I know a few of the leaders designing and developing Microsoft’s AI applications for the Gates Foundation.
I think you’re on the right track, and, alongside the scale of service (reaching more people and more topics with an average level of advice or recognition), there’s a second component to it: scale of analysis. The newly possible solutions that AI advances have created include more than those famous models that answer broad prompts with art, copy, or code.
They also include focused, sometimes incomprehensible tasks which can only be done at an impactful scale due to the creation of deep learning and advances in compute-inexpensive language understanding, computer vision, and audio analysis:
A network of affordable, durable, solar powered, LoRa meshed audio sensors analyzed by a model to diagnose changes in the biodiversity of the Amazon and other rainforests (via ambient bird and animal calls across thousands of species). Visual analysis done on a cheap camera network estimates herd sizes of larger, silent animals.
A model that analyzes satellite imagery to evaluate major shifts in the industrial use of land, including tracking the national development of solar farms to evaluate nations receiving new energy grants.
A social analysis bot that tracks the rapid introduction of propaganda narratives or intentional agitation by foreign state actors (Russian bot farms), including building a map of associated IPs. Sadly, the social networks basically shrugged when given this data, so Msft gave it to LEAs.
These things are being done at a scale that would be incomprehensible to an organization of people.
Scale of analysis tasks are still, IMO the smartest use of AI today, despite the fashionable trend of GPT and the promise of AGI. A few models to spark ideas:
Recognition tasks with a dictionary too deep for human experts to grok when scaled up - like identifying thousands of wildlife
Recognition tasks with a timescale too rapid or sudden for human attention - Amazon Prime Vision predicting a QB sack in a football game before it happens
Recognition tasks when human vigilance or sensitivity would miss an occasional or slight occurrence - measuring eccentricities in electrical signals, vibrations, etc. to predict the failure of industrial equipment