Machine Learning for SMEs: When to Build, When to Buy?

Key Takeaways: The choice between building or buying machine learning is not a binary decision but a spectrum with five options, from standard SaaS to fully custom. For most SMEs, the hybrid approach is most cost-effective: buy the infrastructure, build the models. The WBSO covers up to 40% of development costs for custom builds, and the MIT scheme funds feasibility studies to substantiate the right choice.

The build-versus-buy spectrum

The question "should we build machine learning ourselves or buy it?" is a false dilemma. In reality there is a spectrum of five options, each with its own balance of cost, flexibility and complexity. At one end sits standard SaaS software with built-in AI functionality. At the other end sits a fully custom machine learning platform, built on proprietary infrastructure with proprietary models.

In between lie three intermediate forms. The first is SaaS with configurable AI, where you use a platform that offers machine learning as part of the product but allows you to train models on your own data. Think of a CRM system that performs lead scoring based on your historical sales data. The second is the use of AI APIs such as the OpenAI API, Google Cloud AI or AWS SageMaker, where you integrate existing models into your own applications via a programming interface. The third is a hybrid approach where you use cloud infrastructure and frameworks but develop and train models yourself on your own data.

According to McKinsey, 70% of companies adopting AI spend less than 10% of their digital budget on AI-specific technology. That means most businesses work with modest budgets and must therefore make sharp choices. The right position on the spectrum depends on three factors: the uniqueness of your use case, the importance for your competitive position and the availability of your data. This article helps you make that assessment.

When to buy: standard use cases with off-the-shelf solutions

Buying is the right choice when your problem is not unique. There are dozens of well-functioning SaaS solutions for predictions that virtually every business needs. Churn prevention in e-commerce, lead scoring in B2B sales, demand forecasting in retail, fraud detection in payments and sentiment analysis on customer service tickets are all problems already solved by thousands of companies before you.

The economics are clear. A SaaS tool for lead scoring costs 200 to 2,000 euros per month, depending on the number of users and data volume. You are operational the same day. No development team is needed, no data infrastructure and no maintenance. The vendor handles updates, improvements and compliance. For an SME with a budget of less than 1,000 euros per month for AI applications, this is almost always the wisest choice.

The critical criterion is whether the AI functionality needs to be a source of competitive advantage. If the answer is no, if it concerns a commodity function that all your competitors deploy in the same way, then buying is the right choice. A webshop doing product recommendations with a standard SaaS tool performs just as well as one with a custom recommendation algorithm in 95% of cases. The remaining 5% is only relevant for companies where personalisation is the core product, think Spotify or Netflix.

Do watch the total cost over three to five years though. A SaaS tool at 500 euros per month is 18,000 euros over three years and 30,000 euros over five years. At that point the comparison with custom development becomes more interesting, especially if your needs have grown in the meantime.

When to build: unique data, competitive moat, specific requirements

Building is the right choice when your data or your problem is fundamentally different from what standard solutions address. This applies in three situations. The first is when your competitive advantage stems from the uniqueness of your data. If you possess datasets your competitors do not have, historical process data from twenty years of production, sensor data from patented machines, or customer behaviour data from a unique sales channel, then a standard model trained on generic data is by definition inferior to a model trained on your specific data.

The second situation is when machine learning needs to build a strategic moat around your business. If the prediction model itself is the product or service, or if model quality directly determines your competitive position, then you do not want to hand control to a SaaS vendor. A logistics company that considers route optimisation a core competency builds its own models. An insurer that sees risk assessment as a differentiating factor trains its own models on its own claims history.

The third situation is when you have specific technical requirements that no standard solution fulfils. Integration with legacy systems that do not support standard APIs. Processing of data formats that deviate from what common tools can handle. Real-time predictions with latency requirements under 50 milliseconds. Compliance requirements mandating that data does not leave your own data centre. In each of these cases, custom work is unavoidable.

