Processing tree-felling and public space permits

The permit process handles roughly 1,400 applications per year for tree-felling and public space use, arriving by paper or email. Each application moves through initial validation, assignment to one of two case officers, a substantive review drawing on property and spatial registers, and finally a drafted decision that must be reviewed and formally issued. The whole cycle takes about 35 minutes per case on average, adding up to over 800 hours of manual effort annually. The department head, two case officers, and a shared regional IT function together own the process, but in practice most of the working knowledge sits with the case officers themselves.

The process operates under administrative procedure law, publication duties, and GDPR, which means every decision must be traceable, justifiable on the file, and robust enough to survive an appeal. Roughly 85 to 90 percent of cases follow a recognisable standard pattern, but 10 to 15 percent require individual legal or policy judgment, and those exceptions demand genuine officer expertise. Currently there is no single system of record: email and spreadsheets carry the process from end to end, which creates fragility, inconsistency, and key-person risk. The department has no stated technology preference and is bound by central IT procurement rules, which shapes what any automation solution can realistically look like.

The recommended approach is a Hybrid AI workflow with human-in-the-loop gates, and the case for it rests on two things that pull in opposite directions: volume and compliance. The volume is high enough, and the standard case pattern clear enough, that automating intake validation, register lookups, and draft decision generation is straightforwardly worthwhile. The compliance picture, on the other hand, makes fully automated issuance legally unsafe. Every decision must be defensible on appeal, GDPR obligations attach to the personal data in each application, and administrative procedure law requires a traceable, accountable process. A Hybrid model resolves that tension directly: the automation handles the repeatable work for routine cases, and officers retain sign-off authority on every decision that goes out the door.

The client has no stated preferred technology stack and is constrained by central IT procurement rules. That constraint is taken seriously in this recommendation. The proposed tooling should be built on open-standards or low-code platforms that can pass existing procurement scrutiny, with register integrations handled through standard REST calls against the open-data sources that regional authorities already publish. This avoids a lengthy platform procurement and means the first stages of the build, particularly intake digitisation and automated validation, can begin with tools likely already approved in the region.

Comparing the Hybrid path against the alternatives makes the choice clearer. A fully automated AI Workflow could handle the standard 85 to 90 percent of cases well, but the moment it must issue decisions without human sign-off it crosses a legal line this department cannot cross. As soon as human sign-off is reintroduced, the AI Workflow effectively becomes the Hybrid path anyway. RPA alone could automate the structured, repetitive steps such as data entry and register lookups, but it cannot handle the unstructured paper and email inputs or generate natural-language draft decisions, so it falls short without significant AI augmentation. Traditional Code, meaning a custom-built case management system procured through central IT, would bring auditability and structure but would take far longer to deliver and would not address the judgment-intensive exception tail. An autonomous AI Agent is inappropriate here given the compliance stakes: unsupervised autonomous action in a process where every decision must survive appeal is not a responsible choice.

Education and process improvement alone would not reduce the 817 hours of annual manual effort or fix the structural fragility of knowledge held by two people. Training is a necessary part of the transition to the recommended Hybrid approach, particularly so that case officers can configure, supervise, and improve the workflow rather than simply use it, but it is not a substitute for the automation itself. Staying Manual carries the highest long-term risk: key-person dependency, inconsistent decision quality, and no headroom as application volumes grow.

The Hybrid path is also the most honest fit for the client's stated goal, which is freeing officers from routine work so they can focus on cases that genuinely need judgment. That is precisely what this model delivers: the AI handles the intake, validation, and drafting for standard cases, and the officer's time is reserved for exceptions, quality review, and final sign-off. The result is not a smaller team but a more capable one, with a process that is visible, auditable, and no longer dependent on tacit knowledge that could walk out the door.

