Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
The first prompt only computes and ranks variances; you reconcile totals before a second prompt drafts labelled hypotheses from extract.
Get ready
Compute and rank
Reconcile the totals
Draft the hypotheses
Check and confirm
Use: You
The ledger actuals and budget by cost centre, the transaction detail extract and your materiality threshold, stated in writing.
The materiality threshold is yours to set; actuals and budget must cover the same period and scope; state any mismatch.
Use: ChatGPT
Upload the actuals and budget and have ChatGPT compute variance by cost centre and account and order those above the threshold by size.
What to paste into the tool
Task: Variance table ordered by size with cost centre, account, budget, actual and variance. Compute actual minus budget by cost centre and account. Apply the stated threshold exactly as written. Order material variances by size. Show totals so they can be checked against the source. List rows that appear in only one file.
Use: You
Reconcile the size-ordered table totals to the ledger and budget totals before any hypothesis is drafted.
Explain any gap before moving on. Check your organisation's rules for restricted finance data before uploading it to any tool; this record does not establish that upload is permitted.
Use: ChatGPT
Give ChatGPT the reconciled table plus the transaction extract and ask for one labelled hypothesis per material variance, naming the record that would confirm it.
What to paste into the tool
Task: Draft hypotheses with the confirming record named, and aggregate-only flags. For each material variance propose one hypothesis labelled HYPOTHESIS, citing the transaction lines or patterns it rests on and naming the record that would confirm it. If the extract has no lines for that variance, write AGGREGATE-ONLY and give no pattern claim. State no causes.
Use: You
Open the cited lines for each hypothesis and remove any that cite none, check each draft against your facts, then take them to budget holders.
Do not state hypotheses as causes in the report; confirmed explanations come from budget holders.
Julius analyses actuals, budget and transaction extract in one workspace and drafts each labelled hypothesis from summaries. You check each.
State the rule
Rank and draft
Check and confirm
Use: You
Have ready: the ledger actuals and budget by cost centre, the transaction detail extract and your materiality threshold, stated in writing.
Check your organisation's rules for restricted finance data before uploading it to any tool; this record does not establish that upload is permitted.
Use: Julius AI
Upload all three files; Julius computes and ranks variances, shows the transaction patterns behind each, then drafts labelled hypotheses naming the confirming record.
What to paste into the tool
Task: Variances ordered by size and per-variance transaction summaries. Compute actual minus budget by cost centre and account and order those above the stated threshold by size. For each material variance list the largest transaction lines and any repeating counterparty or timing pattern. Say where no transaction lines exist. Task: Draft driver hypotheses for each material variance. Using only the variance list and the per-variance transaction summaries, write one hypothesis per material variance, labelled as a hypothesis. Cite the transaction lines it rests on and name the record that would confirm it. Where no transaction lines exist for a variance, mark it aggregate-only. Do not state a hypothesis as a cause.
Use: You
Check every hypothesis cites lines from the extract, mark coverage gaps as aggregate-only, then review with the budget holders and confirm.
Check the ordering applies your threshold as written and totals reconcile to source.
The variance numbers live in Excel formulas you can recalculate. You write control totals and reconcile before hypotheses are drafted.
Set the controls
Build the workbook
Reconcile first
Draft hypotheses
Check and confirm
Use: You
Have ready: the ledger actuals and budget by cost centre, the transaction extract and your materiality threshold in writing. Total each source file by hand.
The materiality threshold is yours to set; actuals and budget must cover the same period and scope; state any mismatch.
Use: Claude
Ask Claude to create an Excel workbook with the two sources on their own sheets and formulas for variance, threshold test and size ordering.
What to paste into the tool
Task: Excel workbook with variance formulas and size ordering. Create an Excel workbook with the actuals and budget on separate sheets and formulas for actual minus budget by cost centre and account, a material flag using the stated threshold, and a size ordering. Add a check row comparing totals.
Use: You
Reconcile the workbook variance totals to your control totals and the source files before hypotheses are drafted.
Open the formulas; compare their totals with your control totals. Check your organisation's rules for restricted finance data before uploading it to any tool; this record does not establish that upload is permitted.
Use: Claude
In a new chat, give Claude the size-ordered table and transaction extract and ask for labelled hypotheses on a new sheet, naming the confirming record.
What to paste into the tool
Task: Workbook with a hypothesis sheet naming confirming records and aggregate-only flags. For each variance in the size-ordered table write a row labelled HYPOTHESIS with the transaction lines it rests on and the record that would confirm it. Write AGGREGATE-ONLY where no lines exist. State no causes.
Use: You
Check every cited transaction line exists and matches your own facts, then review with the budget holders and confirm the threshold and hypotheses.
Do not state hypotheses as causes in the report; confirmed explanations come from budget holders.
Best when: Calculation and drafting separately checkable
What it does: ChatGPT computes then drafts; you reconcile between
Strong at: Hypotheses drafted only after you reconcile
Limit: Two uploads and a manual reconciliation
Free plan: Yes, limited (uploads and analysis capped)
Best when: One workspace for analysis and drafting
What it does: Julius analyses and drafts; you check each
Strong at: Each hypothesis names its confirming record
Limit: Plan and credit terms need confirming first
Free plan: Yes, limited (files stored only 7 days)
Best when: Variances as live, auditable formulas
What it does: Claude builds formulas; hypotheses in a new chat
Strong at: Figures are formulas, not pasted values
Limit: Generated formulas must be opened and checked
Free plan: Yes
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Claude for this? See the comparison.
Never paste credentials or full payment details.
Finance manager confirms the threshold and scope match.
Totals reconcile to source before hypotheses.
Hypotheses are labelled and name a confirming record.
Check data-handling rules before uploading restricted transaction data.
Check the result against your original inputs. Approve it before sharing or acting.