Greece and Presight want agentic AI in city services. Put the service meter before the platform

THE SIGNAL IN ONE SENTENCE

The phrase smart city has survived every technology cycle by being pleasantly difficult to measure. A city installs sensors. A vendor adds a dashboard. An official points to a map covered in glowing dots. The press release calls the place more responsive. The bins are still full on Tuesday. Greece's Ministry of Digital Governance and Presight, an Abu Dhabi-based data and AI company in the G42 group, signed a memorandum of understanding in Athens on October 22. Presight says the two sides will explore agentic AI for municipal waste management, water-network monitoring and technology that could improve safety on public transport. The announcement says the work may use synthetic and open data. It does not name a municipality, contract, budget, deployment date, model, operating system, data inventory or performance target. Presight also says the parties still need to define the scope, structure and timeline, subject to final agreements and regulatory approvals. That is the correct place to start this story. This is an exploration agreement. It is not a working service. It is not a procurement. It is not evidence that an AI agent can make a city cleaner, protect a water network or keep passengers safe. The useful opportunity arrives before any platform is chosen. Greece can require a public service meter. The meter would sit above the vendor dashboard and below the political promise. It would show residents what the service did before AI, what changed during the pilot, what the system suggested, what workers decided, what went wrong and how a person can challenge the result. Start with the boring baseline. For waste collection, publish missed pickups, average resolution time, route distance, fuel use, worker overtime, overflowing-bin complaints and the share of service requests closed without a second visit. Split the figures by neighborhood, municipality, season and service type. A model that improves the citywide average while leaving an island, hillside settlement or lower-income district waiting longer has not made the service smarter. It has moved the inconvenience. Then define the agent's job in verbs. Does it predict that a container will fill? Does it rearrange a route? Does it ask a contractor to dispatch a truck? Does it close a complaint? Does it merely draft a recommendation for a municipal supervisor? Those are different systems with different consequences. The word agentic can make a sequence of software calls sound like a tiny city manager. Do not grant the software a job title. Publish the permitted actions, data it can read, systems it can write to, spending limits, escalation rules and steps that always require a worker. Every action should create a record. The record needs the input, model and software version, recommendation, confidence or uncertainty signal, human decision, final action, affected service area and correction. If the system changes a waste route, the driver should be able to see why. If the driver rejects it because a street is closed, a school is dismissing students or the suggested turn is impossible for the truck, that correction should improve the operating record rather than disappear as noncompliance. Workers are not noisy sensors attached to the AI. They are the people who know the pavement. Water networks need an even stricter meter. The public measure is not how many anomalies the system found. It is whether the system helped crews locate real leaks, pressure failures or quality risks faster without flooding the control room with false alarms. Publish true alerts, false positives, missed incidents, time from alert to inspection, time from inspection to repair, water saved, repeat failures and the number of times an operator overrode the recommendation. Separate detection from diagnosis. A pressure change can come from a leak, maintenance, demand, a faulty sensor or a valve state that the data system did not receive. An agent may combine records and suggest a cause. The pipe still gets the final vote. Before an automated action can change pressure, isolate a section or create a work order, require a named human owner and a safe fallback. The fallback should work when the model is unavailable, the network is degraded, the sensor feed is stale or the vendor connection disappears. A city service that cannot operate without the clever layer is not resilient. It is dependent. Public transport safety requires the clearest boundary of all. The Presight announcement mentions technological solutions that could enhance safety on public-transport networks. It does not specify the data or proposed actions. That blank space matters. Safety systems can range from equipment monitoring and crowd alerts to video analytics, incident classification and automated dispatch. Each one affects different people and carries a different privacy risk. Publish the exact purpose before collecting the data. Name the cameras, sensors, vehicle records, location streams, incident reports and other sources. State whether biometric identification, face recognition, behavior scoring or individual tracking is prohibited. If a system flags a person or event, record who reviews it, what evidence they see, how quickly they must decide and what happens when the alert is wrong. A false positive in a maintenance queue is annoying. A false positive involving a passenger or police response can be frightening and consequential. The service meter should count the people who experienced that mistake, not just the alert that the system later marked resolved. Synthetic data needs its own disclosure. Presight says the proposed exploration may use synthetic and open data. Synthetic data can help teams simulate rare leaks, unusual demand, vehicle incidents or combinations that a small historical record does not contain. It can also manufacture a tidy city that behaves nothing like Athens, Thessaloniki, Patras, an island municipality or a rural network. Publish how synthetic data was created, which real distributions shaped it, what scenarios were added, what populations and neighborhoods are missing, and which tests use real operational records. Do not report performance on synthetic incidents as if the city lived through them. Open data needs provenance too. Open does not mean current, accurate, complete or suitable for a consequential decision. Record the publisher, license, collection method, update frequency, known gaps and the service decision that uses each dataset. If a route model relies on a road dataset that is six months old, the meter should say so before a driver meets a new barrier. Next, publish the handoff chain. Agentic systems are often sold as a smooth loop from detection to action. City services are not smooth loops. They include municipal departments, utilities, transport operators, contractors, emergency services, unions, regulators and residents. For every alert, show who received it, who was authorized to act, who actually acted, when the issue was closed and who can reopen it. That final step is essential. An AI system can decide that a complaint is duplicate, low priority or solved because a work order closed. The resident may still have no water. Give every affected person an appeal path that does not require understanding the model. A person should be able to say the bin was not collected, the leak remains, the route change created an accessibility problem or the safety alert concerned the wrong person. The appeal should reach the agency that owns the service, not a vendor help desk. Publish the time to acknowledge, investigate, correct and notify. Count reversals. Count repeat complaints. Count the cases where nobody could explain the decision. Procurement must follow the meter. If this memorandum becomes a pilot or contract, Greece should publish the public problem, procurement route, evaluation criteria, data rights, subcontractors, hosting location, model providers, security duties, audit access, export formats and exit plan. The city must be able to take its records, rules and operating history to another system. Otherwise the most intelligent part of the platform may be the lock on the door. Require open interfaces for service data and decision logs. Require the vendor to document every material model or rules change. Require testing before an update reaches operations. Require a rollback path that municipal staff can execute without asking the supplier for permission. And require a sunset date. A pilot should end unless published evidence justifies continuation. This is not hostility to technology. It is how public services learn without turning residents into unpaid beta testers. The Presight announcement presents the agreement as part of a wider Greece-UAE technology relationship. Other G42 companies signed memoranda covering AI infrastructure, government services and life sciences during the same official visit. That larger relationship raises the stakes. An urban pilot can become the first layer of a national platform. Decisions about data, interfaces, human authority and public evidence made now may shape later systems. So start small enough to measure. Choose one service, one geography and one problem. Publish the baseline before the first model runs. Let workers and residents help define success. Run the system beside the existing process before giving it operational authority. Compare neighborhoods, seasons and failure types. Publish the ugly cases. Then decide whether the tool earned another street. Cities do not become smart because a platform can see more data. They become more accountable when people can see what the platform did with it. Put that on the meter first.