In my last few posts, I’ve been exploring how agents and AI-assisted workflows can support everyday tasks for Excel users.
In this one, I want to focus on what is arguably the biggest risk of using generative AI for data analysis and reporting: getting the numbers wrong.
One of the oldest and most reliable principles in Excel still applies here. If something matters, you cross-check it using a different calculation method. So we are going to compare the total sales returned from our prompt with the total calculated independently using an Office Script. If those two values match, we will treat that as a validation checkpoint and allow the report to move forward.
The dataset is intentionally simple so you can follow along. Upload it to your OneDrive and keep the workbook open, since we’ll need to create an Office Script for it. After that, head over to make.powerautomate.com and we’ll build the flow together from there.
If you’re just getting started with these tools for Excel automation, I’d recommend beginning with my LinkedIn Learning courses to get a solid foundation before diving into this workflow.
If you’d like a more step-by-step walkthrough of how this flow works and why we’re structuring it this way, take a look at this post:
Creating the Office Script
In the workbook, I created a simple Summary sheet where the Sales column is totaled and that result is stored in a named cell called TotalSales.
function main(workbook: ExcelScript.Workbook): number {
const TotalSales = workbook
.getNamedItem("TotalSales")
.getRange()
.getValue() as number;
console.log("TotalSales:", TotalSales);
return TotalSales;
}
The Office Script retrieves the value from that named range, logs it for debugging, and returns it as a number to Power Automate so we can compare it against the AI-generated total as a validation step.
OK, with the workbook ready, let’s move over to Power Automate. I’m going to create a manually triggered flow for now, though you could easily switch it to run on a schedule or another trigger later. Since our data lives in Excel, there are a few extra setup steps involved before everything is wired together.
Get the workbook content
First, search for the action called “Get file content using path.” For this parameter, you simply point it to your workbook. And remember, you can rename each step so the flow is easier to read later.
Run the custom prompt
Next, choose “Run a prompt” from AI Builder”
Open the Excel workbook [Excel workbook]
Locate the table containing sales data.
The dataset contains transactional sales data with fields such as:
Date, OrderID, Salesperson, Region, Customer, Product, Category, Units, UnitPrice, and Sales.
Primary KPI:
- TotalSales = sum of the Sales column.
Contextual calculations (for internal use in commentary only):
- TransactionCount = number of rows contributing to TotalSales
- AverageSale = TotalSales / TransactionCount
- TopCategory = category contributing the highest share of TotalSales
- TopSalesperson = salesperson contributing the highest share of TotalSales
Flash Report Objective:
Return:
- TotalSales (numeric only, no formatting)
- Commentary (1–2 concise executive-style sentences)
Commentary Guidance:
- TotalSales is the anchor KPI.
- Use contextual metrics to highlight meaningful performance dynamics such as:
• Category mix (Hardware vs Electronics vs Service)
• Revenue concentration by salesperson or customer
• Volume vs pricing effects (unit-heavy vs premium items)
• Product mix influence on revenue
- Focus on business interpretation, not mechanics.
- Do NOT restate every metric mechanically.
- Do NOT mention column names, sheets, or calculation steps.
- Do NOT invent trends, time periods, comparisons, or benchmarks not present in the data.
- Neutral, confident executive tone.
Rules:
- Do not filter or modify the workbook.
- Do not fabricate data.
- Return valid JSON only.
- No text outside the JSON object.
Examples of acceptable commentary style:
- "Revenue reflects a balanced mix across Hardware and Electronics, with premium Service transactions meaningfully lifting overall performance."
- "Sales are moderately concentrated among top contributors, though revenue remains diversified across product categories."
- "Performance is supported by steady transaction volume, with higher-priced offerings contributing disproportionate revenue impact."
- "Revenue mix suggests balanced regional contribution without material dependence on a single product line."
Output format:
{
"TotalSales": number,
"Commentary": string
}You can rename this prompt to something clearer, like ‘Sales Flash Report Generator.’ Before saving it, we should test it, and to do that we need to upload a sample workbook. Go to the three dots, open Settings, and turn on the code interpreter.
Next, upload a sample workbook. This file is only for this test run. In the prompt, locate the placeholder [Excel workbook], delete it, type ‘/’, and choose Image or document. That dynamic field is where you will upload the workbook for testing. Later, when we build the flow, we will pass this in as a variable instead of uploading it manually.
Also, under the output panel on the right where it says “Model Response,” we want the result returned as JSON. There are a few setup steps to get there, but once everything is configured you should see a test result come back. Each run may vary slightly, but ideally the sales total is correct. We will validate that by comparing it to the value returned from the Office Script, and the commentary field will just provide some additional narrative context.
We are structuring the output as JSON because it makes it much easier to extract specific pieces for a downstream report. It also gives us a clean way to verify that the TotalSales figure is accurate before pushing anything forward.
