|
Do you start from first principles, standards, or an overall strategy? It has to grow organically. You need baseline standards - like data handling. What data should go into AI, and what shouldn’t. What roles should AI handle, and what it shouldn’t handle. Some things should always stay human. Strategies matter for big projects. But encouraging staff to look at their everyday work and ask how AI could make it easier often reveals low-hanging fruit. Smaller projects don’t always need heavy strategy, but clear standards really help. For many people, AI just means ChatGPT. How do we look beyond that? Often the conversation stops at integrating Copilot into Microsoft 365. I find that boring. I like looking at what others are doing and asking how we could adapt those ideas here. I was flying into Dubai recently and they had a video saying, “This is our future”, and it was flying cars and AI drones for law enforcement, all by 2050. Here, most cities take a very cautious approach, slowly rolling out Microsoft 365. But we should be looking at how other cities and the private sector use AI in big ways, then easing communities into accepting those kinds of ideas. For example, cameras that detect if someone falls and automatically log an incident. Or virtual community members you can test ideas with before public engagement. If a policy is terrible, the virtual community would flag it before release. That’s similar to synthetic data in marketing - building a model of our customers or community to test ideas with. I’ve actually built a system like this. Using thousands of survey responses, you can build AI personas. You can test ideas with them, and they keep learning as new data comes in. From my experiments, it works really well. Does it make it harder because councils and government often store data in lots of different places that don’t talk to each other? You need a central solution. Many governments have five or six survey platforms, all working separately. Ideally you have one main approach, or systems that integrate well. Data must be accessible - you need to be able to extract and reuse it. If you can’t find one system that does everything, then strong API integration plus manual CSV imports is the next best thing. Is there a risk that automation reduces insight compared to manual review? If you have thousands of data points, manual review gives you very little insight. And if you only have 10 or 12 responses, that’s not enough to build anything useful. Data-driven personas need thousands of data points. What can these personas help with? They flag big risks. Is this policy likely to cause harm? Is it controversial, and how risky is that? You don’t have to follow AI advice exactly - but if it predicts real harm, you should rethink or reword.
|
