How AI Builder helped me solve a very practical Power Platform problem

Dec 30, 2025  •  5   • 975 

AI Builder is a feature within the Microsoft Power Platform that allows you to create and use AI models to automate and optimize business processes. You’ll often see impressive demos: invoices, receipts, images, predictions.

But this blog is not about a complex AI use case. It’s about a very down-to-earth, real-life problem that you may recognize.

And that’s exactly why I’m sharing it.

Context: AI-first… but still practical

At the company I work for, we have an “AI-first” strategy. Sounds great (and it is) but if I’m honest, my first instinct is often:

Can I solve this with the Power Platform techniques I already know?

Expressions. Conditions. Switch cases. Maybe a bit of regex if I’m feeling brave.

In this scenario, I tried. And quickly realized: this would become ugly, fragile, and hard to maintain.

That was the moment where AI actually became the simplest solution.

The real-life scenario

We receive lines of free-text input coming from users, emails, or uploads. Each line describes a commercial agreement or contribution and somewhere in that text, a period is mentioned.

Examples like: The scenario is a situation in which the input are line of text with multiple details in it.

• Retailer A 42204504 contribution 4% Q3 2025
• Retailer B 4224054045 promotional allowance 2% Oct-Nov 25 Region X
• Retailer C DK 047891233 bonus Nov 25
• Retailer D 404204042 anniversary bonus 25 years 2026
• Retailer D 57278287007 recycling 0.25% Jan-Feb 25
• Retailer F 40646707 partner alliance EH 2% Q4 2024
• Retailer G 4564080/456045670 online category 1% Q3 2025
• Retailer G 176104/70930118 online 1% Q1 2025
• Retailer H FR 0452424524 dynamic 3% 2024 Concept Y

If you work in finance, retail, or operations: you’ve seen data like this a lot.

The challenge

From each line, we needed to extract two clean values:

• from_date
• to_date

The desired outcome

The result examples for each lines should be like displayed in the table below.

Why classic Power Platform logic falls short

Yes, technically you could solve this with:

• Endless contains() checks
• Nested conditions
• Month mappings in variables
• Special cases for quarters, years, apostrophes, and languages

But take a step back and think about what that solution would actually look like in practice. Would the flow still be readable a few months from now? Would it be easy to extend when a new date format shows up? And when another developer opens the flow for the first time, would they understand it?

For me, the answer to all of those questions was a very clear no.

Enter AI Builder – Custom Prompts

Instead of building logic for every edge case, I used AI Builder → Run a prompt.

The idea was simple:

Let the AI interpret the text like a human would and return structured dates.

The prompt

The key was being very explicit about:

• What the input looks like
• How the AI should reason
• What the output must look like

Task
Determine the correct "from date" and "to date" based on the provided information.

Input Data (JSON Format)
{
  "description": "string" 
}

the input:
/jsoninput 

Processing Logic (Priority Order) 

1. Identify Period Indicators (if no explicit dates found) 
Look for these keywords in the description: 
"Q3'25", "okt25", "okt-nov25" 
Quarter indicators: - "Q1", "Q2", "Q3", "Q4", "1st quarter", "2nd quarter", etc. → Use quarter date ranges 
Half-year indicators: - "H1", "H2", "S1", "S2", "first half", "second half" → Use half-year date ranges 
Year indicators: - Year mentioned (e.g., "2025" or "25" )  

Additional Considerations 
- Account for leap years (February 29) 
- Use correct days per month (28/29/30/31) 
- Handle various date formats and separators 

Output Format 
{ 
"from_date": "YYYY-MM-DD", 
"to_date": "YYYY-MM-DD"
 } 

Examples
Example 1 - quarterly: 
Input: { "description": "Supermarkt 12345345 bonus 4% Q325
" }
Output: { "from_date": "2025-07-01", "to_date": "2025-09-30" } 

Example 2 - multiple months: 
Input: { "description": "Bakker 654321 extra 2% okt-nov25 Hoog " }
Output: { "from_date": "2025-10-01", "to_date": "2025-11-30"} 

Example 3 - one month: 
Input: { "description": "Slager 123789 party 1,5% nov25 Laag " } 
Output: { "from_date": "2025-11-01", "to_date": "2025-11-30" } 

Output Requirements 
- Return ONLY valid JSON 
- No explanations or additional text 
- Ensure all dates are in YYYY-MM-DD format 
- Validate before returning

Using it in Power Automate

The flow itself is very simple:

1. Run a prompt (AI Builder)
2. Pass the description text
3. Receive clean JSON
4. Parse JSON
5. Use from_date and to_date anywhere in the flow

Because the prompt is forced to return valid JSON, parsing is trivial and reliable.

Output for run a prompt:

Output for ParseJSON:

What I learned

Talking to colleagues and reading similar examples, a few patterns stand out:

• Testing matters more than prompt length: Try weird inputs early.

• Tell the AI also what not to do.

• Examples are gold: They anchor the behavior far better than abstract rules.

• This doesn’t need to be complex to be valuable: Small AI use cases often have the biggest impact.

Final thought

This is not a groundbreaking AI solution. It won’t end up in a keynote.

But it removed complexity, made the flow cleaner, and saved future maintenance time.

And honestly? Those are often the best AI wins.

If you’re working with semi-structured text and you’re forcing Power Automate to behave like a language parser, it might be time to let AI do what it’s actually good at.