The AI trick that only works if your data is half-decent

One of the most useful ways finance teams can use AI is also one of the least glamorous. Not to write strategy papers. Not to create a beautiful dashboard. Not to “transform the finance function” by Friday afternoon, because apparently every problem now needs a transformation label before anyone is allowed to fix it. A much more practical use is this: Use AI to review messy operational data and ask:   What should a sensible person check before they trust this?   That sounds small, but it matters. A lot of finance reporting pain starts in ordinary files that nobody quite owns. Branch sales exports. Product margin reports. Labour rosters. Inventory adjustments. Cost centre dumps from systems that were meant to be replaced three years ago. The file is useful enough to run the business, but not clean enough to trust without checking it carefully. That is where AI can help finance teams today, provided the job is kept narrow.

Use AI to Build a Better Exception List

The mistake is asking AI to “analyse the business” from a spreadsheet and expecting magic. A better approach is to ask AI to help build a clear exception list. For example, you could upload a CSV or Excel export and ask:

Review this branch sales file. Find missing fields, duplicate rows, suspicious dates, margin outliers and anything that should be checked before this goes into a management report.

That is a very different request from:

Tell me what is happening in sales.

The first request gives AI a control job. The second invites it to sound confident before the data deserves confidence. And finance teams already have enough confident nonsense to deal with. Usually in meetings.

Why Data Quality Matters Before Reporting

A pattern I often see in growing businesses is that reporting becomes more polished before the underlying data discipline catches up. The board pack looks better. Power BI looks cleaner. The monthly commentary sounds sharper. But underneath the reporting layer are manual adjustments, inconsistent product names, old branch codes, sales cut-off issues, duplicate rows, missing fields and costs sitting in the wrong bucket because the chart of accounts has become a museum of past decisions. AI will not fix that ownership problem on its own. It will not automatically know whether “Sydney Metro” and “SYD-M” should be treated as the same branch unless someone gives it the business context. It will not replace a proper data model, approval process, master data owner or month-end review. But it can make the mess visible faster.

A Practical AI Workflow for Finance Teams

Here is a simple workflow a finance manager could use before sending a report to the CFO. Start by uploading the export and the basic business rules. Not the whole ERP. Not every table the company has ever created. Just the relevant rules. For example:

• Valid branch names
• Product categories
• Expected date range
• Normal gross margin range
• Known exclusions
• Required fields that must not be blank
• Any specific business rules for the report

Then ask AI to produce three things:

1. A data quality check
2. An exception list
3. A set of commercial questions

A useful prompt might be:

Here are the valid branch names, product groups and expected margin ranges. Check this export against those rules. Show me the top exceptions and explain why each one matters commercially.

That last phrase matters:   Why each one matters commercially.   Finance teams do not need more lists for the sake of lists. They need to know which issues could distort margin, revenue, cash, labour productivity, stock availability, pricing decisions or the story being told to the leadership team.

Separate Data Issues from Business Questions

One useful trick is to make AI separate data problems from business performance questions. For example:

Split your findings into two sections: data issues that need correction, and business performance questions that need a manager’s explanation.

This stops the review becoming a random pile of observations. A missing product code is not the same as a genuine margin collapse. A duplicate row needs fixing. A sudden fall in margin may need a conversation with operations, sales, procurement or whoever owns the decision that caused it. That distinction is important. Finance teams should not treat every exception as an error. Some exceptions are real business signals. The job is to work out which is which.

Use AI as a Review Assistant, Not the Final Approver

Another good habit is to challenge the first AI response. Ask:

What could be wrong with your interpretation? Which findings might be caused by timing, coding, master data issues or one-off operational changes?

This pulls AI back into the role it should usually play in finance:   A thinking assistant, not the person signing off the numbers.   AI can help prepare better questions. It can highlight unusual patterns. It can draft review checklists. It can identify areas that need a human explanation. But it should not be treated as the final owner of the result. The finance team still needs to understand the business, validate the data and decide what matters.

AI Can Also Review Reports, Dashboards and Management Packs

This approach is not limited to CSV or Excel files. It can also work with Power BI screenshots, PDF reports, board packs or management reporting exports. For example, you could give AI a report and ask:

Identify the five questions a CFO should ask before accepting this result.

That can be genuinely useful, especially where finance teams are reviewing sales, margin, labour, inventory, working capital or cost performance. If the business uses governed semantic models, approved Microsoft Copilot tools or controlled internal AI environments, even better. Sensitive financial and operational data should stay inside approved data access controls, not be thrown into random tools because someone saw a productivity video on LinkedIn and lost all judgement.

Where Finance Leaders Should Start With AI

For sceptical finance leaders, this is where I would start. Not with an AI transformation program. Not with a giant agent that promises to run month-end. Not with a slide deck about the future of finance that somehow takes longer to create than fixing the actual issue. Start with one recurring file that already causes rework. Use AI to:

• Find data quality exceptions
• Highlight unusual movements
• Draft review questions
• Create a repeatable checklist for next month
• Identify where ownership is unclear

If it saves the team thirty minutes and catches one material issue earlier, that is useful. If it exposes that nobody owns the branch mapping table, that is useful too. Because the trick is not pretending AI makes messy data clean. The trick is using AI to show where the mess is, what it might affect, and who needs to make a decision before the numbers become a story. For finance and operations teams in growing businesses, that is a very practical place to start.