Monitoring council minutes and resolutions

The monitoring of council minutes and resolutions is a public-sector intelligence task carried out by two case officers and a department head. Several times a week, councils and municipalities publish new minutes and resolutions on their public websites and open-data registers. The team must check those sources, read through the documents, decide which topics are relevant to their department, assess the urgency of any action required, summarise the key points, and then alert the right people in time to respond. At roughly 600 runs per year and an average of 40 minutes per run, the process consumes around 400 hours of staff time annually, almost all of it spent on low-value reading rather than on the judgement and response work the team is actually paid to do.

The process sits across email and spreadsheets with no single system holding it together. Some public registers publish open data, but council documents are largely unstructured text, and the formatting varies significantly from one source to another. About a quarter of the documents the team reviews involve ambiguous or cross-cutting topics that genuinely require human judgement, whether because a resolution spans multiple jurisdictions, because the relevance to the department is not immediately obvious, or because the consequences of getting the call wrong are significant. Administrative procedure law and GDPR require that every decision is traceable and justifiable on the file, and that any alert or action taken can survive an administrative appeal. That compliance context shapes the entire design of any solution.

The core problem is latency and concentration of knowledge. When documents arrive in bursts, the team can fall behind, and a missed resolution or a late alert can mean a missed statutory deadline. The process depends on the availability and attention of two or three specific people, which creates fragility. The goal of automation here is not to remove human judgement but to remove the manual scanning and reading that currently consumes most of the time, so that officers spend their attention on the cases that actually need them.

The recommended path is Hybrid: an AI workflow with human-in-the-loop review gates. The case for this is straightforward. The process has a large routine component, roughly 75 percent of documents can be reliably classified and summarised without a human reading them in full, but it also has a well-understood and persistent exception pattern that requires human judgement and generates a compliance obligation. A Hybrid design handles both halves: the AI pipeline does the scanning, classification, and summarisation, and the human gate catches the ambiguous cases, logs the officer's decision, and produces the audit trail that administrative procedure law requires. The design matches the client's own stated goal of freeing officers from routine reading while keeping them accountable for the calls that matter.

The client has no stated technology stack preference, which is not unusual in a public-sector setting, but the constraint here is more specific than a simple lack of preference. Central regional IT is slow to change, and procurement is governed by public-sector rules that can add months to any project that requires new infrastructure. This shapes the recommendation directly. The solution should be built on a cloud-hosted, procurable platform with open APIs, one that can connect to public data sources without requiring bespoke infrastructure or a lengthy IT approval process. A well-documented workflow tool with a native LLM integration, something that can be procured through an existing framework rather than a new tender, is the realistic fit. Bespoke infrastructure, however technically elegant, would almost certainly stall in procurement and is not recommended.

The alternatives were considered carefully. A fully automated AI Workflow, scored at six out of ten, could handle ingestion, classification, and summarisation competently, but removing the human gate entirely creates legal and operational risk that the client cannot accept. Decisions must survive an appeal, the exception rate is roughly 25 percent, and GDPR obligations require a documented basis for processing and a clear chain of accountability. Full automation without oversight does not satisfy those requirements. An AI Agent operating autonomously scores lower still, at four, because the guardrails needed to manage the exception variety and the compliance obligations would effectively recreate a hybrid design with additional complexity and less transparency.

RPA and Traditional Code both score four, and for similar underlying reasons. Council minutes are unstructured text. RPA can scrape structured data from consistent sources but cannot detect topic relevance or produce summaries, and the inconsistency of public document formats would cause frequent pipeline failures. Rule-based keyword matching can catch some obvious topics, but it would miss nuanced or novel resolutions and generate false negatives at exactly the moments that matter most, when a council uses unfamiliar language to describe something the department needs to know about. Maintaining a keyword ruleset as topics evolve would also create ongoing IT overhead that the slow central IT function cannot absorb. Neither path is viable as the primary approach, though lightweight source-monitoring scripts could play a supporting role in the ingestion layer.

The honest tradeoff in the Hybrid recommendation is cost and timeline. The build cost estimate runs from roughly 18,000 to 35,000 GBP, and the payback period at a conservative 65 percent automation rate is around 24 months. That is not a fast return, and it assumes the ingestion pipeline is reliable enough that officers trust it. If source formats are more inconsistent than expected, or if the topic taxonomy takes longer to calibrate, the automation rate will be lower and the payback longer. The roadmap addresses this by running the hybrid system in parallel with the manual process for four to six weeks before decommissioning the manual routine, which gives the team real data on precision and recall before committing. The compliance gate is not optional; it is the feature that makes the solution legally defensible, and it should be treated as a design requirement from the first day of the build, not a feature added at the end.

