Civil servants already work with AI. Now it is the office’s turn
Civil servants did not wait for reform. They turned to artificial intelligence on their own because they want to work faster and better. That is valuable capital: people are ready, they know what AI can do, and they already have their first good practices. Now the office can give them something better than a private chatbot account.
This is the last post in the series AI in public administration: the last mile. In part 1 we wrote about Polish language models, and in part 2 about digital sovereignty. Today it is about people: how to move AI in public administration from the grey zone to the official desk.
Tomasz and forty pages of documentation
Tomasz (a fictional but typical character) has worked in a county architecture department for fifteen years. Applications keep growing, staff numbers do not. A year ago a colleague showed him how to summarise forty pages of documentation in a few minutes. Since then Tomasz has done it every day, on his private phone, in the free version, because there is no official tool yet.
He does not hide it, but he does not talk about it either. He does not know whether it is allowed. This is not a story about breaking rules. It is a story about people who got ahead of their institutions, and about institutions that can now quickly catch up.
Shadow AI in numbers
Surveys of employees and surveys of institutions show two different pictures:
| Who asked | Whom and when | Findings |
|---|---|---|
| NASK | 6,093 public sector employees, October-November 2025 (published March 2026); pilot study | 46% of respondents use generative AI at work |
| Association of Polish Cities and Fundacja Miasto | Civil servants and cities, published 1 October 2026 | Half of city employees use AI; 8% of cities have an AI strategy |
| Citizens Network Watchdog Poland | Questions to 5,480 institutions, summer 2025; 74% responded | 9% of institutions report using AI; 20 of the 2,027 responding municipalities have rules for publicly available AI tools |
These numbers cannot simply be subtracted from one another, because they concern different people, questions and dates. They do show, however, that AI in local government and central administration is already part of everyday work, even if official declarations do not reflect it yet. Read more about this in our Shadow AI guide.
How a workaround becomes the norm
No one decided to “use AI quietly”. Simplified, the process has four steps:
- Workaround. One civil servant finds a way to get the work done on time. The tool is at hand, free and it works.
- Habituation. Nothing bad happens, so the workaround stops being an exception. Sociologist Diane Vaughan called a similar phenomenon the normalisation of deviance.
- Imitation. Others see that it works and do the same.
- Dependence. The office starts to rely on the pace AI provides, even though it officially does not use it.
In Polish public offices, three things encourage this: civil servants have not yet received a tool approved for official data, offices have not received money for one, and responsibility rests on the individual employee. The Head of the Civil Service’s guidelines rightly point out that AI does not release a person from responsibility. That principle works best when the civil servant has an official tool at hand.
What everyone gains
Civil servant: works faster, openly and safely. Tomasz does not need convincing about AI. He needs permission and a tool. His colleagues think the same: in the NASK survey, 87% of respondents want training, 73% expect a clear legal framework and 70% want internal guidelines. When AI comes out of the grey zone, good practices move from one phone to the whole department.
Citizen: a decision based on real law. Joanna (a fictional character) receives a refusal citing a provision she cannot find, because it does not exist. Watchdog Poland described three municipalities that issued decisions with non-existent legal bases, most likely generated by a chatbot. A tool that shows the source of every answer lets the civil servant check the provision before signing the decision. That matches what civil servants themselves expect: 91% of NASK respondents want a person, not an algorithm, to decide citizens’ cases.
Office: sees, protects and can prove it. When residents’ data goes to foreign services on private accounts, the office, as the data controller, is responsible for GDPR compliance. Poland’s Act of 3 July 2026 on artificial intelligence systems has applied since 11 August and establishes the Commission for the Development and Security of Artificial Intelligence. From 28 October its provisions on inspections, proceedings and penalties will apply, and the Ministry of the Interior and Administration is already reviewing how offices use AI. An office with a usage log will answer the question “how does AI work here?” calmly and concretely. What the act means for public offices is described in our guide Act on artificial intelligence systems.
A tool works better than a ban
The first reflex is a ban. A ban without an alternative, however, pushes AI deeper into the shadows: in a Software AG survey, 46% of employees said they would not give up their own AI tools even if their employer banned them. The countries that have done best gave civil servants a tool they may use with real data:
- Singapore gave civil servants the Pair assistant, approved for restricted and sensitive data. In the first two months it was used by over 11,000 people in more than 100 agencies.
- The US House of Representatives banned staff from using Copilot in 2024, and in September 2025 began making up to 6,000 licences available to them, with heightened legal and data protections. From ban to tool in less than two years.
- France opted for a state assistant for civil servants. In summer 2026 the number of users grew from 10,000 to 43,000, and local government employees reach for ChatGPT less often.
Corporate data says the same. When employers buy official licences, the share of employees using AI on private accounts falls, from 72% to 47% within a year in Netskope data. So ChatGPT in public offices does not have to be a problem to fight. It is enough to offer a better, legitimate alternative.
Poland has an advantage here that no one on this list had at the start: its own language models that already work, as we described in part 1. Others are still waiting: in South Korea, the national model project was described in September as “adrift”.
Four steps out of the grey zone
Leaving the grey zone does not require a revolution. It requires four things:
- Clear rules. An AI use policy, a register of systems and responsibility assigned to roles. Our AI use policy template is a good place to start.
- Polish models. Bielik and PLLuM running locally, so that residents’ data never leaves the office.
- A tool with control. Every query passes through a controlled gateway, described in more detail in part 2. Answers cite a specific document: a resolution, a procedure, a BIP page. And when the documents do not contain an answer, the system says so plainly.
- Methodology. Compliance checked before deployment, not bolted on after the first inspection.
We deliver these four steps in SAIE: TWIN:DESK is the AI digital desk, PROXY:AI is the control gateway with a usage log, and our deployment methodology starts with compliance. For Tomasz, this means he still summarises documentation in minutes, only now on his work computer, openly and with his manager’s knowledge.
See you at the 2nd Forum on AI in Public Administration
We close the series with an invitation. At the 2nd Forum on AI in Public Administration in Wrocław (4-6 November 2026) we will discuss how to bring AI out of the grey zone in municipalities, counties and central offices. We are a Special Partner of the Forum, and on 5 November, in our talk “SAIE in a public office”, we will show how to give civil servants an assistant with access control, usage logs and proof of compliance. See Forum details.
AI training for civil servants
A tool without training will not work, and training without a tool ends up on a private phone. The AI Act (Article 4) requires public bodies that use AI to support their staff’s AI literacy. Our AI training for civil servants was designed as a practical response to the NASK report and the Ministry of Digital Affairs guide, in other words to what civil servants themselves expect. See our training offer.
Summary
Civil servants have already shown that they want to work with AI. Joanna can receive a decision based on verified law, and the head of office can answer the question “how does AI work here?” with a log. Now it is the offices’ turn.