accumulation of new learning data from other data sources.
The open source frameworks Apache PredictionIO , Hydrosphere.io and Seldon are examples of systems that implement the aspects mentioned in different forms.
From planning to implementation
When implementing AI applications in the industrial context considered here, it makes sense to fall back on the rich pool of existing implementations of AI methods, unless basic research work is to be carried out.
As in every implementation project, the question arises of the quality, timeliness, stability and security of the frameworks or libraries used. In some cases, an individual implementation tailored to the application domain can be significantly more efficient than the use of generic libraries. In any case, it must be taken into account that the selection of usable technologies may be limited by company requirements, IT governance or the target operating environment.
If the legal framework permits the processing of data in the cloud, the appropriate repertoire of AI methods from the cloud providers can be used. This has the advantage that the customized hardware resources are also available there.
Another point to consider is how to deal with updating AI models. If continuous learning takes place, the system behaves non-deterministically from the user’s point of view, as it can return different results if the output data is staggered but identical in content. This makes the presence of an explanatory component all the more important. For performance reasons, however, depending on the AI process, it can be prohibitive to log an explanation every time the AI forecast is calculated, and instead to have a necessary explanation calculated in a separate request.
Operation and maintenance
Following the DevOps concept, integration and deployment processes should also be automated as far as possible for AI systems, so that the system can be updated easily. In contrast to classic systems, the CI / CD process includes the preparation and display of a prepared AI model. Accordingly, integration tests must test the interaction between the AI core and a new AI model before it can be deployed to the productive environment.
Even normal application monitoring monitors the health and normal behavior of all system components. In the case of AI systems, it is also advisable to collect further metrics in order to obtain information about the quality of the predictions of the AI system. This can be difficult if the prediction quality is delayed or cannot be determined for each individual case. To make matters worse, the decline in the quality of predictions can have different causes. These range from the quality of the input data, to the parameterization of the AI process, to a faulty or insufficiently trained model. It follows that the operation of an AI system requires special expertise, which generally go beyond the qualifications of the employees of the typically encountered, classic IT operations. Rather, operating AI applications requires even more overlap between operations, development and technical support than is already being sought in the DevOps context.
AI methods not only bring a new quality to the possibilities of data processing, they also place completely new demands on the implementation and operation in productive corporate use. Which AI method is used not only depends on the technical problem, but is also determined by numerous other framework conditions, organizational and technical in nature. The key success factors are acceptance and ultimately the business benefit that leads to the sustainable use of an AI application.