01
From source material to a reviewable workpaper
In my research work at Foresight, I designed and developed a multi-agent workflow for document review, financial-statement cleaning and workpaper preparation. A number needs its source, period and unit before it can support further analysis.
- 01Disclosure and reference
- 02Ledger and definition checks
- 03Excel workpaper and analysis
02
Open a row to inspect the source and calculation
The example below uses an entirely fictional restaurant business and invented figures. It does not represent Yum China or findings from my internship.
Open any row to check what the figure represents.
01Revenue£120m
Example disclosure A · FY2025
Fictional annual report, illustrative page 12. Unit: GBP millions. Period: full year 2025.
02Operating profit£18m
Example disclosure A · FY2025
Same period, currency and consolidated scope as revenue. This is operating profit, not net income.
03Operating margin15.0%
18 ÷ 120 × 100
Derived from the two rows above. Align units, periods and company scope before calculating a ratio.
03
Four AI research questions
Alongside company research, I have written reports on the following topics. These questions describe the scope of that work.
| Topic | Question |
|---|---|
| Capabilities and white-collar automation | Which tasks can a model complete, and which steps still require people? |
| Consumer AI monetisation | How does product usage become demand that people will pay for? |
| Historical technology transitions | How do adoption, competition and investment shape economic returns? |
| AI for science | Who verifies a generated candidate, and what does verification cost? |
04
Keeping the process when evaluating agents
When examining deployed customer-support agents, I look at the conversation over multiple turns: what the user asked, what the system did, whether the issue was resolved and whether a person still had to take over. Those questions distinguish producing a reply from completing the task.