Case Study: Institutional Memory and AI Writing for a Hong Kong Think Tank
Civic Exchange is one of Hong Kong's longest-running independent think tanks, publishing on climate, energy, transport and liveability since 2000. By 2026 it held twenty-five years of research in an archive its fourteen staff could not search. Researchers were rebuilding positions the organisation had already published. Those using AI did the numbers by hand, because general chatbots dropped figures or invented them. Reports written by several authors were aligned by hand at the end. An earlier attempt by another firm to build a searchable archive had not worked: it invented information and could not tell important material from keyword matches.
12 notebooks · 2,350 documents · 4 writing assistants · 12 weeks · about USD 50 a month
The approach
AI Simply began by interviewing eleven members of staff about how they actually worked, before designing anything. Every structural decision that followed traced back to those conversations. The whole archive was then mapped and each folder assessed file by file to decide what belonged, with a sensitivity screen keeping HR, financial and board material out entirely. A town hall gave the whole team the chance to challenge the design before it was built, and staff co-designed the final category structure.
What was built
Twelve searchable notebooks holding around 2,350 curated documents, one per research area. Each answers only from Civic Exchange's own reports, and every answer carries a citation that opens the exact passage it came from. When the answer is not in the documents, it says so rather than guessing.
Four writing assistants, each set up for a different kind of output: a report writer carrying the house style extracted from the organisation's published work; a translator working from its own 390-term Chinese glossary; a slides assistant; and a communications assistant for web and social copy.
A 39-page owner's manual, so the team can run and maintain everything without outside help.
After a first look at the tools at a town hall in July, the organisation's data scientist wrote:
"I was really impressed by what you have built and how thoughtfully it addresses our day-to-day challenges."
-- Kay Huang, Data Scientist, Civic Exchange
Results
The notebooks were tested with sixty-seven questions before handover. None produced invented content when used as designed, and every citation checked led to a real supporting passage. The writing assistants refused to make things up every time they were given too little material. A live test drafted a complete funding proposal against the funder's own template, cited throughout, and showed exactly where human judgment remained essential.
It runs for around fifty US dollars a month at nonprofit pricing, with no developer, no custom code, and no data leaving the organisation.
Staff described the work as valuable, said it addressed their day-to-day challenges, and said they were looking forward to building on it. AI Simply continued supporting the team after handover.
After the funding-proposal test, the programme lead whose proposal it was wrote:
"This is really valuable work that you've done. Thanks again for building this, it's going to be super helpful."
-- Lionel Mok, Sustainable Finance Lead, Civic Exchange
Why this took twelve weeks
Most AI Simply engagements automate a single workflow and are delivered in under a month. This one was different in kind. It changed how an entire organisation reaches its own knowledge, which meant interviewing every member of staff, curating twenty-five years of files one by one, giving the whole team a say in the design, and testing thoroughly before anyone relied on it. A nonprofit software application and a regional availability problem specific to Hong Kong also had to be worked through. The pace was set by getting it right rather than getting it done.
What carries over
The software was right for this organisation at this moment, and another organisation might well choose differently. What carries over is the method: understand how people actually work before designing anything, curate rather than bulk-load, test with real questions, and hand over in a session where people use the tools on their own work. Download the full case study (PDF):
Comments