Custom AI model: when it is worth building and what you own at the end
When we talk about your own AI, we mean a model that is trained on your company's data and belongs to you. The alternative is rented off-the-shelf software, which is built for many customers at once and is available without a development project on your part. On this page, you will find the criteria for choosing between the two options and the cases in which your own model is not the right choice. You will also find out what you receive from us when you have your own model developed.
Build or Buy in AI: Six Criteria for Your Own Model or Rented Software
Renting an AI application or having one developed as a custom AI solution is a build-or-buy decision. You can only make that decision for a specific use case. The table lists six criteria and, for each one, the circumstances that favor one option or the other.
| Criterion | In favor of rented off-the-shelf software | In favor of your own model |
|---|---|---|
| Task | The task arises in the same way at many companies, for example translating texts or extracting data from invoices and receipts. | The task depends on patterns found only in your data, such as when your customers order or how they respond to the season. |
| Data | Your company has no historical data on the question that could be used to train a model. | The value the model is supposed to predict was recorded at your company in the past. |
| Costs | You pay a recurring fee that often depends on the number of users or the volume of use. There is no development project. | You pay for development once. After that, you incur operating costs, mainly for the working time needed for monitoring, review, and approval. |
| Time to first use | The software can be used once the contract is signed and the setup is complete. | We need four to six weeks for a prototype on your data. How long the pilot and the integration into your systems take depends on your processes. |
| Operation | The provider is responsible for operation and further development. | The model has to be monitored and retrained when needed. Your team or a service provider takes on this task. |
| End of the contract | Use of the software ends when the contract is terminated. | The model we train on your data remains your property. You can keep using it after the project and retrain it with new data. |
Hybrids of the two options also exist, for example off-the-shelf software that can be adapted with your data. For that reason, the table is no substitute for a case-by-case assessment. Before every model decision, we check whether the volume and quality of your data are sufficient for your use case.
When a Custom AI Model Is Not the Right Choice
In the following four cases, the arguments against your own model outweigh those in its favor. If our check shows that an AI project is not worthwhile for you, we tell you so and explain why.
The task is a standard task
For tasks that arise in the same way at many companies, ready-made software is often available. If you mainly want to introduce such tools and train your employees to use them, software vendors and training providers are the better choice in our view. Our focus is on models that we develop from your own data.
The data is missing
Your own model is trained on the data your company has recorded so far. If the value it is supposed to predict was not recorded at all or only roughly, there is no basis for training. In that case, it makes more sense to improve data capture first.
The simple solution is enough
Every model has to be better than what you already do today, for example the current rule of thumb or last year's value. If a prototype misses the threshold agreed beforehand, we recommend ending the project at that point. Sometimes analyzing the data is enough to show that an existing metric is sufficient for the decision.
Nobody will look after the model
After the project, a model has to be monitored and retrained when needed. If nobody at your company is responsible for this task and you do not intend to hand it to a service provider either, rented software that the provider operates is the better choice.
How the Costs and Accuracy of a Custom AI Model Develop
Rented software and your own model generate costs in different ways. With rented software, the price often depends on the number of users or the volume of use, so costs rise when more employees use the software. With your own model, you pay for development once, and after that the costs come mainly from the working time needed for operation.
A model does not become more accurate with new data automatically. Without monitoring and retraining, the quality of the predictions tends to decline in operation because the data changes, for example through new customers or a replaced machine. A model can become more accurate when four prerequisites are met.
Actual values are reported back
What actually happened is recorded as an actual value and compared with the prediction. You can measure how well the model performs in operation only once you have the actual values.
Limits are set in advance
The limits for warning and retraining are set before operation begins. Monitoring compares the same metrics that the model previously passed in validation.
Retraining follows a fixed rule
The model is retrained when a limit is breached. For most projects at mid-sized companies, a partly automated workflow in which a person approves every switch to a new model is enough.
The new version has to be better than the old one
A newly trained model version goes into operation only if it performs better in testing than the current one.
When these four prerequisites are met, a model can learn from more recent data with every retraining. We cannot promise in advance how much more accurate the predictions will become as a result. In operation, however, you can measure that difference against the actual values.
Computing time is rarely what makes retraining expensive. The greater effort lies in the steps around the training, namely collecting the actual values, testing the new model against the old one, and approving and documenting the switch.
Which option is cheaper over several years depends on the number of users, the period of use, and the operating effort. For that reason, run the comparison with the figures from your own quotes.
For details on how monitoring and retraining work, see the page MLOps Explained: Deployment, Drift, Retraining.What You Own at the End of a Custom AI Model Project
The model we train on your data remains your property. The documentation also belongs to you.
With rented software, use ends when the contract is terminated. You can keep using your own model after the project and retrain it with new data.
