What are Assistants
AI chatbots customized for specific tasks like drafting policy papers or answering HR questions. Set up once with prompts and documents, then reuse for consistent results. Available on Pair Chat (secure ChatGPT for government officers), similar to GPTs.
Problem
Government officers were drowning in repetitive tasks, pulling them away from higher-value work. While Assistants were intended to help automate these tasks, only 5.61% of users had created one. Our challenge: make automation so easy that officers could finally automate their own work.
My role
Sole designer (with 3 engineers, 1PM, 1 Product Operations): Led end-to-end redesign, research, data analysis, prompt engineering, QA, and in-person propagation efforts to drive AI transformation across government agencies.
Research: Understanding the Barriers
I interviewed government officers across different agencies, doctors, policy officers, teachers, to understand why Assistant creation wasn't working:
- 🔍 Poor discoverability or motivation to try: 90.8% of users had never tried any Assistant. Most didn't know the feature existed and even if they did, they lacked the incentive or time to try something new, defaulting to regular chat/old ways of working.
- 🕐 No time for complex setup: Subject matter experts lacked time for complicated development. As one doctor explained: "Only doctors know how to write the rules for the Assistant, but we have no time."
- ⚡ Couldn't find relevant tools: Even when users knew about Assistants, they couldn't quickly find the right one when needed, defaulting to regular chat instead.
Removing Three Critical Barriers
Solution 1: Showing Value First
Barrier: Users weren't motivated to try Assistants because the value wasn't apparent. New users faced empty lists with nothing relevant, leading to high drop-off rates.
The Marketplace: I created a public marketplace with pre-built Assistants targeting common pain points from user research—Meeting Minutes Writer, Email Improver, and Translation tools. I personally identified, prompted and tested these 10 Assistants. In addition, we showcased top Assistants from the user's agency, adding to more relevant Assistants immediately ready for use.
Impact: 11,300 users experienced immediate value from pre-built Assistants. This encouraged them to create more agency-specific Assistants, kicking off a flywheel of inspiration and creation that led to a thriving marketplace. The number of Assistants published to the marketplace increased 25 times.
Solution 2: Simplify the Building Process
Barrier: Officers found building Assistants difficult. Subject matter experts didn't have bandwidth for prompt engineering or extensive iteration.
Generating A Draft Prompt: I designed a system where users describe their needs in plain English, then automatically generates high-quality instructions. This allowed non-technical users to work from a solid first draft rather than starting from scratch. The system also suggests reference documents to improve Assistant quality.
I created and tested the base prompt against top use cases like document analysis and writing tasks until it worked reliably. Better initial prompts meant higher-quality Assistants overall.
Live Previews: I designed a split-screen interface where users can test and iterate on actual AI responses as they build. This eliminated guesswork and provided immediate feedback. To clearly show the real time relationship between the left and right panels, I added clear change indicators and disabled editing during testing.
Impact: These features that make creation easier has led to a 770% increase in Assistants created (3,548 to 27,455).
Solution 3: Improve Discovery at Point of Need
Barrier: Even daily Chat users were unaware of relevant Assistants that could help their work. Those who did found it cumbersome to remember or search for them at the point of need.
Intelligent Routing (WIP): The system detects when a Chat message aligns with an Assistant's expertise, it gives users the choice to use the Assistant.
By providing just-in-time nudges, users can quickly get linked up to the relevant Assistants without having to remember or search for them. This is currently in the works. We aim to complete our trials with one agency by the end of the year.
Creating a Self-sustaining Assistant Ecosystem
Tools to give creators full ownership of their Assistants
- Creator Dashboard: Assistant builders can monitor usage statistics and receive direct feedback from users, including access to chat logs when users opt in to share them with creators.
- Collaboration Tools: Multiple people can work on the same Assistant, and ownership can be transferred
Community Building: We conducted one-on-one sessions, workshops and hackathons helping officers transform workflows into Assistants. The biggest breakthrough came when peers showcased their own creations. Seeing colleagues solve shared problems was far more inspiring than our external demonstrations. This peer-to-peer inspiration culminated in schools organizing their own Assistant hackathons, inspiring other schools to host their own events.
Real-World Applications
SGH Doctors: Helps SGH's Anaesthesiology team search surgical antibiotic guidelines , reducing a 200-page manual search to a 5-minute review task, freeing 90% more time for patient care.
Teachers: 150 teachers save time drafting and vetting end of year reports for each student, with 1,300 uses for brainstorming quizzes and exporting Kahoot-ready formats.
Hospital front line staff: Assists Singapore General Hospital staff in providing quick answers regarding patient arrivals and emergencies, used over 5600 times to streamline daily operations and speed up work by up to five times.
17,293 users
using Assistants monthly (103% increase from 8500)29,510 Assistants
createdKey Learnings
Transforming government AI adoption requires turning domain experts into builders. The sequence mattered: first demonstrate value to create interest, then simplify creation for accessibility, improve discovery for continued use, and finally enable peer-to-peer inspiration where colleagues showcasing their own solutions drives organic spread.
The most meaningful outcome wasn't the usage metrics, it was witnessing teachers teaching their colleagues to build classroom management tools, or watching policy officers from different ministries collaborate on procurement Assistants. This validated our core hypothesis that empowering domain experts creates more valuable and sustainable AI solutions than top-down development.