The two closest alternatives to the recommended Hybrid path are the AI Workflow and RPA paths, and it is worth being clear about why neither quite fits. An AI Workflow without human gates could, in theory, handle the routine 85 to 90 percent of cases with reasonable accuracy. The problem is not the technology; it is the legal context. Administrative procedure law requires that every permit decision be justifiable on the file and capable of surviving appeal, and that accountability sits with a named officer, not a system. An AI Workflow that issues decisions autonomously removes that accountability link in a way that creates direct legal exposure for the municipality. The moment you add a genuine human sign-off gate to fix that problem, you have rebuilt the Hybrid path anyway, so the AI Workflow path without oversight is not a serious option here.

RPA scores lower for a different reason. The process inputs are unstructured paper and email, and a meaningful portion of the work involves natural-language reading, judgment about eligibility, and generating coherent draft decisions. RPA handles structured, deterministic repetition well, but it cannot reliably parse a scanned paper application or draft a legally coherent refusal letter. You would need to add significant AI components to cover those gaps, and at that point you are building a more complex and brittle version of the Hybrid path rather than a simpler one. RPA is not wrong as a component, but it is the wrong foundation.

The Hybrid path wins because it is honest about what the process actually requires. The standard pattern is clear enough to automate the intake, validation, register lookup and drafting steps, and at 1,400 cases a year that automation delivers real time savings. But the exception rate is too high, the compliance stakes too serious, and the decisions too consequential to remove officers from the loop. A well-designed Hybrid keeps officers focused on the 10 to 15 percent of cases that genuinely need their judgment, gives them a reviewed and validated draft for the routine cases, and preserves full legal accountability at every step.

The first and most important step is not technical. Before any tooling is selected or integrated, the standard permit pattern needs to be documented in enough detail to distinguish it reliably from exception cases, and the intake channel needs to be converted from unstructured paper and email into a structured digital form. A tool like Typeform, Microsoft Forms, or a simple government-approved web form serves this purpose, depending on what has already cleared regional IT procurement. The form captures all mandatory fields up front, which means every subsequent automated step operates on clean, consistent data rather than trying to parse a scanned document. This intake digitisation step is the dependency that everything else rests on, and it is primarily analytical and change-management work rather than engineering.

Once structured intake data is flowing, automated validation and register lookups can be layered in. The validation logic checks completeness and eligibility against the rules the case officers already apply manually, and it calls available open-data registers for spatial and property data using standard REST API calls. A workflow automation platform such as n8n, Make, or Microsoft Power Automate handles this orchestration and is well within reach of a low-code procurement approval. Cases that pass all checks are flagged as standard and move forward; cases that fail a check or trigger a defined exception rule are placed in a case officer queue with the specific reason for escalation surfaced clearly. This triage step alone removes a significant portion of the daily coordination burden.

For the standard cases that clear validation, an AI-assisted drafting layer generates a structured draft decision by combining the validated application data with a library of approved decision templates. A large language model accessed via API, such as the OpenAI API or Azure OpenAI Service (which carries stronger data residency and compliance options relevant here), takes the structured case data and a relevant template and produces a draft that the assigned case officer reviews before sign-off. The draft is explicitly a starting point, not a finished product, and the interface should make that framing visible to officers. The template library needs to be signed off by the department head and reviewed by the regional legal team before the drafting layer goes live; the quality of those templates directly determines the quality of every automated draft.

The case officer experience is tied together in a simple case management dashboard, which can be built in Power Apps, Retool, or a comparable low-code tool depending on what is available in the region. The dashboard shows each officer their active queue, flags exception cases with the escalation reason, and tracks each case through validation, drafting, review and issuance. This replaces the current spreadsheet and email coordination with a single view that is also visible to the department head. Running the full hybrid workflow in parallel with the existing process for four to six weeks before full cutover gives the team time to calibrate exception routing rules, improve template quality based on real cases, and build confidence in the system before the manual fallback is retired. Every automated action is logged with a timestamp, actor reference and outcome to satisfy administrative procedure law and provide a clean audit trail for any appeal.