The Government Digital Service (GDS) is testing whether artificial intelligence can help government information management professionals cope with a backlog of email so large that manual review is no longer sustainable. Project Kestrel, led by GDS's AI Data Readiness Incubator team in collaboration with the Government Knowledge and Information Management (KIM) team, is exploring whether machine learning can make the review process faster, more consistent, and easier to document.
Email has accumulated across government for decades, encompassing business decisions, administrative threads, sensitive exchanges, and personal messages. KIM specialists can be required to work through hundreds of thousands of items, deciding what should be kept, what meets legal and compliance requirements, and what may be sensitive enough to require careful handling. GDS said that at the scale at which government operates, carrying out that work manually "simply is not sustainable."
Synthetic data for safe experimentation
Project Kestrel faced an immediate obstacle: it couldn't safely use real government email for experimentation. Those messages can contain personal information, sensitive material, and legally protected content, creating risks if used to develop and test new tools.
To work around that, the team built Tiger Heron, a synthetic email generator that uses large language models to produce realistic examples resembling the content government teams encounter without including real or sensitive information. This allows teams to develop and test technology without accessing genuine government inboxes.
Machine learning brings order to email
The second component of Project Kestrel is Cuckoo, a machine learning capability designed to address the scale of the email management challenge. In a controlled test environment, Cuckoo reads large volumes of emails and groups them into clusters according to their content. Rather than opening every message individually, KIM professionals can use an interactive visual interface to see emails grouped around similar topics, threads, and types of content.
The approach is intended to help reviewers focus their attention, identify patterns, and make decisions that are easier to document and defend. GDS said the objective isn't to replace human judgement, but to support KIM professionals by making their work "faster, more consistent and easier to explain," while keeping people responsible for important decisions.
An experiment, not a finished product
Phase one has also involved user research with KIM professionals, ethics input, engagement with stakeholders across government, and the development of best practice materials. GDS said the phase was designed as an experiment rather than to produce a finished product, and further testing is still required.
For those working in government, Project Kestrel offers an early look at how AI can be applied to one of the public sector's less visible but critical workloads. The project's approach to synthetic data also points to a broader lesson: agencies can experiment with AI on sensitive datasets without compromising privacy or security. For AI for Government professionals, the project demonstrates a practical template for testing automation in compliance-heavy environments - one that keeps human reviewers in charge of final decisions while using machine learning to manage scale. It also aligns with the growing use of AI Agents & Automation in public sector workflows, where the goal is augmentation rather than replacement.
Why this matters for government professionals
If Project Kestrel progresses beyond the experimental phase, it could reshape how KIM teams allocate their time - shifting effort from triage to judgment. The immediate takeaway for government staff is that AI-assisted information management is moving from concept to tested practice, and the skills involved in reviewing, validating, and documenting AI-assisted decisions will become more valuable as these tools mature.
Your membership also unlocks:
