Introduction
Performing a flux or preparing a leadsheet and explaining MoM variances are the most consistent thing we face under a controllership. Trade Payables, AR, Prepaids, Accruals, FA, every month someone has to explain why balances moved. Staff can rush through it, explanations get inconsistent, and reviewers end up rewriting half the file or sending it back with my favourite: review notes (these were called coaching notes back in 2010 @ PWC 🙄 )
Consider this:
Building a sub-workflow agent that is set up with very specific context and rules, built with a pre-emptive context building prompt, that layers into an execution prompt.
This gets us closer to consistent outputs. Building account-specific prompts as sub-workflows inside your orchestration of workflow agents can become a powerful business process to have if we’re thinking automation-first.
That means a Trade Payables variance prompt built for your AP working paper. An AR prompt tuned for reporting packs. A slightly different one for an audit file or board deck. The structure doesn’t change much, but the audience and context do.
Why Two Parts
We split the prompt into intake and execution because context and analysis are different steps.
Intake (framework prompt): collects account-specific rules, audience, and context. - This becomes your building block or your engine building phase.
Execution (customized prompt): tells the agent how to run the variance analysis and draft the explanation. This is the recurring step that is front-facing.
This two-step design matters because:
Each account has its own policies and drivers
Each audience (management, auditors, board) expects different levels of precision
Explanations have to run on a recurring schedule with minimal oversight
The output is not meant to be “final.” It’s a draft, reviewed by a capable party. But the draft is consistent, structured, and built off the rules you set.
The Framework Prompt
The framework prompt is built for a specific account inside a specific process. For example: Trade Payables in the AP working paper.
Framework instructions:
You are a variance explanation assistant. Before creating an execution prompt, gather context for the Trade Payables GL account in the AP working paper. Do not draft explanations until intake is complete.
Intake questions:
Which account level should the agent analyze? (parent, child, GL string)
Which company policies apply? (cutoffs, vendor payment terms, accrual rules)
Should the explanation include vendor, region, or BU detail?
Who is the audience? (internal reporting, auditors, board)
Are seasonal trends relevant? (year-end supplier payments, quarterly cycles)
Are there one-time items this period? (settlements, large vendor payments)
Should the agent recommend drill-downs?
The Execution Prompt
Once the intake is complete, the GPT builds the execution prompt.
Example: Trade Payables (AP working paper)
You are a controller reviewing Trade Payables in the AP working paper. Analyze the current variance using trial balance and historical balances.
Use the following steps, in order, to analyze trade AP for the current month as compared to prior month, and provide insights/analysis over the change using the steps provided, and context included from the framework prompt.
Context:
• Policies: cutoff rules, vendor payment terms, accrual methodology
• Include vendor detail where material
• Audience: Accounting Review (controller, senior manager, etc)
• Seasonality: year-end supplier payments
• One-time: settlement with Vendor X this month
• Drill-down: flag vendor-level analysis if exceptions are large
Instructions: identify transactional drivers, separate recurring activity from anomalies, reference policies directly, draft an explanation suitable for auditors, and recommend drill-down when needed
Goal: To get the workflow to run through each step individually, factoring as much as possible before providing an output to the user.
Why Account-Specific Matters
The same balance sheet account looks different depending on context:
In the AP working paper, Trade Payables needs cutoff rules and vendor-level detail.
In a reporting pack, the same account might focus on cash flow impact and business unit exposure.
In an audit file, the explanation has to be evidence-driven and tie back to policy references.
This is why we don’t build one master prompt. We build framework prompts for each account and each recurring use case. Once the rules are baked in, the workflow can run with limited oversight.
Where This Fits
This should be looked at as one piece, a sub-workflow inside a broader agentic orchestration for flux.
The orchestration might include:
pulling balances from ERP
calculating flux at parent and child levels
drafting variance explanations
routing drafts to reviewers
logging evidence for audit
This sub-workflow is one block in that chain. The two-part prompt makes sure the block works the same way every time.
Conclusion
Variance explanations drain time when they’re inconsistent and rushed. By building account-specific prompts in two steps: intake and execution… you create a sub-workflow that agents can run on schedule with minimal oversight. Again, the execution oversight, not the review of the output. There’s a very distinct, and important difference there.
The framework captures context. The execution applies it. The output is structured, repeatable, and ready for review.
This design works because it acknowledges the reality: Trade Payables in AP is not the same as AR in a reporting pack. Audience and context matter. Build the prompts accordingly, review the output, and let the orchestration handle the rest.
