AI software and LLMs, a decision guide: buy or build in-house?

Is it worth developing your own large language model (LLM), or is it strategically smarter to use existing models and buy AI software? LLMs have given artificial intelligence its breakthrough: applications like ChatGPT proved in a very short time that they have the potential to multiply company productivity. For decision-makers at large and mid-sized companies, the question has long since stopped being whether and is now only how best to integrate generative AI into their value chain.

The decision is complex and anything but black and white. With this article I would like to give you an overview of the advantages and disadvantages of the various strategies and of the factors you should weigh up in making your decision.

‍

Basic considerations before introducing LLM-based AI software

Which option turns out to be right for you depends not least on the resources available to you, your goals and your specific requirements. The following questions will help you start your research with clarity:

• Which option turns out to be right for you depends not least on the resources available to you, your goals and your specific requirements. The following questions will help you start your research with clarity:
• Strategic frame: what role should LLMs play in your business strategy in the long term? The more central they are to your future market positioning, the easier it is to justify high internal development costs.
• Customization: how important is it to you to be able to adapt the functions of the AI software? If standard functionality already delivers leaps in productivity, buying AI software may be enough.
• Copyright: how important is it to you as a company to keep the rights to the data being processed? When you buy an LLM, the rights usually belong to the software provider, whereas developing your own makes it possible to keep the intellectual property in-house.
• Security: which security standards do you expect from LLM providers? Check whether LLM providers meet your internal requirements or whether those could be implemented better in a custom development.
• Budget: what budget is available for AI integration? In-house development does come with higher upfront costs. But if you buy AI software, licence fees can become a cost driver as usage grows.
• Internal expertise: what internal knowledge of LLMs and AI exists in your company? If you first have to build up internal know-how, integrating AI into your processes in the short term is realistic only with an external partner, or the buy model is the better alternative for the transition.
• Legal expertise: do you have the expertise to reliably settle all legal questions around developing or using an LLM, on data protection, licensing rights and liability for example?
• Data availability: do you have relevant internal data in sufficient quality and quantity to train an LLM yourself? If the data basis is missing, that argues more for buying a pre-trained LLM.

Advantages and disadvantages of the various options

In fact you are not facing the decision "buy AI software or build it yourself". You have these four possible options for introducing an AI solution:

• You can use pre-trained AI applications.
• You can train an LLM yourself or with support from experts.
• You can adapt open source LLMs.
• You can use another kind of machine learning for your AI project.

Let us take a closer look at the advantages and disadvantages together.

1 Buy AI software or an LLM

The fastest way to integrate AI into a particular process in your day-to-day business (process digitalization) is to buy AI software. You then use commercial LLMs in tools such as OpenAI's ChatGPT or Google DeepMind's AlphaFold. For a subscription fee you can integrate the LLMs into your own applications via API or use the software directly in the browser.

In many areas (marketing, R&D, development) AI use of this kind is becoming the standard. To avoid efficiency disadvantages in the medium term, your company should at least keep pace here. The scope for individual adaptation with off-the-shelf LLMs is limited, though: the models are not trained on your data, and prompting with your own data can improve the quality of results but cannot change the model itself. Major competitive advantages or leaps in innovation with standard LLMs are rather unlikely.

‍

Advantages:

• Use out of the box
• Simple usability, no deep technical knowledge required
• Well suited to many standard use cases
• Support and updates from the provider
• Providers usually ensure legally compliant and secure use

Disadvantages:

• Less transparency about the model's sources and how it works
• High licence fees as user numbers grow
• Handling internal company data can be a challenge
• Very limited scope for adaptation

2 Develop a language model internally

If you want sensitive data to be processed securely by an AI and you have an industry-specific use case in mind, you will quickly reach the limits of commercial LLMs such as GPT-4 or BERT. Developing your own language model makes AI results more reliable, improves data security and allows more versatile use within your company's digital ecosystem.

The disadvantage here: developing a high-performance LLM needs a team of specialists and, even then, is not a short-term project but can stretch over months. And as with hosting, your company pays for the transparency and control of the LLM with ongoing costs for maintaining and optimising the system.

Advantages:

• Tailored to your individual data and use cases
• Full control over model, architecture and performance
• Sensitive data stays in-house
• Flexible integration into existing systems and databases

Disadvantages:  

• A specialised team is required
• Of little use without sufficient volumes of clean data
• Comparatively high cost in money and time
• Maintenance and continuous upgrades have to be handled by you
• Responsibility for meeting security standards and legal obligations

‍

Wondering what potential AI could unlock in your company? Request an AI workshop now.

‍

3 Use a pre-trained open source LLM

You have neither the time nor the other resources to develop everything yourself? Then build on the work of experienced AI specialists. There is now a huge number of open source LLMs. On Huggingface.co alone more than 3,000 models are listed. Among the best known are Google's PaLM and Meta's LLaMA.

Do not let the big names dazzle you, though. Depending on the use case you have planned, it can be worth turning to a lesser known but specifically developed model. Well known or not: thanks to open source you can modify the LLMs freely for your purposes and train them with your own data from the start, instead of "only" buying AI software.

Part of the truth, however, is that the quality of many open source models still lags behind the commercial ones and that adapting them is more complicated than building with Lego. Your company will depend on experienced developers, whether internal or external.

Advantages:  

• Shortens the development process
• Fine-tuning the LLM on the basis of your own data
• Usually no licence costs
• Full control over infrastructure and architecture

Disadvantages:

• Requires your own ML team or cooperation with a technology partner
• Building your own infrastructure and maintenance processes
• Performance still lags behind the commercial LLMs, for now

4 Use a traditional ML model

It is fascinating that we can communicate with AI in natural language. With all the hype around LLMs, though, it is easy to forget that there are also value-creating AI use cases beyond natural language. Often value creation can be increased significantly simply by implementing traditional machine learning (ML) there.

One advantage of classic ML tools: you are not bringing a black box into the house, but can check exactly how results were arrived at at any time. To validate hypotheses in a mathematical context you do not even need machine learning. Here statistical inference is still clearly superior to large language models.

Advantages:

• Better suited to mathematical calculations
• In some cases cheaper to develop

Disadvantages:

• No communication in natural language
• Possibly lower user acceptance at first

Conclusion

In the end your decision is not just make or buy. You have a range of possibilities for putting artificial intelligence to work for your company, with and without LLMs. Convenience and speed argue for buying AI software or commercial LLMs. Developing your own, on the other hand, lets you cover very individual use cases and reach the last 20 percent of performance advantage that standard tools do not offer.

Despite the high AI pace you may be sensing in your industry: prepare your decision thoroughly and strategically, and start with an MVP for one selected first use case. That way you set the course for your company's future success.

If you are looking for a sparring partner for your decision-making or would like to be inspired by AI use cases, do take a look at our AI services page.