Costs for custom machine learning development vary considerably. A relatively simple prediction model, think demand forecasting for a retailer with structured sales data, costs 20,000 to 35,000 euros for development and initial deployment. A more complex system with multiple models, real-time scoring and integration with existing systems comes to 40,000 to 60,000 euros. The timeline is three to six months for a first production-ready version. Annual maintenance costs of 15-25% of initial development costs should be added on top.

The hybrid approach: buy infrastructure, build models

For most SMEs that decide standard SaaS is insufficient, the hybrid approach is most cost-effective. You buy the infrastructure and build the models. Concretely this means using cloud platforms like AWS, Google Cloud or Azure for compute and storage, using open-source frameworks like scikit-learn, PyTorch or TensorFlow for model development, and training, validating and deploying the models yourself (or with your partner) based on your own data.

This approach combines the advantages of both worlds. You do not need to manage data centres, purchase GPU clusters or develop basic infrastructure. At the same time you have full control over your models, your data and your intellectual property. Infrastructure costs are variable and scale with your usage. A typical cloud bill for an SME machine learning project runs 100 to 500 euros per month, depending on training frequency and data volume.

The open-source ML community has drastically lowered the barrier to entry. Hugging Face offers more than 500,000 pre-trained models you can use as a starting point. Transfer learning allows you to fine-tune a model trained on millions of data points to your specific dataset of just a few thousand records. What five years ago required a team of five data scientists and a budget of half a million euros is today achievable with one experienced ML engineer and a budget of 30,000 to 50,000 euros.

Cost comparison: SaaS versus custom over 3 years

Let us put the numbers side by side for a concrete scenario: a wholesaler with 50 employees wants to deploy demand forecasting to improve inventory optimisation.

The SaaS route. A specialised demand forecasting SaaS costs 800 to 1,500 euros per month. Over three years that is 28,800 to 54,000 euros. Implementation costs of 5,000 to 10,000 euros for configuration and integration are added. Total over three years: 33,800 to 64,000 euros. The solution is operational within four to eight weeks.

The custom route. Development of a demand forecast model based on proprietary sales history, seasonal patterns and external variables costs 25,000 to 40,000 euros. Cloud infrastructure runs 200 to 400 euros per month, or 7,200 to 14,400 euros over three years. Maintenance and continued development costs 5,000 to 8,000 euros per year, or 15,000 to 24,000 euros over three years. Total over three years: 47,200 to 78,400 euros. The first version is operational after three to four months.

At first glance SaaS appears cheaper. But two factors shift the balance. First, a custom model typically performs 10-30% better than a generic SaaS model because it is trained on your specific data and patterns. For a wholesaler with 5 million euros in revenue, a 2-3% improvement in forecast accuracy can lead to 50,000 to 100,000 euros in inventory cost savings per year. Second, with custom development you own an asset after three years, while with SaaS you are back to zero when you cancel the subscription.

Timeline: from idea to production

Timeline is an underestimated factor for many SMEs. SaaS implementations are fast. Four to eight weeks for configuration, data integration and user training. You see results quickly and can adjust course rapidly if the tool does not meet expectations.

Custom development requires more patience and a clearer plan. A realistic timeline looks as follows. Weeks one to two: data inventory and feasibility analysis. Weeks three to six: data engineering, cleaning and preparing data for model training. Weeks seven to ten: model development, training and validation. Weeks eleven to fourteen: integration with existing systems and user acceptance testing. Weeks fifteen to sixteen: deployment and monitoring. Total: four months for a first production version.

After this initial version the iterative improvement process begins. A machine learning model improves as it processes more data and as you feed practical feedback back into the model. After six to twelve months a well-maintained custom model typically performs significantly better than its first version. This continuous improvement process is inherent to machine learning and one of the reasons why custom models can be more valuable in the long term than static SaaS solutions.

Data ownership and vendor lock-in risks

A factor many businesses consider too late is data ownership. When you use a SaaS platform for machine learning, that platform processes your data. In many cases the vendor uses anonymised versions of your data to improve its models, which means your competitors indirectly benefit from your data. Read the terms of service carefully: 42% of SaaS vendors retain certain rights to data processed through the platform according to a Gartner analysis.

