Power Automate prompt miscomparing dates

(0) Share

Report

Posted on by RB-11061800-0

We are analyzing a json with ids and dates and want to identify the best candidate based on a natural language question.

JSON:

"elements": {
    "58": {
      "dataInicial": "1991-01",
      "dataFinal": "1999-07"
    },
    "655": {
      "dataInicial": "1999-08",
      "dataFinal": "2006-06"
    },
    "1419": {
      "dataInicial": "2012-01",
      "dataFinal": "2019-12"
    },
    "1692": {
      "dataInicial": "1989-07",
      "dataFinal": "1990-12"
    },
    "2938": {
      "dataInicial": "2006-07",
      "dataFinal": "2011-12"
    },
    "6691": {
      "dataInicial": "2014-11",
      "dataFinal": "2026-06"
    },
    "7060": {
      "dataInicial": "2020-01",
      "dataFinal": "2026-06"
    }
  }

We questioned which elements encapsulate the period of 2019 until today and asked for an explanation. In this case, the element 6691 should be chosen, but the agent responds "no element covers the time period" and gives the following explanation:

"6691": "Coverage from 2014-11 to 2026-06, doesn't cover from 2019-01 to 2014-10"

We tried:
- Comparing logically the inicial dates and the final dates separately and to return only when the dates in the question are between the dates in the elements (a more code based approach with intermediate variables with boolean value)

- Using a natural language approach

- Using dates in format YYYYMM and YYYY-MM

What are the best practices when handling dates and comparing numbers with prompts? Is it possible to achieve 100% accuracy in these cases?

Categories:

Calling actions from Copilot Studio

I have the same question (0)

All responses (3)

Answers (0)

Sort by

Suggested answer

sannavajjala87 282 Super User 2026 Season 1 on at

Like (1)

Report
Copy link

Link copied!

Hi,

The honest answer to your last question: no, you won't get 100% by having the model compare the dates itself. That weird explanation you got ("doesn't cover from 2019-01 to 2014-10") is the tell. The range is backwards, so the model didn't really compare anything, it just made up text that looks like reasoning. LLMs are bad at this kind of number/date math when it's mixed into a generation task, and prompt tweaking won't fully fix it.

What works is to stop making the model the calculator. Let it read the question and pull out the target range (2019-01 to today), then do the actual comparison in a Power Automate flow or Power Fx, where the rule is just dataInicial <= start AND dataFinal >= end. Do that and 6691 matches every time, while 7060 (starts 2020-01) correctly drops out.

Two quick tips: convert the dates to YYYYMM integers before comparing (201901 instead of "2019-01"), and pass in today's date as a variable instead of trusting the model to know it.

So it's very doable, just not as a pure prompt. Parse with the model, compare in code.

Was this reply helpful? Yes No
Suggested answer

11manish 2,837 on at

Like (0)

Report
Copy link

Link copied!

Do not rely on prompts alone for date comparisons, numeric calculations, or business-critical filtering. Use code, Power Automate, functions, or plugins to

perform the comparison, and use the LLM only to interpret and explain the results.

Was this reply helpful? Yes No
chiaraalina 2,318 Super User 2026 Season 1 on at

Like (0)

Report
Copy link

Link copied!

Hi

Are you using AI Builder Run a prompt. Please share your flow and your prompt. It's easier to help!

Was this reply helpful? Yes No