Freedom of information (FOI) requests represent a vital pillar of democracies worldwide, but one that governments have found increasingly challenging to fulfill for citizens.

The US government has a backlog of more than 200,000 FOI requests, a record high. The UK government’s response rate has sagged in the face of a record number of requests. In Australia, a Senate inquiry called its system “dysfunctional and broken.”

While these backlogs put pressure on public resources and are a source of frustration for citizens, the highly manual and database-centric nature of FOI work plays to generative AI’s core strengths and represents an opportunity for AI to have a positive impact on a process vital to maintaining healthy democracies.

FOI laws (known as FOIA in the United States, ATIP in Canada, the Public Records Act in the UK, and Regulation 1049 in the European Union) provide a crucial means for citizens to access government documents and communications, thereby serving to promote transparency in government.

AI agents can be used to enhance the agility and responsiveness of FOI systems, drastically reducing processing times, enabling records officers to handle more cases, and significantly lowering error rates. Agentic platforms, like North, can be deployed privately and securely, addressing a key consideration for governments, which have so far been cautious in adopting AI compared to the private sector.

Governments we work with are actively exploring the use of AI (North for government) to enhance nearly every aspect of the current FOI process, from triaging requests to assessing their sensitivity and completeness, to retrieving records. It’s just one way in which public sector productivity and services stand to benefit from AI over the long term.

Current FOI bottlenecks

The basic premise of FOI laws is simple: If you ask to see a government record, the government has to either show it to you or give a reason for keeping it secret.

Unfortunately, the reality of public access often falls short of the ideal. Backlogs and delays are an issue in nearly every country that maintains a FOI law, and the slow pace of access blunts its effectiveness as a tool for accountability. A recent survey by the EU ombudsman found that requests routinely take over a year to be filled. Canada’s national archives office met only 61% of requests on time in 2023/24.

There are many reasons for this slow pace, but the primary challenge is understaffing and a reliance on manual processes. Faced with competing demands and limited resources, governments rarely prioritize FOI. Responding offices are starved for resources, and transparency suffers as a result.

How the process works now

Most FOI systems start with a portal where citizens can make requests. While some countries have a single portal, others, such as the United States, maintain separate intake systems for each agency, resulting in a tangle of different record offices with varying levels of effectiveness. Even where systems are unified on the front end, each department typically manages its own records and maintains its own public records officer for handling requests.

From there, each request is assigned to a specific officer who communicates directly with the filer. Often, there are particular notification requirements (e.g., when a request is received or when a certain amount of time has passed), but communication is typically handled manually through email or even phone calls. The officer may sometimes propose modifications to the request, narrowing it in ways that result in faster fulfillment or fewer extraneous records; however, these modifications must be approved by the requesting party in advance.

The biggest pain point is redaction. Once relevant documents have been assembled, records officers will review them manually to remove trade secrets, classified information, or other sensitive material. However, this is typically performed manually, and as requests often span hundreds of pages, it can be highly labor-intensive.

How AI can transform the FOI process

The good news for governments struggling to meet citizens’ transparency expectations is that these processes align extremely well with AI’s strengths.

A significant portion of handling each FOI request involves repetitive, rule-based tasks, such as reviewing emails, memos, and reports to identify relevant documents and verify them against statutory exemptions. This initial phase of document discovery is the kind of task that AI models can perform in seconds.

Another challenge is that the rules of document discovery must be applied across a large and inconsistent information landscape. One common request, for example, is for “all communications” from a particular office holder over a specific span of time, limited by specific keywords or subjects. While the phrasing of the request is simple, it could involve collating records from half a dozen different systems over the course of years. It’s the kind of request that can be frustrating, tedious, and time-consuming to perform manually.

But AI models are adept at navigating multiple systems in this way, which means they’re much better at the initial phase of document discovery. Handing off that part of the process to AI systems would allow requests to be fulfilled much faster, freeing up staff to handle the parts of FOI work that require more expertise and judgment.

AI models' natural language capabilities also enable them to translate written requests into structured queries. By training them on a large corpus of filed requests, agencies can tune models to capture the nuances of different types of requests and various redaction rules.

Since FOI requests are limited to unclassified data, the AI system wouldn’t need access to any highly sensitive information. Governments will also be able to run AI-powered FOI systems safely through private deployments, ensuring that none of the data being accessed by the AI model risks leaking out of their secure environment. Permission and access controls can be used to add firewalls between different departments and job functions.

Deploying AI agents across FOI workflows

Agentic AI systems can be tailored to specific tasks and existing FOI workflows while ensuring that data and sensitive information remain within the appropriate departments.

With human oversight and authentication reviews, AI agents can perform assigned tasks for each step of a workflow. Consider, for instance, a records department that deployed fine-tuned (meaning tailored) agents to manage each of the following tasks:

AI agent for department routing and triage: Records requests often require manual routing to the relevant departments for processing. A workflow-embedded agent could automatically identify the appropriate departments and recommend interdepartmental consultations as needed. Over time, it could also alter the overall routing system to optimize processing time.

AI agent for sensitivity assessment: Records vary significantly in their sensitivity, and much FOI work involves determining which requests are sensitive enough to warrant additional oversight. An AI agent could assess sensitivity based on previous decisions and flag potentially sensitive requests in advance, giving departments far greater flexibility in how they handle more delicate records. The same AI agent could also flag records that would be exempt from the request, based on the country’s laws.

AI agent for clarification and information gathering: Records requests are often ambiguous, requiring labor-intensive clarifications between the requesting party and the records officer. These clarifications typically address the scope of the request, including whether it should be time-gated or otherwise limited. (As veteran requesters know, getting thousands of pages of redacted or unresponsive documents can be worse than getting no response at all.) An embedded agent could be explicitly tailored to assess the completeness of the request and automatically follow up with any necessary questions.

AI agent for record search and retrieval: A more advanced AI agent could be deployed for the search and retrieval of records, automatically converting natural-language requests into structured queries. Instead of starting from scratch, records officers could begin each request with a pre-filled selection of relevant documents, requiring only the manual selection and redaction of documents before release. The same worker could handle many more requests, enabling agencies to make significant progress in reducing their backlogs.

In each case, the agent could be trained on actual work tasks to learn the nuances of how they are processed. By being embedded in workflows in this way, the agent can learn not only the unique demands of the task but also the unique practices of each workplace.

This agent-empowered system would produce major productivity gains across the FOI process. We estimate processing times could be cut by 50-75% while record officers would be able to handle 2-3 times more requests than in current systems. Errors in redaction and document discovery could also fall by around 30-50%. These improvements enable departments to reduce backlogs by an estimated 40-60% within a year of implementation.

See visual below for the entire workflow.

Example of AI agents with human-in-the-loop for processing FOI requests.

A gateway to wider government AI

Governments face a unique set of challenges and concerns around implementing AI, not least of which is winning over public trust in the technology. In this context, the overwhelmed FOI request systems represent a powerful opportunity to show how AI can deliver tangible improvements to a vital, core citizen service without undermining privacy or security. It can act as a gateway to scaling AI securely across other government services.

Get started with North for government today.

Request a demo!

David Ferris