Audit of an accounting office
The full AI audit of a sample accounting office: how the team works, what we chose, how it is set up and how it is used every day.
The starting point
A sample accounting office, with the daily routine of many firms.
The demonstration starts from a fictional accounting office that handles tax, payroll and bookkeeping for small businesses and the self-employed, and keeps track of their paperwork.
Three jobs keep coming back: answering the same client enquiries, comparing insurance proposals page by page and reconciling invoicing with the accounts every quarter. Some people were already using ChatGPT on their own, with no shared instructions or data rules.
- Team
- Administration, the professional team and accounts.
- Tools
- Email, Excel, invoicing software, accounting software and ChatGPT, used individually.
- Goal
- Use AI for specific tasks, with clear rules and a person who reviews every result.
Five tasks assessed, three chosen
We choose up to three tasks where AI can help at a manageable risk. We don't add weak ones to make up the numbers.
| Task | Who does it | Frequency | Decision |
|---|---|---|---|
| Replies to clients' administrative enquiriesThe same questions (channel, opening hours, whether documents arrived) are answered by hand every day, and everyone words them differently. | Administration | Daily | Chosen · P01 |
| Comparing insurance proposals and document versionsComparing two long proposals page by page is slow, and it's easy to miss an exclusion. | Professional team | Weekly | Chosen · P03 |
| Quarterly check of issued invoices against the accountsMissing invoices, duplicates and mismatched amounts are hunted down by eye. | Accounts | Quarterly | Trial · P06 |
| Reminders for outstanding documentsThere isn't yet a validated to-do list per client; without it, reminders aren't reliable. | Administration | Monthly | Postponed |
| Internal questions about proceduresThe approved internal knowledge is still small. It will be reviewed once the procedures are documented. | The whole team | Weekly | Postponed |
How it is set up
A ChatGPT project dedicated to the office. The team writes short prompts; the project already knows how to work.
- Standing instructions
Project_Instructions.mdRole, sources and authority, what to do when information is missing, professional limits, human review, data and the format of replies. - Approved knowledge
Knowledge_Pack.mdThe office's facts with a version and validity: for example, that the channel for receiving documents is the Document Portal (K01 v2.0). - Workflows
Selected_Playbooks.mdHow each of the three chosen workflows runs, and what must never be done. - Reconciliation engine
p06_reconcile.pyCompares two lists row by row with verified code. The AI explains the result; it doesn't calculate from memory.
How the team asks for it
- P01Reply to this client using our information.Reviewed by: Head of Administration
- P03Compare these two policies.Reviewed by: Head of the professional team
- P06Find what's missing between these two spreadsheets.Reviewed by: Head of Administration
The step-by-step guide
Six steps to get the workspace ready, each with what to do and how to know it worked.
- Step 1 · Open
Create a new project
In the sidebar, under “Projects”, create a new project.
The dialog for creating the project opens.
- Step 2 · Name
Give it the office's name
Type the office's name, for example “Office · AI workspace”, and choose to keep its memory to this project only.
The project appears in the sidebar.
- Step 3 · Paste
Paste the instructions
Open the project settings and paste the whole of the instructions. Then save.
The whole text fits: don't shorten any of it.
- Step 4 · Add
Add the project files
Upload the approved knowledge, the workflows and the reconciliation engine to the project's sources.
Each file appears only once.
- Step 5 · Test
Run the first test
Open a new chat inside the project and run the first three tests, one per workflow.
The guide describes what you should see in each test.
- Step 6 · Check
Complete the checks
Before working with real data: complete instructions, no duplicate files, passing tests and clear reviewers.
If any test doesn't go as expected, the team doesn't work with real data.
The prompts to copy
Each prompt has a button that copies it exactly, with no headings or explanations. The first three are for testing the workspace; the rest are for everyday use.
Reply to this client: “Hi, where can I send you this quarter's invoices?”
Compare these two insurance proposals: premium, limit, excess and exclusions. Don't choose either.
Find what's missing between these two third-quarter invoice spreadsheets.
Reply to this client using our information.
What changed between v1 and v2?
Show the mismatched amounts with their rows.
What data goes in
Three clear lists so nobody has to wonder. When something arrives that shouldn't be there, the workspace stops and asks for a trimmed version.
Safe
- Synthetic or anonymised data
- Public information
- Approved internal knowledge
- Invoice exports with only number, date, amount and currency
Minimise first
- Invoices, contracts, policies and emails with names, tax IDs or addresses
- Spreadsheets and CSVs with unneeded columns or rows
- Insurance proposals: remove the policyholder's name, tax ID and address
Never upload
- Passwords and credentials
- Digital certificates
- Bank details and credentials
- Full payslips
- Bulk client lists
- Entire mailboxes
Never: send, file, pay, sign or change anything. A person always does that.
An audit for your team
Yours would be different: your tasks, your information and your rules.