Artificial intelligence (AI) can be an amazing tool for improving operations. But agencies are rightly concerned about how best to implement it.
Getting value out of any new technology requires expertise from multiple domains, said Siddhartha Chaturvedi, Director of Product Management at Microsoft. You need experts in the technology itself and experts in the data, processes and goals of the agency, working together. “We bring our expertise, which is technology, and work with the [agency] experts, who know their ecosystem,” he said.
An Innovation Model for Government
While Facebook’s original motto, “Move fast and break things,” has ruled technology development in the private sector, that’s never worked for digital transformation in government, where regulations, security, and the public mission require more caution and less breakage.
Faced with challenging new tech, agencies may have the opposite reaction: to freeze and fall behind. Focusing on the potential risks, even if they are unlikely to occur, can prevent adoption of new and useful tools. That’s playing out now with AI.
“It’s true, there are edge case scenarios that we can’t anticipate today,” Chaturvedi said. “The ideal is to test it out in safe environments.”
Get Real — Safely
To address concerns about what could happen, AI adopters need to step away from the hypothetical, Chaturvedi observed. “We need to get out of the theoretical into the real world and try it out. Experimentation is key.”
Safe experimentation means setting up a “sandbox,” a testing environment where technology can be put through the paces and revised as needed before being deployed in the production environment.
Sandboxes aren’t the only guardrails that help keep AI safe. “Like every other technology that’s come before, there are things that it’s good at, and there are things it’s not. And then putting boundary conditions in place that allow us to control certain things.”
Generative AI is based on large language models (LLMs) that use massive amounts of data drawn from public information to learn to recognize patterns and generate statistically likely outputs, the training data doesn’t include data specific to a corporate or government user.
The “boundary conditions” guide outputs by adding context — essentially, giving the AI access to the data you want it to use and specifying how you want the AI to use it. For example, prompting is a method that refines queries to the AI. Grounding provides appropriate contextual guidance on how the AI answers and the data from which it draws its responses. “It’s looking only where you want it to look,” Chaturvedi said.
That means AI responses tailored to an agency’s priorities.
Where to Start
We’re still learning what AI can do for government, according to Chaturvedi, but some relatively easy first steps for agencies have emerged.
Document and information formatting. Every agency has its own styles and forms. AI can replicate those patterns while incorporating new information. “You can tell Microsoft 365 Copilot, ‘Write an email that sounds like me.’ Or it can format new information into standardized documents.” It can also summarize and generate content into existing formats.
Code. “[You may] have legacy code and we need to modernize it, we need to optimize it, but [it] might have been written in a language that’s now become redundant,” Chaturvedi said. “How do you modernize it without upskilling people and incurring the cost?” That’s something AI can do. “And then it helps you to write better code, which adheres to the requirements and standards you’ve laid out for your agency,” he said.
Chaturvedi also noted that as with written communications, every organization has its own way of writing code. AI can help new hires adhere to those standards and learn them over time.
How Microsoft Can Help
The idea of shared journey informs Microsoft’s approach to working with government. “We are here to help you get started [with AI],” Chaturvedi said. ““We want to be a partner as this technology evolves.”
The company offers co-innovation workshops, “where we sit down across the [table] and talk to you about what [AI] is doing in other ecosystems to inspire you, sharing some of the patterns we see emerging.”
Chaturvedi and other Microsoft experts have written a “Decoding AI” blog series that provides basic information, and the company hosts “office hours” for answering questions.
