Public sector
Public Administration
The problem
The assistant writes a polite letter to a citizen that promises the outcome of the decision, mentions their complaint on social media or cites articles nobody verified. The letter is registered, may be published, and can be requested by a journalist or an auditor.
30 courses · 5 specific to this profession
A process from your work
How much of your day goes on finding information and writing emails?
- Step 1The tool finds the relevant documents and drafts the email or the letter.
- Step 2You check every fact at its source before using it, and rewrite the text for its reader.
- Step 3The letter goes out under your name, and the responsible person takes the decision.
- Where the tool goes wrongThe facts, names, dates or references the tool finds may not match their source document.
- The verification stepFor every fact, name, date and reference, open the document it comes from and compare.
More: what was measured and the full check
What changed, as measured
Compared with a group of non-users, staff with the assistant saved an estimated 19 minutes a day across routine tasks; the biggest savings were in searching for information and writing emails, 26 and 25 minutes.
Who was measured: Staff of a UK government department who received an office AI assistant, compared by survey with colleagues who did not.
The caveat: Times are self-reported, and licences went to volunteers first, not at random, so early adopters may overstate the benefit.
The check, step by step
- For every fact, name, date and reference the tool found, open the document it comes from and compare.
- Do not promise any outcome for a request before the decision is taken.
- Before you ask the tool, remove any personal data the task does not need.
- Keep the version you sent and the sources you checked on file.
The course covers drafting, editing and checking text with AI, with no invented facts and no personal data in public tools.
We can help you with this process: “Public administration AI: Documents and communication”, 8 hours.
A documented failure pattern
Undelivered leaflets, filed as “Discrimination between neighbourhoods”
- What the tool doesA thematic summary of comments from a municipal public consultation, produced by an AI tool for the civil servant who prepares the summary of recommendations that accompanies the draft act.
- Where the tool goes wrongThe tool turns a complaint about leaflets into “discrimination”, an accusation nobody made.
- The verification stepFor each theme with a legal term, open the original comments and check the citizen's words.
More: why it is easy to miss and the full check
Where the tool goes wrong, and why it is easy to miss
The citizen reports a notification gap: the leaflets did not reach their neighbourhood. The tool has lifted this into a legal category. Albania's anti-discrimination law (Law 10221/2010) defines discrimination as a distinction based on a ground listed in its Article 1, and place of residence is one of those grounds, so "discrimination between neighbourhoods" reads as a plausible legal claim rather than an error. The underlying fact is true, only the qualification is added, and the theme name carries no quote from the comment, so a reader who sees only the list of themes has nothing to compare it with. If it reaches the summary of recommendations under Article 19 of Law 146/2014, the institution answers an accusation nobody made, while the real request, repeating the notification in that neighbourhood, is lost.
The check that catches it
- For every theme that contains a legal term (discrimination, violation, abuse, conflict of interest), open the original comments the tool mapped to it and check whether the citizen uses that term or describes facts that support it.
- Ask the tool for the comment numbers and one verbatim quoted sentence per theme. A theme without a quote does not go into the summary.
- Rename the theme in the citizen's own facts ("leaflets were not distributed in neighbourhood X") and separate complaints about the notification process from comments on the substance of the draft act.
- Before the Article 19 summary, check notification complaints against the distribution record and answer them as procedural issues.
Where the failure pattern is documented: Government evaluations of AI tools for public consultation analysis: ASIC, Australia (March 2024), UK Department for Transport (December 2025), UK i.AI (August 2025).Sources: ASIC, Australia, 2024; UK Department for Transport, 2025; UK i.AI, Consult evaluation, 2025.The case above is ours, written for work in Albania; the laws it names were checked against the official text.
What it looks like
The citizen's comment"The leaflets about the consultation were not distributed in our neighbourhood, so many residents never heard about it."
The tool's summaryTheme 4: Discrimination between neighbourhoods. Residents of one neighbourhood complain of unequal treatment by the municipality.
Translated from the Albanian original.
Specialist courses for this profession
Public administration AI: Orientation
Public administration AI: Workflow
Public administration AI: Documents and communication
Public administration AI: Data and analytics
Public administration AI: Quality and evaluation
What you leave with
The letter checking sheet follows the six criteria, from procedure to legal limits, and records each problem found and each citation still awaiting verification in the official text. The responsible person reads it before signing the letter.
The workflow
You check the letter against six criteria: it follows the procedure, keeps transparency in mind, treats citizens equally, does not prejudge the outcome, is ready for audit, and states the legal limits clearly. Every citation is verified in the official text, and deadlines are taken from the Code of Administrative Procedures, not from the assistant.
Public Administration
The catalogue lists all 30 together, with filters for format and time.