AI in government processes: maintaining control while experimenting responsibly

Appeals and Objections Processes Under Pressure? This Is Where AI Can Make a Difference

Appeals and objections are precisely the processes where AI can take a significant amount of work off people's hands. These workflows typically involve large volumes of cases, extensive textual communication, and repetitive tasks. At the same time, every file demands meticulous care and proper context.

In the podcast, Deloitte describes how large public sector organizations sometimes process hundreds of thousands of objections. Manual assessment is highly time-consuming and puts immense pressure on available capacity. AI can assist by automatically identifying grounds for objection, finding similar cases, and labeling or clustering issues.

"You're easily talking about hundreds of thousands of objections. If you have to process all of those manually, you'll be busy for quite a while." — Rutger Rienks, Deloitte

There are also significant opportunities in communication—for example, when generating draft letters or checking whether an objection is complete before the formal process even begins. This not only accelerates processing times but also helps make processes more consistent.

At the same time, AI introduces new challenges. Citizens themselves are now using AI to draft objections. The podcast describes how ChatGPT effortlessly generates professional, legally-sounding objection letters that look convincing but are by no means always factually or legally correct. Nevertheless, these objections still have to be formally processed.

This means AI is becoming both part of the solution and part of the problem. That is precisely why designing smarter processes is more important than ever.

Safe and Sovereign: How Government Maintains Control Over AI Safely and Securely

Many organizations are now experimenting with generative AI tools like Copilot or Gemini. However, public consumer AI solutions are not automatically suitable for sensitive government processes.

The discussion here revolves around more than just privacy; it is primarily about control and sovereignty. Where does the model run? Which data is being used? Who can look behind the scenes? And what happens if a foreign entity decides to restrict access?

"You don't want it to end up in a black box, or have a US government decide to temporarily shut down this solution." — Dion Bonné, Nebul

The podcast highlights that this is precisely where governments' greatest concerns lie—especially when processes involve sensitive personal data or are directly tied to legislation and regulations.

At the same time, working fully on-premises is simply not realistic for many organizations. Modern AI requires massive computing power and specialized infrastructure. This makes the decision highly complex: becoming completely dependent on Big Tech feels undesirable, yet building everything yourself is often not scalable.

According to Nebul, the solution lies in a sovereign ecosystem: open-source models running on Dutch infrastructure, combined with specialized AI that can be specifically trained on Dutch laws and regulations.

That last part is crucial. A generic language model mostly possesses general knowledge. A government, however, needs AI to understand how Dutch legislation works and which rules take priority within a specific context. By fine-tuning models on these specifics, AI becomes not only more relevant but also far more reliable.

Deploying AI Alone Is Not Enough for Reliable Decision-Making

Yet, the most important takeaway might lie elsewhere. According to Blueriq, the greatest risk arises when organizations assume AI can take over the entire process.

A language model does not deliver fixed, predictable outcomes the way a rule-driven system does. It predicts words and patterns, but it does not operate on explicitly defined legal rules. This makes AI highly powerful for communication, interpretation, and support, but less suitable for hard, binding decision-making.

Therefore, the podcast draws a clear distinction between generative AI and rule-driven process logic. For legal rules and transparent decision-making, fixed business rules remain essential.

You don't need to leave a simple legal rule like "someone must be eighteen years or older" to a language model. You want to explicitly capture that kind of logic in rule engines and decision logic. AI then supports the knowledge worker around those guardrails.

It is precisely this combination that keeps processes manageable. AI assists with analysis, communication, and summarization, while rule-driven technology guards the boundaries, process steps, and compliance.

"Don't view it as something that will take over everything, but see it as an addition to your IT toolkit that can support you incredibly well." — Yuri Versluis, Blueriq

AI in Appeals and Objections Sounds Good. But Does It Work in Practice?

"By applying AI to objection procedures, you can reduce processing times by 70 to 80 percent." — Rutger Rienks, Deloitte

The collaboration between Blueriq, Deloitte, and Nebul was not limited to theory. Together, the three parties developed a proof of concept around appeals and objections, enriching an existing process with AI and running it on Dutch sovereign infrastructure.

According to the participants, this delivered remarkably fast results. Deloitte notes reductions of 70 to 80 percent in processing time for specific parts of objection procedures. Just as important is what this time savings yields: more room to focus on cases that truly require human judgment.

This also shows what the discussion surrounding AI in government should actually be about: how to organize processes smarter in the face of ever-increasing workloads, an aging workforce, and growing complexity.

Experimenting with AI Feels Exciting. Yet, This Is the Time to Start

The technology, the infrastructure, and the first working examples are already here. We can see that AI works. For governments, the primary question is how to get started safely and manageably. According to the podcast participants, this simply requires the courage to experiment, combined with the right expertise and clear boundaries.

The technology is available, the first results are proven, and European regulations around AI usage are becoming increasingly concrete. This makes now the perfect time to learn, run pilots, and build experience—before the pressure to change outweighs the room to do so in a controlled manner. To do this, you start small and controlled, utilizing a combination of sovereign infrastructure, specialized models, and rule-driven decision-making.

After all, AI within government does not have to be a black box, but that is only possible if you maintain control over the processes surrounding it.