What AI does here, and what it does not.

01What this covers

The AI Lab designs and builds AI automation, AI agents, AI chat and voice interfaces, workflows, integrations, dashboards and the business systems around them, and advises on where those are worth building at all. This policy explains how AI is used in that work.

It applies to systems we deliver to clients and to the tools we use internally while building them. Anything specific to a single engagement is agreed in that engagement's own contract, which takes precedence over this page.

02Where AI is used, and where it is not

AI is a good tool for a specific class of problem: understanding messy language, and producing a draft that a person can check quickly. That is where we reach for it.

• Reading and classifying incoming text, such as sorting enquiries by intent or urgency.
• Extracting structured fields from unstructured documents, so a person is not retyping an invoice.
• Summarising long material into something a human can act on.
• Drafting replies, descriptions and documents that a person reviews before they go out.
• Routing work to the right person or system based on what the content actually says.
• Conversational and voice interfaces that answer common questions and hand over cleanly when they cannot.
• Assisting us while we write and review code during a build.

Where we deliberately do not use it

A language model is the wrong instrument whenever the answer must be identical every time, provably correct, or defensible line by line. Arithmetic, pricing, tax, thresholds, eligibility rules and anything with a legal consequence belong in deterministic code that we can test.

If a normal automation, a rule or a plain integration solves your problem better than AI, we will say so, even when that makes the project smaller.

03Human oversight

Every AI step we deliver has a named human who is accountable for its output. Automation moves the work, it does not move the responsibility.

We design so that the strength of the check matches the size of the consequence. Something that only saves reading time can run unattended and be sampled. Anything that reaches a customer, moves money, changes a record or affects a person's rights is reviewed by someone before it takes effect.

• Every AI step can be inspected: you can see what went in and what came out.
• Every AI step can be overridden and corrected by a person without waiting for us.
• Every AI step can be switched off, and the process keeps working without it.
• Low confidence, unusual input or a failed check escalates to a human rather than guessing.

We will not deliver a system whose output nobody is able to check. If a process cannot be supervised in a realistic way, that is a reason to build it differently, not a reason to trust it more.

04The limits of AI output

Modern language models are genuinely useful and genuinely unreliable in specific, predictable ways. Knowing those ways is the difference between a system that helps and one that quietly causes damage.

• Output is probabilistic. The same question can produce a different answer on a different day.
• A model can be confidently wrong. Fluent, well-structured text is not evidence that the content is true.
• Models invent details that look plausible, including references, figures and names that do not exist.
• Small changes in how something is phrased can change the answer more than they should.
• Models reflect biases present in their training data, which can show up as uneven treatment of people or cases.
• A model knows nothing about events after its training cut-off, and nothing about your business unless the system explicitly supplies it.
• Providers update models. Behaviour can shift without anything in your system changing.

05Precision where it matters

The way to build a trustworthy AI system is to give the model only the part of the job it is actually good at, and to hand the rest to code that behaves the same way every time.

Our own product CalcAI is the clearest example we can point at. A language model reads a job description written in ordinary language and works out what is being asked. A hand-built calculation engine then does every sum, deterministically, the same answer every time. The AI never gets to invent a price. That split is the pattern, not the exception.

06Your data and your responsibility

A system we build processes the material you put into it. That material stays yours, and so does the responsibility for it being lawful to use.

• You are responsible for having the right to use the data you supply, including any personal data in it.
• You are responsible for the lawful basis for the processing, and for informing the people whose data it is.
• You are responsible for the accuracy of what you supply. A system fed wrong information produces wrong output faster.
• Special categories of data, such as health, biometric or criminal record information, may only enter a system where the agreement provides for it and the safeguards are in place.
• Do not put credentials, API keys or passwords into prompts, chat fields or uploads. Nothing that reads text should ever be handed a secret.

The output of an AI step is a draft until a person or a deterministic check has accepted it. Where you choose to reduce the checking, you are choosing to accept the risk that comes with that, and we will tell you so plainly before it is configured.

07Third-party AI providers

This website uses no AI provider at all, so there is nothing to disclose about it here. That is the honest answer, and it will change on this page first if it ever stops being true.

In client work, model providers are chosen per project together with the client, because the right choice depends on the task, the sensitivity of the data and where it may lawfully be processed. We do not keep a single default provider that we quietly apply to everyone.

• Whichever providers are used are named in the engagement's processing agreement, before anything is connected.
• The processing region is part of that choice, and we prefer European processing where the project allows it.
• Where a provider offers a setting that excludes submitted data from their own model training, we select it.
• Where a provider does not offer that, we tell you before it is used, so the decision is yours.

08Confidentiality

Client material is confidential and stays inside the tools and accounts agreed for that engagement. We do not paste client data into consumer AI products, personal accounts or free tiers whose terms allow the provider to train on the input.

Where we use an AI assistant while writing code, it works against the project's own codebase under the terms agreed for that project, and client secrets and production data stay out of it.

09Being honest with your users

When we build a chat or voice interface, we build it to say that it is an automated system rather than to imply a person is typing. We recommend the same in the copy around it, and we will push back on a design intended to make users believe they are talking to a human.

Every conversational system we deliver has a route to a human. An automated assistant that cannot help and cannot hand over is worse than no assistant at all.

Where AI materially shapes something a person receives, we advise saying so. Users are more forgiving of a system that admits what it is than of one they catch out.

10Regulation

AI systems in the European Union sit under the AI Act as well as the General Data Protection Regulation, and the obligations depend on what the system does rather than on the technology inside it. A tool that drafts internal summaries is not treated like a tool that screens job applicants.

We assess where a proposed system is likely to land before we build it, and we agree in writing who carries which obligation. Where a use case looks like it belongs in a high-risk category, we say so early, because that changes the budget and the timeline rather than a checkbox at the end.

11What we will not build

Some requests we turn down, and it is fairer to say which before you spend a meeting on them.

• Systems designed to deceive people about whether they are dealing with a machine.
• Systems that impersonate a real, identifiable person without their clear permission, including voice clones.
• Systems that score or rank people in ways that decide their access to work, housing, credit or services with no meaningful human route.
• Systems built to collect personal data from sources where doing so is unlawful or against the source's terms.
• Systems whose purpose is manipulation, harassment or the mass production of misleading content.

12Telling us something is wrong

If an AI system we built produced something incorrect, unfair or harmful, we want to hear about it quickly. Write to hello@thecreativitylab.nl with what happened and roughly when, and a person will look at the actual records rather than send you a form.

Where the system runs for one of our clients, they own the process and we support them in correcting it. Where the problem is in something we built, fixing it is our job.