How AI Sees an Entire City: Behind the Scenes of Scale AI Solutions.

When we think about AI, it is easy to imagine it solving everything with a single click. In reality, between an idea and a system that actually works, there is an entire path: breaking down the problem, understanding the customer limitations, choosing an architecture, and solving a thousand small problems that no one thought of in advance. This guide is the written summary of session 31 in the AI Master club, featuring Galit Levin, a senior manager at Deloitte and part of the Deloitte and NVIDIA Alliance. Galit brings about 15 years of experience in deep learning, following about a decade at Intel and about seven years at Mobileye developing autonomous vehicles, with patents in the AI vision field. In the session, she pulled back the curtain on what AI projects look like at the scale of entire cities and mega enterprises.

The team that Galit leads at Deloitte works in two tracks: vision worlds (image and video analysis) and agentic worlds (AI agents connected to organizational data). The separation is mostly conceptual, because in practice, agentic systems also need to understand visual data like charts and documents.

Smart cities: thousands of cameras, one brain

What the system actually detects

• Cleanliness and maintenance · identifying garbage accumulation, hazards on the sidewalk or road (even a fallen tree branch), and alerting for immediate treatment
• Crowds · identifying crowd density that can lead to violence, pickpocketing, or transportation congestion
• Graffiti and vandalism · identifying in real time, when in European cities the fine for graffiti reaches 3,000 to 5,000 dollars
• Traffic management · smart traffic light control according to actual congestion instead of pre-set times

An interesting insight from the field: tourism is not only a positive thing for a city. Major cities in Europe turned to Deloitte precisely because tourism brings with it pickpocketing, crowds, large amounts of garbage, and graffiti. A smart city is first and foremost an answer to these problems.

Cloud or Edge: the decision that shapes the entire project

Privacy is not an asterisk at the end, it is an entry condition. Municipal camera data is sensitive information: cities do not share it outside, not even with Deloitte itself. The team sends people to sit physically at the client site, to install and test everything there. Any work on the data requires face blurring, and the models are tested against the standards and laws that the municipality is obligated to.

When the client asks for one thing and needs another

The problems found only in production

• Lighting and weather · dark streets at night, rain and snow that obscure the image
• Winter clothing · people covered in hats and coats, with most of the face and body hidden
• Reflections in store windows · the system can count the same person twice, once real and once in the reflection
• Occlusions · a baby in a stroller where only a piece of a face is seen, a person behind a tree, a hand or forehead peeking out from a dense crowd
• Differences between cities · a European city looks different from an Israeli city, and local culture changes how the data looks

How do you get the client to trust the system? You cannot expect them to count thousands of people with their eyes to verify that the number is correct. The solution: the system marks a red dot on every person it detects, so you can see exactly who it counted, including difficult cases like a covered baby in a stroller. Transparency is what builds trust.

From alerts to predictions: the collected data is worth money

Zora: AI agents working in the largest organizations in the world

These systems are already in production at global clients, with thousands of logins a day. HP, one of the biggest clients, stated that 40 percent of the manual work that finance teams used to do is done today in seconds by the agentic system.

Four layers of value: from Q&A to recommendations

• Layer 1 · Question and Answer: "What was my revenue this month?" and the agent retrieves from all organizational data
• Layer 2 · Analysis and Root Cause: Not just how much, but why. Which region dropped, which market surprised, how the currency affected things
• Layer 3 · Forecast: Revenue forecast based on data, market condition, and competitor analysis
• Layer 4 · Action Recommendations: Which product is not worth continuing to sell, where you should focus next quarter

The AI Master club opens the door for you: a live meeting every week, recordings, and a community. 47 NIS per month, cancel anytime.

Good morning, here is what matters to you today

How to trust an agent that touches money

The three principles of trust

• Full transparency · every answer in the system can be broken down: from which sources it was taken, from which PDF and even from which pages, from which tables, columns, and rows, and what the logic of the agent was along the way
• Read-only permissions · the agents are on Read and not on Write: they analyze, predict, and recommend, but do not execute actions and do not change data without human approval
• External control and standards · the product undergoes controls and tests by an external party, in Certified to Sell and Certified to Distribute processes, before it is released to clients

The ability to let agents perform actions themselves (Write) is on the table, but the approach is careful: to define precisely what permissions each agent has, what it sees, and what it is allowed to change, and to release this gradually. This is an important lesson for anyone building agents for their business.

