Make or Buy AI projects are rarely to be created alone
In projects related to machine learning (ML) and artificial intelligence (AI), the question arises as to how much companies want to develop themselves and where they turn to external help. The decision on this is multifaceted. A structured approach can help.
(Image: Initiative for Applied Artificial Intelligence)
AI projects differ from other software projects in different areas and so do the necessary make-or-buy decisions. A paper of the Initiative for Applied Artificial Intelligence, developed at the Entrepreneur branch of the Technical University of Munich, UnternehmerTUM, deals with the content and structure of such decision-making processes.
The authors recommend first classifying projects according to strategic value and “unfair advantage”. The latter are company-specific skills and data that others do not have. Projects that meet both conditions should definitely be done by yourself, projects that do not meet either of the two conditions may be discontinued.
The other two cases – i.e. either high strategic value, but missing skills or data, or skills and data available, strategic value uncertain – require further considerations.
Basically, three approaches to make-or-buy are conceivable: The complete in-house development including in-house training of the ML model is very complex. Hybrid projects use pre-trained ML models or advanced modules, such as those offered by AWS or Google, and combine them with their own development efforts. Finally, the external purchase of complete AI applications is also possible. Here, however, the integration question arises.
There are four different levels to consider in AI projects, each of which requires its own make-or-buy decisions: The infrastructure level includes systems and processes for the development, training, deployment and maintenance of AI applications. The data level describes the existing or required data and their possible sources, including acquisition or synthetic generation.
The level of ML skills includes the basic functions of AI/ML products, namely artificial vision, hearing, speech comprehension and movement, as well as the ability to discover (for example, of connections), search and plan, predict and create new things. Depending on the project, only a few of them will be needed. If the company does not have them itself, a supplier must be found for this.
The top level is the concrete application, whether it is visible to the users or not. Applications are based on the resources of the three lower levels.
Whether you buy or develop it yourself is influenced by six factors in every make-or-buy decision:
- How big are the competitive and strategic advantages of the project (efficiency gains, cost reductions, new functions …)?
- How important is the control of the ML model (e.g. lock-in risks, regulatory requirements)?
- How much can the company learn from the project for its own benefit?
- Are resources (data or skills) used in the project that significantly distinguish one’s own company from competitors?
- How efficient are the envisaged solutions of external suppliers in the long term and in relation to the own project goals?
- What are the costs of the different delivery options?
At each stage of the product life cycle, these decisions have to be made anew, but they are of different relevance depending on the phase. So the idea generation can most likely be carried out completely on your own – but you should definitely keep in mind whether there is perhaps already a finished ML product for the intended purpose. This also applies to the PoC, because an already existing tool can be an important sign of basic feasibility here. In addition, one should include possible learning effects in the considerations. They can partly compensate for the lack of practical benefits.
Closer to the application are the creation of a Minimum Viable product (MVP) and finally the scaling of the product. In the MVP phase, a solution alternative available on the market can make the decision easier. For scaling, the make-or-buy question is the most important. Above all, the long-term costs of possible alternatives must be taken into account.
How to approach scaling to make a make-or-buy decision depends on how strategically valuable a project is. Projects with low strategic value should be connected to strategically important projects in some way, if possible, instead of investing resources there.
If a project is rated sufficiently high, one should decide whether the company should own the ML model itself, for example, in order to keep important knowledge in-house. If this is answered in the affirmative, the degree of one’s own development efforts depends above all on the availability of one’s own data and capability resources as well as the costs. Depending on this, a complete in-house development, a hybrid development or the use of an external application can be considered, if otherwise the cost disadvantages become too great.
Partners also for self-creation
Even with self-creation, you usually need partners, for example internal ones, with whom you share existing AI resources. Even external developers (freelancers) or research institutions are more often involved in such projects. In the case of external parties, care must be taken to secure intellectual property. Academic research is often still far from practical maturity. Here, particularly clear agreements must be made about who is entitled to what results in the end.
Those who prefer hybrid or external creation usually work with other types of partners. For example, with cloud providers that provide AI tools or process chains. Questions of data ownership and the integration of solutions into your own infrastructure can pose problems.
Other important partners are start-ups, which may also be taken over as a strategic resource, larger solution providers and integrators as well as suppliers of software that can be embedded in their own products.
All suppliers should be carefully qualified. In the PoC and MVP phase, the most important thing is the mastery and quality assurance of tools, sufficient insurance and documentation. In the scaling phase, sufficient hardware resources, maintenance processes and tools, certificates, a secure handling of the data as well as an orderly handover and training phase must also be taken into account in the selection.
Benchmarks are only of limited significance
Benchmarks for evaluation are usually based on the processing of generic data and are therefore only of limited significance. With external solutions, your development potential, for example in the form of a clear roadmap, may be more important than the current performance. Finally, you should avoid vendor lock-in, for example by training the model used with your own data or ensuring the transferability of the model. If all this is taken into account, long-lasting and in-depth partnerships between AI users and AI suppliers can certainly arise.
At the end of the search for a partner there is usually a contract. As a rule, it should cover five topics: data acquisition, the parameters of a feasibility study, a profitability calculation, the desired quality and scope of the project and, finally, its integration into the rest of the IT environment.
In addition, appropriate performance standards for the completed development project should be defined and determined by which methods they are to be measured. It is also necessary to regulate who gets access to the training data or how to distribute the generated value. This also includes the right to buy out intellectual property rights to models and data under certain conditions. Finally, the aspects of data protection and risk distribution as well as risk prevention should be regulated in the contract.