Fix the step where cases wait before you add AI to the front door.

Backlogs are counted, not explained

Applicants for a permit, a license or a benefit wait without knowing where their case sits. People who complain may never learn what happened. Managers see the backlog but not always the step that causes it, and two officers can read the same legal rule differently.

We think the fix starts with the timestamps the case system already keeps.

What we would model, and with which data

Open data comes first, and internal records are added only where they change the answer.

• Event logs from case systems, which record when each application or complaint entered and left each step.
• Service requests. New York City publishes its 311 requests as open data, and cities such as Helsinki, Chicago and Toronto implement the Open311 standard for sharing them.
• Grievances. India's national portal, CPGRAMS, gives each grievance a registration number the complainant can track.
• Staff rosters and workloads, so the model can tell a staffing gap from a broken step.
• Published datasets. The OECD found that only 47% of OECD governments' high-value datasets are openly available, on average.

How the model is built and checked

Process mining rebuilds the real path of each case from its timestamps, including loops and rework missing from the official process map. A queue simulation replays a past year of applications under a change, such as a new triage rule or two more reviewers, before staff and applicants live with it. Rules as code turn eligibility and permit criteria into tests that run against past decisions, and an officer checks every case where code and decision disagree. A language model sorts complaints by topic and urgency, and staff label a sample by hand to measure its accuracy. Each model is back-tested on a past period first.

What decision-makers get

• A map of each service that shows where cases wait and why.
• Costed options to cut waiting time, each tested in simulation first.
• Complaint trends by place and topic, linked to the service that caused them.
• A decision record with the rule, the evidence and the officer who signed, published as open data where the law allows.

We would judge the work by one number, the change in waiting time. A first project picks one service with a visible backlog, such as building permits. We would rebuild its process map from a sample of event logs, back-test the queue model on a past year and hand over a scorecard. The public body then decides whether to continue.

How it runs safely

In the OECD survey, about four in ten people were confident that government AI could make services more tailored, and fewer that it would be used with transparency, fairness and protection of personal information. Until that confidence is earned, AI should sort and draft, and staff should decide every case, with a named official signing off.

• Applicants are told when AI sorted their case or drafted a reply, and they can ask for a human review.
• Personal data stays in the public body's systems and is masked before any language model sees it.
• Fairness checks compare waiting times and outcomes across neighborhoods and groups.
• In the EU, the AI Act classes AI that judges eligibility for essential public assistance benefits and services as high-risk, and the GDPR limits decisions based solely on automated processing. This page is not legal advice.

Open-source building blocks

Open tools let anyone check a public body's process maps and coded rules.

• PM4Py, a Python library for process mining.
• Catala, a research language from Inria for turning legislative text into code. Its authors note the compiler is not yet stable.
• FixMyStreet, mySociety's software for reporting street problems on a map and sending them to the right authority.
• Decidim, a participatory democracy framework.
• CKAN, a data management system for open data portals.