The next station: Physical AI

Frequently asked questions about AI Vision and scale AI solutions

AI Vision (or Computer Vision) is the field of AI that analyzes images and video: identifying people and objects, measuring density, identifying hazards and defects. This is a long-standing field that preceded the GenAI era, and today it integrates with language models and agents to provide complete solutions.

CityShield is a smart city system developed at Deloitte. It connects to thousands of existing cameras throughout the city, in real time and past recordings, grades each scene according to criteria defined by the municipality (cleanliness, crowds, graffiti, and more), and triggers alerts. Instead of two or three people trying to watch thousands of screens, the system watches everything and surfaces only what requires attention.

It is not always possible and not always right. A city without full internet infrastructure cannot connect cameras to the cloud, and then the calculation runs on a small chip inside the camera itself (like NVIDIA Jetson for a few hundred dollars) with small models trained for a specific problem. Even when there is cloud, cost and latency considerations dictate things. The architecture is derived from the reality of the client, not the technology.

Municipal camera data is sensitive information that does not leave the municipality, not even for the development team. The work is done physically at the client site, with mandatory face blurring on all data, and testing of the models against the standards and laws that the municipality is obligated to.

An AI agent is a smart model that also has hands: access to tools and systems that allow it to do things, not just answer. Like an octopus where each tentacle is a capability: calculator, files, organizational systems. The difference between a generic agent and an organizational agent is knowledge: ChatGPT has the internet, an organizational agent also has the internal knowledge that the company has accumulated, and that is the real differentiation.

Through three mechanisms: transparency (every answer is broken down to sources, down to the page level in the PDF and the column level in the table), read-only permissions (the agent analyzes and recommends but does not perform actions on data), and external control with Certified to Sell standards before release. These are principles that any business building agents can adopt.

The guest of the session

The bottom line of the session: AI does not solve problems with a single click. Those who succeed at scale are those who know how to break down the problem together with the client, choose an architecture that fits reality (and not a demo), build trust through transparency, and progress layer after layer. And this is true both for a municipality with thousands of cameras, and for a small business building its first agent.

Frequently asked questions

What is AI Vision anyway?

AI Vision (or Computer Vision) is the field of AI that analyzes images and video: identifying people and objects, measuring density, identifying hazards and defects. This is a long-standing field that preceded the GenAI era, and today it integrates with language models and agents to provide complete solutions.

What is CityShield and how does it work?

CityShield is a smart city system developed at Deloitte. It connects to thousands of existing cameras throughout the city, in real time and past recordings, grades each scene according to criteria defined by the municipality (cleanliness, crowds, graffiti, and more), and triggers alerts. Instead of two or three people trying to watch thousands of screens, the system watches everything and surfaces only what requires attention.

Why not run everything on large models in the cloud?

It is not always possible and not always right. A city without full internet infrastructure cannot connect cameras to the cloud, and then the calculation runs on a small chip inside the camera itself (like NVIDIA Jetson for a few hundred dollars) with small models trained for a specific problem. Even when there is cloud, cost and latency considerations dictate things. The architecture is derived from the reality of the client, not the technology.

What about the privacy of the people captured by the cameras?

Urban camera data is sensitive information that never leaves the municipality, not even for the development team. The work is done physically at the client site, with mandatory face blurring on all data, and model testing against the standards and laws that the municipality is obligated to follow.

What is the difference between an AI agent and a regular chat like ChatGPT?

An AI agent is a smart model that also has hands: access to tools and systems that allow it to execute tasks, not just answer. Like an octopus where each tentacle is a capability: calculator, files, organizational systems. The difference between a generic agent and an enterprise agent is the knowledge: ChatGPT has the internet, while an enterprise agent also has the internal knowledge the company has accumulated, and that is the real differentiator.

How can you trust an AI agent with financial data?

Through three mechanisms: transparency (every answer is broken down into sources, down to the PDF page and table column), read-only permissions (the agent analyzes and recommends but does not execute actions on the data), and external control with Certified to Sell standards before release. These are principles that any business building agents can adopt.