.png)
BY NOVOVILLE
Inside the HumAIne Smart Cities pilot: AI-assisted parking permits
Every municipality that issues resident parking permits faces the same bottleneck. Citizens submit a form and a supporting tax document. Someone has to check that the two agree, application after application. It is repetitive, it is error-prone when done at the end of a long day, and it is exactly the kind of work where an AI system can help, as long as a human stays in charge of the outcome.
That was the brief for the Smart Cities pilot of HumAIne, the EU-funded project on hybrid human-AI decision support. Novoville built the pilot for a Greek municipality, and ran it on real applications with real municipal operators. The video above is the full walkthrough. This post tells the story behind it.
Three actors, one workflow
Citizens apply through a public form, with no login, and upload their E1 tax document. Council operators run the review. The AI checks each application against the document and flags the cases that need a closer look. Applications are collected in sessions, so the operator closes a session, the AI processes it, and the operator then reviews and decides.
One principle holds throughout, and we kept it visible in the design: no AI decision adverse to a citizen is final. Every flagged or rejected application is adjudicated by a council operator. The AI recommends; the operator decides, and the two are recorded as separate things in the audit trail.
Built on the HumAIne platform
The pilot does not stand alone. It plugs into shared HumAIne platform tools: a data hub on MinIO, a workflow scheduler with Node-RED flows that move data between tools, the Benchmarking Suite from Harokopeio University, which turns operator behaviour into metrics, and the Swarm Assistant from Innov-Acts, which lets anyone question the data in plain language. The two layers meet only through an anonymised, offline export, so citizen data stays in the city and the AI runs securely on municipal systems.
Three models, each one steered by evidence
The pilot went through three model versions, all evaluated on the same set of real applications and reviewed by the same municipal operators.
IBM, the first model, drove the operator interface through browser automation (RPA). The Benchmarking Suite showed that waiting for the AI dominated the review session.
NOVO replaced the UI navigation with a direct API integration. The AI response time dropped sharply.
IBM2 added autonomous fuzzy-match correction. The number of flags fell noticeably, with fewer cases sent to operators for no good reason.
This is the point of the platform in one sentence: each upgrade was chosen because the benchmarking showed a specific weakness, not because it seemed like a good idea.
Two key exploitable results
The pilot produced two results that other municipalities and vendors can reuse.
The parking permit interface. A session-based workflow, developed by Novoville, with privacy by design, role-based access and a full audit trail. A session moves from open to stopped, AI check, verifying and closed. Citizens can apply only while it is open and are notified only at the end. In every branch, the operator has the last word.
PDF information extraction and matching. Developed by our partner IBM, a system that fuses rule-based matching with flexible AI document parsing: a deterministic extractor for the standard Greek E1 tax form, an LLM fallback for anything that is not a clean PDF, a fuzzy comparer, and a transparent verification result. A city official can add a simple text hint, and the system remembers it. No coding expertise is needed from municipal staff.
Asking the data in plain language
In the second half of the video, the Swarm Assistant answers questions about the pilot without anyone writing code. We show six prompts in order: a data discovery question, a comparison of AI and operator processing times that needs several tools, a documentation question, a request for the HAIC benchmark results, a request to define the HAIC metrics, and an out-of-scope question, which the assistant declines politely. In a capability test with a fixed set of prompts, in scope and out of scope, two small fixes to the routing rules and the tool descriptions were enough for the assistant to handle every prompt as intended.
Did it work?
All the technical targets were met in every round of the evaluation.
Precision stayed very high and identical across the three models.
Recall comfortably cleared its target and held steady.
Processing time fell to a small fraction of the manual baseline, well inside the target.
Residual operator effort kept shrinking from one model to the next.
The human side tells a consistent story. Operators spent less time per case, a larger share of cases was handled automatically, and the operators who answered our questionnaire were overwhelmingly positive about the platform. Their view of the AI module itself was positive to neutral, which is a fair reflection of a first deployment, and it gives us a clear direction for the next iteration. Citizens see only the outcome of an application, so bringing their feedback into the loop is a natural next step.
Risks, handled openly
The pilot keeps an AI-aware risk register for the risks specific to the AI-assisted system. The main ones, such as no time saving, a partner failing to deliver, and false positives with no remedy, are closed. Operator distrust and staff feeling threatened are mitigated. Explaining decisions to citizens is the one to keep watching, and its severity is limited by the architecture, because the decision a citizen would contest is a human decision, open to the municipality’s existing appeal procedures.
Beyond the project
The approach scales with compute, not headcount. A session costs some extra machine time and saves considerably more operator effort, a clear net saving per session. Other Greek municipalities use the same national tax form, so the same extraction applies; for other formats or languages the model adapts. Many municipalities have been engaged in three tiers: one live, several trained and the rest through demos. Along the way we ran webinars with live polls, and the roadmap points to IEEE 7001, ISO/IEC 42001 and the EU AI Act.
The structure is common in public administration: a structured submission, a supporting document, a rule-governed decision, and an expert human who keeps authority. That is why we think the lessons travel.
What each did for the other
The pilot gave every platform tool real data, real operators and a real decision flow to prove itself on. The platform turned operator behaviour into evidence, and evidence into better models.
With thanks to our partners
The pilot stands on the work of HumAIne partners. The PDF information extraction and matching system was developed by IBM, the Swarm Assistant by Innov-Acts, and the Benchmarking Suite by Harokopeio University. The Smart City Permits interface and the pilot itself were built by Novoville. Thank you to all the teams involved.
Watch and get in touch
The full demo is available at: https://www.youtube.com/watch?v=zalQX04dLCU
The data shown is synthetic and anonymised. If you run a municipal service and would like to see this in action, or if you are a researcher or vendor working on human-AI decision support, contact Aristodemos Pnevmatikakis, Chief Innovation Officer at Novoville, at aristodemos@novoville.com. You can also follow the wider project at https://humaine-horizon.eu/.
HumAIne (Hybrid Human-AI Decision Support for Enhanced Human Empowerment in Dynamic Situations) is funded by the European Union under Horizon Europe, project 101120218. Novoville is funded by UK Research and Innovation under the UK government’s Horizon Europe funding Guarantee, project 10087361. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the funding organisations or granting authorities. Neither the funding organisations nor the granting authorities can be held responsible for them.