A listed company's 70-page annual report, prepared, designed and checked by one system
AI supports the drafting in the company's own voice, code carries every figure, and its people sign off every section. The report reached the auditor 14 weeks after kickoff.
& Accounts70 pages · designed by the system
Results
- 4 weeksfinance team time back
- 1–2 weeksexecutive time back
- £4–7kdesign costs saved
- 5 daysahead of the deadline
Time and money saved are the client's own estimates, against the previous year's manual process.
Project goals
The goals were agreed at the start of the project. The seventh is for next year.
- Less drafting and formatting workDelivered
- Drafts in the company's own voiceDelivered
- Every figure locked to the sourceDelivered
- People sign off every sectionDelivered
- Secure setup, no model trained on client dataDelivered
- Print-ready design, built by the systemDelivered
- A system that learns every yearNext year
Method
In the company's own voice
Four years of filed reports taught the system the structure, the rules and the house voice of every section. Drafts arrive with every open point marked, and each edit the client makes becomes a rule for next year.
Code carries the numbers
Code reads every figure from the finance team's files and adds every total again on every build. Not one figure is written by AI.
Every change is checked again
Late changes are rebuilt the same day. The finished PDF is compared with the signed-off file word for word, all 2,795 text blocks, with nothing lost.
Under the hood
An AI agent works a fixed tool chain in Python.
Learn moreShow less
- 01
Sources
Four years of filed reports, this year's figures from the finance team, and the signed-off Word file.
- 02
Reporting agent
An AI agent built on Claude, in a contracted setup where client data is never used for training. It supports the drafting, merges feedback and runs the tools.
- 03
Python tool chain
One combined file for markup, a three-way merge of returned markups, every total added up again, a word-for-word check, and the build of the designed PDF.
- 04
Outputs
The designed 70-page PDF and an editable Word file.
What the AI does
Probabilistic: it proposes, people decide.
- Drafting narrative from last year's text
- Sorting feedback: new, done, or conflicting
- Suggesting wording and flagging gaps
- A second reading as independent QA
What the code does
Deterministic: same input, same output, every time.
- Reading and carrying every figure
- Adding up every total, every build
- Identifying which version came back
- Laying out pages, tables and signatures
- Comparing the PDF word for word with the signed-off file
The system is built to run again every year. Next year starts from this year's signed-off file, its rules and its house voice. A direct link to the company's finance data can follow, so figures arrive without a file in between.
In the client's words
The week before sign-off used to be hell. Every late change meant going through the whole report again by hand, and something always slipped through. This year the system checked everything, I haven't seen a single thing that was wrong so far, and it took so much stress off us.
Andreas and the team at Chim took the pain out of producing our annual report! [...] Andreas' proactive approach allowed us to process many significant late changes while maintaining a high level of confidence and control over the outputs.
Film
95 seconds, with voiceover.
Your annual report
The method fits listed companies, PLC or AG, that publish a statutory annual report. We build it once and roll it forward every year, and it can extend to the reports around it.
- FY2026Built and signed
- FY2027Next
- Interim accountsCandidate
- Subsidiary accountsCandidate
Running the reporting of a listed company, or advising those who do? We build this with your finance team, or with you for your clients.
AI for language.Code for numbers.People for decisions.
Let’s build yours.