You also receive the code for data preparation and training so that your team can develop the model further.
The finished model runs in a container that bundles it with its runtime environment. As a result, it can be operated unchanged in a different environment, for example with another provider or in your own data center.
You can find the four points for checking any offer on the page Digital Sovereignty in AI.What Production-Ready Means and What You Receive at Handover
We call a model production-ready when it is versioned, automatically tested, integrated into your business application, and continuously monitored in operation. A model that runs only in a notebook, that is, in a data scientist's development environment, does not yet meet these conditions.
We put a model into operation only if it outperforms a simple baseline solution and the limits for warning and retraining have been set beforehand. We treat the handover to your team, that is, the knowledge transfer, as a separate project step. In addition to the model, your team receives three deliverables that it needs for operation.
The model in operation
The finished model is integrated into your systems as a versioned container. On request, it runs entirely on-premises in your infrastructure.
The documentation
The documentation is yours and describes what the model does, which data it needs, which metrics it passed, and which assumptions apply. It also records which data version and which code version produced a given result.
The runbook for incidents
The runbook, a set of written instructions for incidents, sets out what your team does when monitoring raises a warning. It describes the checks to run, names the contacts, and specifies how the model is rolled back to the previous version and how the decision on new training is made.
The training session
The training session is intended for the people who work with the predictions. Afterward, they should be able to judge what the model can do and in which cases it does not deliver a reliable result.
The goal of the handover is for your team to run the model itself and be able to develop it further. Until then, we help run it. We agree with you case by case whether and for how long we stay involved beyond that point.
How Your Business Unit and the Technical Team Work Together and How We Measure the Result
An AI project needs the knowledge of your business unit just as much as the work on the data and the model. We measure the result in euros because a metric such as accuracy shows how well the model calculates but leaves open how your company benefits.
Results in euros
Before the project starts, we agree on how each model metric is converted into euros. For example, we measure the result by the cost of downtime hours, scrap, or tied-up inventory.
Comparison with the simple solution
The model is compared with whatever makes the decision at your company today, such as the experience of your planner or a rule in the ERP system. Together, we set in advance how much better it has to be and record that margin as the threshold.
Business Understanding: Translating Business Goals into Measurable Questions
What your business unit contributes
The business question and the success criteria come from the people who work with the data. They also know what a field in the inventory management system actually means. That is why the first workshop includes the people who will later use the prediction, and one person at your company is responsible for the project.
What we contribute
We develop the model and integrate it into your processes. We support operation until your team takes over. Throughout, we work with the systems you have and do not use a language model where a simpler method is enough.
Before the project starts, it should be settled who develops and integrates the model and who runs it after the project ends. In our projects, the roadmap sets out these responsibilities.
The page Data Science Lifecycle: the Six CRISP-DM Phases shows what knowledge each phase requires.Frequently Asked Questions About Building a Custom AI Model
When is it worth building a custom AI model?
Building your own AI model can be worthwhile when the task depends on patterns found only in your data and when the value the model is supposed to predict was recorded in the past. In addition, the model has to be better than the solution you work with today. A prototype on your data shows whether it meets that condition.
When is off-the-shelf software better than a custom AI solution?
Rented off-the-shelf software is the better choice when the task arises in the same way at many companies or when nobody would look after the model after the project. It also counts against your own model if the necessary data is missing or a simple rule is enough. For introducing ready-made AI tools, software vendors and training providers are the right partners in our view.
Who owns the AI model you train on our data?
The model we train on your data remains your property. The documentation also belongs to you. You can keep using the model after the project and retrain it with new data. Because it runs in a container, it can be operated unchanged in a different environment, for example in your own data center.
What do we need in order to have a custom AI model trained?
You need a concrete problem with noticeable costs, processes that already generate data on that problem, and a person at your company who is responsible for the project. What matters is that the value the model is supposed to predict was recorded in the past. You do not need your own data science team or a data center to get started.
Does a custom AI model get better with new data automatically?
No. Without monitoring and retraining, the quality of the predictions tends to decline in operation because the data changes. A model can become more accurate when actual values are recorded, the limits are set in advance, the model is retrained when a limit is breached, and a newly trained version goes into operation only if it performs better than the current one.
What does it cost to have a custom AI model built?
The intro call is free. You start with fixed-price packages, from the Data Quality Snapshot for 650 euros to the AI Readiness and Data Quality Assessment for 3,490 euros (net). We quote the implementation individually afterward because effort and benefit depend on your data. In addition to the development costs, there is the ongoing effort for monitoring and retraining.
Let us talk about your data
Free 30-minute initial call. Briefly describe your dataset and your plan; optionally you can state a preferred date.
These services do not replace legal data protection advice.