Handling freedom of information requests with AI: faster searching, not faster judging

By · Updated · 3 min read

Short answer

Freedom of information requests run on short deadlines: in the Netherlands four weeks, extendable once by two. Most of the work is finding, sorting and redacting, not deciding. AI can find documents, bundle them, remove duplicates and flag passages that may fall under an exemption. The assessment per passage and the decision stay with an employee, because wrongful disclosure cannot be undone.

Where the work sits

The deadline is short. Under the Dutch Open Government Act, for example, a request must be answered within four weeks, extendable once by two weeks where the size or complexity of the information requires it. For a request touching thousands of emails and documents, nearly all the time goes into four steps:

  1. finding what falls within the request, spread across mail, case systems and network drives;
  2. sorting and removing duplicates;
  3. assessing per passage whether an exemption applies;
  4. redacting and giving reasons for the decision.

What AI takes over, and what it does not

StepWhat AI doesWhat the employee does
FindingSearch by subject and period, including attachmentsDecide what falls within scope
SortingBundle, recognise duplicates and versionsSpot-check completeness
FlaggingPoint out personal data and possible exemptionsAssess and weigh each passage
RedactingPrepare proposed redactionsConfirm, add, finalise
DecisionDraft text referring to the documentsGive reasons and sign

Always have flags point to the source passage. See Summarising confidential documents with AI.

The risk runs one way

With redaction, mistakes are not symmetrical. Redacting too much brings a complaint; redacting too little brings a data breach or a disclosed trade secret, and that cannot be taken back. So let AI propose and never publish on its own.

Mind the documents themselves too: such a file nearly always contains personal data of officials and third parties. Uploading them to a public AI service is processing you cannot justify. Processing belongs in an environment of your own, with permissions that follow the source system. See Access control for AI.

A workable approach

  1. Start with requests of average size, not with the largest file.
  2. Measure the current lead time and the number of extensions.
  3. Let AI search, bundle and flag; the lawyer assesses.
  4. Track how often a proposed redaction is changed; that is your quality measure.
  5. Record the application in your AI register, with purpose, sources and owner.

With werqly this runs on a dedicated node in your server room, so the documents never leave the organisation. See AI for local government and Generative AI in local government.

Frequently asked questions

Questions on this topic

May AI redact documents for a disclosure decision?

AI may propose redactions, but the assessment per passage and the decision stay with the public body. Wrongful disclosure cannot be undone, so have an employee confirm every proposal before anything goes out.

What is the deadline for a Dutch Woo request?

Four weeks from receipt, extendable once by two weeks where the size or complexity of the information requires it. That extension is communicated to the requester in writing, with reasons, within the first period.

May such documents go into a public AI tool?

No. A disclosure file nearly always contains personal data of officials and third parties, and sometimes commercially confidential information. Process those documents in an environment that belongs to your organisation alone.

How do we measure whether it helps?

Measure lead time, the number of extensions and the hours per request beforehand. Measure the same afterwards, plus how often employees change a proposed redaction. That last figure says something about the quality of the support.

Sources

  1. Dutch Open Government Act (Woo), Article 4.4 on the decision period
  2. Government of the Netherlands — What happens after a Woo request (Dutch)

This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.

About the author

Tim Hoogstrate

Founder of werqly

Tim Hoogstrate is a software developer and the founder of werqly. He writes the articles in the knowledge base: what the GDPR, the AI Act and sector rules require when you use AI, and how to arrange that in an environment of your own.

Calculate your private AI costs.

Enter your work email and number of employees. We work out four setups for you straight away: in your own server room or in our datacentre, self-managed or with werqly Pro management. werqly is not a cheap AI subscription: you pay for your own hardware and full control.

  1. 1Enter your email and number of employees
  2. 2Compare four proposals
  3. 3Request a quote straight away

We email the calculation to this address and use it only to contact you about your request.

A no-obligation quote for your own AI node, in your inbox within a minute.

Request a quote →