Skip to content

Start an AI pilot

One use case, success and stop criteria set upfront, and a small-scale implementation using real data. We help GVB choose which use case will have the greatest impact.

Through data strategy, we help GVB identify the use case with the greatest impact for the least relative effort

At GVB, a new data application involves several departments and teams that have often never worked together before. We support GVB in defining, selecting and delivering data use cases, and work together to shape its move towards data-driven working.

Read the case study

Does our AI idea really help in practice, and how can a pilot show that?

A demo with a language model or a model in a notebook can be convincing. Whether it improves your work only becomes clear with real data, real users and an agreement upfront on what counts as success and when to stop.

Arrange a conversation

After an AI pilot, you have a decision: proceed, adjust or stop, backed by your own data

Discuss your pilot

One clearly defined use case

One task or decision, one group of users, one metric. We help GVB identify which use case will have the greatest impact with the least effort.

Success and stop criteria agreed in advance

Before we start, we agree on what the pilot needs to demonstrate, what we will compare it against and which results would mean stopping.

Testing with your own tasks

For AFWC, we created AI exercises from tasks participants brought from their own work. We use the same approach when pilot users test.

A decision and a next step

At the end, you have a go or no-go decision. If it’s a go, you know what production requires: data feeds, integrations, management and an owner.

We define the criteria and data check before building with real data

  1. Use case and criteria

    What is the task, who are the users, what counts as success, and what outcome would make us stop?

  2. Data check and small-scale build

    We check whether the data is available and accurate, then build the smallest version that answers the question.

  3. Measure and decide

    Users work with it in their day-to-day work. We measure the results against the criteria, and you decide whether to go ahead or stop.

Ready for a pilot, or does our idea need something else first?

Talk it through with us

A good fit if

You have one concrete idea, someone who owns the process and users willing to help test it. The data exists, even if it is not perfect yet.

Not a good fit if

There are ten ideas and no one is choosing between them, or the data for the use case does not exist. Start by choosing a use case or looking at the data source.

What determines the scope

How many sources the pilot needs, whether it uses your own model or a language model, how many users take part, the privacy requirements and how close to production you want to test.

What if the pilot fails?

A decision not to proceed, based on what you agreed in advance, is also an outcome. You then know why it does not work and what it would take to make it work. That is better than a pilot that runs on indefinitely.

A proof of concept, a pilot or straight to production: which step makes sense now?

A proof of concept

Choose this if you don’t yet know whether it’s technically feasible, for example, whether a model can extract the right information from your documents. Test this on a fixed dataset.

A pilot

Choose this if it’s technically feasible and you want to know whether it helps in practice. Our advice: choose one use case and define the criteria for stopping before you begin.

Straight to production

Choose this if the application is well established and its value is already clear. You can skip the pilot and read more about AI implementation.

We set up a pilot from the outset so it can scale into production

Data Science Lab has worked with data and AI since 2016 and is a Databricks and Microsoft partner. We also test new AI tools in our own team through pilots, such as GitHub Copilot.

Book an introductory meeting

The first conversation starts with one idea and the colleagues who will work with it

Stef Stoelwinder

Stef Stoelwinder

Business Development

Stef is your first point of contact. In our first conversation, we’ll discuss your idea, the data and what the pilot needs to demonstrate.

Which AI idea would you like to test in a pilot?

Tell us what idea you want to test, what data you would use and who would be involved. We’ll reply by email or arrange an initial conversation.

Enter a valid email address.

Let us know how we can help you.

Frequently asked questions

A small, clearly defined trial in which an AI application uses real data with real users to show whether it helps them in their work. You decide in advance what counts as success and when to stop. The outcome is a decision, not a demo.

A pilot is one step in our work. You can find the steps before and after it under services.

An AI proof of concept, or AI PoC, answers whether something is technically possible, usually using a fixed dataset. A pilot answers whether it helps: with real data, in the work of a small group of users.

If you already know it is technically possible, you can skip the PoC and start a pilot straight away.

You define how you will tell whether it works compared with the current way of working: less time spent searching, fewer manual checks, or better suggestions according to users. You also set the threshold below which we stop, and what users must be able to do without help.

We also record what the pilot does not test, so no one interprets the decision differently later.

As small as possible, as realistic as necessary. One use case, the data sources it needs and a group of users who use it in their own work. Better to test a narrow task fully than a broad task only partly.

The data check can be broader: it is better to find problems in the source data during the pilot than in production.

Usually not because of the technology, but because of what surrounds it. Three things need to be in place: a reliable data supply, a team that can maintain the model and a process in which the output is actually used. If one is missing, the pilot remains a demo.

Read more in from pilot to production.

Yes. At Data Science Lab, for example, we started a GitHub Copilot pilot involving several colleagues and different data science projects. Before we began, we reviewed the privacy terms.

Read how we set up that pilot in DSL Pilot: GitHub Copilot.

The next step is production: running the application on live data, connecting it to your systems, with monitoring and someone on your side responsible for it. The pilot decision already sets out what that requires.

We describe this step under AI implementation, and what comes after under management and continuous improvement.

When no one will use the outcome, when the data for the use case does not exist, or when a fixed rule in your system already handles the task well. Without someone to make a decision after the pilot, it also remains just a trial.

If you are not yet sure which idea to pursue first, start with AI consultancy.