Pass the file contents to the prompt
Go ahead and save the prompt. You will notice there is one more parameter we need to configure. For the purposes of the flow, we have to specify where the workbook will come from. The file we uploaded earlier was just a one-time sample for testing.
In the actual flow, the workbook should be passed in dynamically based on its location, which we can retrieve from the previous step. That said, this action expects the file content as raw base64. We can generate that using the base64() function and pass it directly into this step.
base64(body('Get_file_content_using_path'))
Run the Office Script
The prompt has now been run against our workbook, but before we report anything out, we need to confirm the numbers are correct. There are a few ways to validate this, and if you are even slightly risk averse, you should absolutely be checking these totals yourself before sharing them. Even so, this automated cross-check is a smart safeguard.
Our goal is simple: verify that the Total Sales calculated by the generative AI matches the Total Sales calculated by the Office Script. What happens next in the flow will depend on whether those two values agree. Add the step called “Run script” to your flow, then specify which Office Script to execute and which workbook it should run against.
Set up the condition
Now we want to cross-check that the Total Sales calculated by the prompt’s Code Interpreter matches the Total Sales calculated by the Office Script. If the numbers tie out, we will send the flash report. If they do not, the report will not run and instead I will send myself an email to investigate.
To set this up, add a Condition step to the flow and configure it to compare the TotalSales value from the Office Script with the TotalSales value returned by the prompt. The expressions you will need for this comparison are shown below.
outputs('Run_script')?['body/result']outputs('Run_a_prompt')?['body/responsev2/predictionOutput/structuredOutput/TotalSales']
I used to prepare my content and presentations several weeks or even months ahead. Now I don’t bother with that and only give myself a couple of weeks, because things move so quickly anyway. I don’t want to put too much effort into something that’s just going to go obsolete or change substantially.
If false, send an email
There are plenty of ways to handle this, but I am going to keep it simple and just use the “Send an email” action. You will of course want to use your own email address here, not mine. The idea is simply to send yourself a quick, hard-coded message letting you know the numbers did not match so you can take a closer look.
If you wanted to get fancy, you could include the TotalSales value from the Office Script and the value from the prompt in the email body so you can immediately see the discrepancy. That could be a helpful starting point. Personally, I assume I will need to do some old-fashioned analyst digging anyway, so I am keeping it simple for now.
If true, post to Teams
Now let’s look at the better outcome. If the two values do match, we will go ahead and post the Sales Flash to Teams.
We will post it as an Adaptive Card. An Adaptive Card is simply a structured JSON layout that Teams understands how to render as a clean, formatted message instead of plain text. Because our inputs from the prompt are already in JSON, it actually makes things easier to stay in JSON. That way we keep everything structured and consistently formatted from calculation all the way through presentation.
Of course, that means we need to define a bit more JSON to design the card itself. We will use the Adaptive Card schema framework to do this. Think of it as a template that tells Teams how to display titles, labels, and values. The full card definition is shown below.
You will also notice that I apply a small amount of formatting to the TotalSales variable so it renders as currency, for example $3,760.00 instead of 3760. That final polish makes the flash report feel much more like a proper financial summary.
OK if it’s ACTUALLY working it will go out to Teams! We’ll use adaptive card
{
"type": "AdaptiveCard",
"version": "1.5",
"body": [
{
"type": "TextBlock",
"text": "Sales Flash Report",
"weight": "Bolder",
"size": "Large"
},
{
"type": "FactSet",
"facts": [
{
"title": "Total Sales:",
"value": "@{formatNumber(float(coalesce(outputs('Run_a_prompt')?['body/responsev2/predictionOutput/structuredOutput/TotalSales'], 0)), 'C2', 'en-US')}"
}
]
},
{
"type": "TextBlock",
"text": "@{outputs('Run_a_prompt')?['body/responsev2/predictionOutput/structuredOutput/Commentary']}",
"wrap": true,
"spacing": "Medium"
}
]
}Go ahead and run the flow. Ideally, you should see something like this. If not, you will receive an email alerting you that there was a variance.
Conclusion
There is a lot more we could layer on top of this. Just because the total ties out does not mean the explanation is solid. The math could be right while the commentary is completely off. You might decide to add additional summary calculations to triangulate the results, or tighten the prompt so it is more deterministic and less free to hallucinate details that are not actually in the data.
You could also introduce a formal approval step so the report routes for signoff or rejection before it ever posts to Teams. That would make the entire process more controlled and auditable, especially in a finance setting.
I would love a clean way to directly guide or override the prompt’s commentary within the flow itself, but at least right now I am not seeing a simple, built-in way to do that.
Even so, this demonstrates that with a bit of structure and creativity, you can absolutely cross-check AI-generated reports. There is real opportunity here, and I will keep experimenting and sharing what I discover.