The two strongest contenders are Hybrid and AI Workflow, and the gap between them comes down to compliance accountability rather than technical capability. An AI Workflow could handle the ingestion, classification and summarisation layers just as well as a Hybrid design, and for a process with a low exception rate and no regulatory accountability it would be the cleaner choice. Here, though, roughly a quarter of cases require contextual judgement, and administrative procedure law demands that every consequential decision is traceable to a named officer who can defend it on the file and in an appeal. Removing the human gate does not just increase operational risk; it creates a structural compliance gap that no amount of prompt engineering can close.

The Hybrid design wins precisely because it preserves that accountability by design. Routine cases, those the classifier handles with high confidence, flow straight to an alert without officer involvement, which is where most of the 400 hours of annual effort gets recovered. Ambiguous or low-confidence cases are routed to a review queue where a case officer confirms, corrects or escalates before anything is sent. That gate is not a workaround for AI limitations; it is the mechanism that makes the solution legally defensible. The AI Workflow score of six reflects the fact that the detection and summarisation layers are genuinely strong candidates for full automation, and the Hybrid design inherits all of that, it just adds the review gate on top rather than relying on the AI alone to decide what is consequential.

The remaining paths, AI Agent, RPA and Traditional Code, all score four and are not realistic contenders. An AI Agent would need so many guardrails to satisfy the compliance requirements that it would effectively recreate a Hybrid design with more complexity and less transparency. RPA cannot handle unstructured text, and the inconsistency of public council document formats would cause frequent pipeline failures with no semantic fallback. Traditional Code based on keyword matching would generate unacceptable false-negative rates on novel or nuanced resolutions, and maintaining rules as topics evolve would impose an ongoing IT overhead the central IT function cannot absorb.

The implementation starts with a source and taxonomy audit before any code is written. The case officers, with light facilitation, map every public council website and open-data register they currently monitor, assess how consistently each source publishes new documents, and document the topic categories they actually care about. This phase produces two outputs: a source inventory with reliability ratings, and a written escalation policy that defines what separates a routine alert from a case that needs human review. Neither output requires IT procurement approval, and having them documented is what makes the classifier trainable and the human gate operable.

The ingestion and classification pipeline is built next, using a scheduled fetch process, something like a Python-based cron job or a low-code workflow tool such as Make or n8n, that polls each source for newly published PDFs or HTML pages. New documents are passed to an LLM-based classifier and summariser, realistically GPT-4o or a comparable model accessed via the OpenAI API or Azure OpenAI Service, which is procurable under most public-sector framework agreements. The model is prompted to assess relevance against the documented topic taxonomy, assign a confidence score and produce a short structured summary covering the resolution, the decision body and the implied action or monitoring requirement. Outputs are written to a lightweight structured store, a simple database table or even a well-structured spreadsheet at pilot stage, alongside the source URL and a timestamp for the audit trail.

The human-in-the-loop gate is implemented as a review queue, which at minimum can be a filtered view of that same structured store with a web form or a tool like Airtable or Notion layered on top, accessible to the case officers without specialist IT involvement. Items below the confidence threshold, or flagged by the classifier as cross-jurisdictional or ambiguous, appear in the queue with the AI summary and a link to the source document. The officer reviews, confirms or corrects the classification, records a brief rationale, and approves the alert. That logged decision is the compliance record. Only after approval does the alert travel outward, delivered via email using a service such as SendGrid or through the team's existing Microsoft 365 setup, keeping the IT footprint small and avoiding a new system for recipients to learn.

The final phase runs the Hybrid system in parallel with the manual process for four to six weeks, comparing AI-generated alerts against what officers would have caught manually, measuring missed topics and false positives, and using those findings to adjust the confidence threshold and refine the prompts. The department head signs off on the calibration before the manual parallel process is retired for routine cases. Source health checks, simple automated pings that alert the team if a monitored website stops returning new documents, are built in from the start so that a source going offline does not result in silent monitoring gaps. A quarterly review cadence is established to catch topic drift, source format changes and any shifts in the department's area of interest that the taxonomy needs to reflect.