Vendor lock-in is a real risk. The longer you use a platform, the more your processes, integrations and workflows depend on that specific vendor. Switching to a competitor or to a custom solution becomes more expensive as time progresses. Ask yourself: what happens if the vendor doubles the price, if the vendor is acquired by a competitor, or if the vendor discontinues the product?

With custom development, ownership is clear. The models, the code and the training data are yours. You can switch cloud providers without losing your models. You can transfer development to another partner without starting from scratch. This ownership is not just a legal matter but a strategic one. Your machine learning models, trained on your unique data, become a business asset that increases in value over time.

To minimise lock-in, even with a SaaS choice, follow these three principles. Demand data portability: you must be able to fully export your data at any time in a standard format. Avoid proprietary integrations: choose platforms that work with open API standards. Always keep your raw data under your own control: the SaaS tool may process your data, but the source data stays in your own systems.

WBSO for custom machine learning development

Developing custom machine learning models is precisely the type of work the WBSO was designed for. The WBSO compensates part of the salary costs for technical-scientific research and technical development. Building a company-specific prediction model trained on your data, validated against your business context and integrated into your systems qualifies as technical development under the WBSO.

Concretely, the WBSO provides a payroll tax reduction of 40% on the first 350,000 euros in R&D salary costs and 16% above that threshold. For a custom ML project of 40,000 euros in development costs, this means an effective saving of 10,000 to 16,000 euros, depending on the exact split between salary and material costs. That saving makes the cost difference between SaaS and custom considerably smaller.

The WBSO application must be submitted before you begin development work. This requires a project plan describing the technical uncertainties you will resolve and the development activities you will perform. The application process is well documented and approval rates are high: more than 90% of applications are approved. More information can be found on our WBSO page.

MIT scheme for feasibility studies

Still uncertain whether building or buying is the right choice for your specific situation? The MIT scheme (SME Innovation Stimulus for Regional and Top Sectors) offers subsidies for feasibility studies that answer exactly this question. A feasibility study for deploying machine learning in your business can be subsidised up to 35% of costs, with a maximum of 20,000 euros.

Such a feasibility study examines what data is available and of what quality, which machine learning techniques are suitable for your specific problem, what the expected return on investment is and whether building, buying or a hybrid approach is most suitable. The result is a substantiated recommendation on which you can base a well-considered investment decision.

The combination of MIT for the feasibility phase and WBSO for the development phase offers an attractive financing model. You minimise risk by first investigating whether the project is feasible with subsidy support, and then maximise the subsidy on actual development. For businesses considering integrating AI as a core competency, the AI subsidy page is a good starting point for an overview of all available schemes.

A decision framework for your situation

The right choice depends on your specific circumstances. Use the following three questions as a decision framework.

Question one: is your use case standard or unique? If your problem is comparable to what thousands of other businesses have, such as churn prevention, lead scoring or email optimisation, then buying is almost always the best choice. If your problem is specific to your industry, your data or your business process, the balance shifts towards building.

Question two: is the AI functionality a source of competitive advantage? If the answer is yes, invest in ownership. A vendor selling the same AI functionality to your competitors delivers no competitive advantage by definition. If the answer is no, buy the functionality as a commodity and allocate your development budget to things that are differentiating.

Question three: do you have unique data of sufficient volume and quality? Custom models are only superior to generic models if they are trained on data that the generic models do not have. If your data is not unique or if the volume is too small for reliable model training, typically fewer than 5,000 to 10,000 relevant data points, then a generic SaaS model likely performs better than a custom model.

Three times "standard, no, not unique" means buy. Three times "unique, yes, sufficient" means build. Everything in between points to the hybrid approach. And when in doubt: start with a subsidised feasibility study via the MIT scheme. Better to invest two months in a well-substantiated decision than two years in the wrong direction.