top of page

Find the Utility Bills That Look Wrong: Anomalies and a Winter Gas Forecast with the Copilot Analyst Agent

7 hours ago
5 min read

Utility bills are the most ignored numbers in property management. They arrive every month, AP pays them, and nobody looks closely unless one is enormous. But estimated reads, catch-up bills, leaks and stuck equipment all hide in those bills, and by the time someone notices, you've paid for months of it.

The problem isn't the data. It's that checking it properly means adjusting for the weather. A gas bill that doubles in January is normal. A gas bill that's 36% higher than January should be is not, and you can't see that difference by scanning a spreadsheet.

In this guide we give three years of utility bills to Analyst, the data analysis agent built into Microsoft 365 Copilot, and ask two questions: which bills look wrong, and what will gas cost us this winter? Two prompts, about five minutes, and every number checked.

Here it is, sped up:

The Analyst agent in Microsoft 365 Copilot finding anomalies in three years of utility bills and forecasting winter gas costs

What you'll need

  • A Microsoft 365 Copilot licence. Analyst is in the Agents list in Copilot Chat. We pinned it so it's one click away.

  • Your utility bills in Excel, one row per bill: property, utility, billing period, usage, unit, amount and whether the read was actual or estimated.

  • Monthly heating and cooling degree days for your city. Environment Canada publishes them, and your energy consultant or utility portal probably has them too.

Step 1: Put the bills in one table

Our workbook has 324 bills from October 2023 to September 2026: electricity, natural gas and water for Harbourfront Office Tower, Northgate Distribution Centre and Lakeview Medical Plaza. A second sheet holds Toronto's monthly heating and cooling degree days.

That second sheet is the important part. Without it, any analysis will tell you that January is expensive, which you already knew.

Step 2: Ask the two questions

In Analyst, select the plus sign, upload the workbook and ask:

Analyst agent prompt with the utility bills workbook attached
This workbook has three years of utility bills (electricity, natural gas, water) for our three properties, plus monthly heating and cooling degree days for Toronto. 1) Find any bills that look wrong or unusual. Normalize gas for heating degree days and electricity for cooling degree days, so cold or hot months are not flagged by mistake. For each anomaly give the property, utility, month, how far it is from expected (in units and dollars) and the most likely cause. 2) Forecast natural gas usage and cost for each property for November 2026 to March 2027, assuming a normal winter, and show the method.

Notice the second sentence. Telling Analyst how to normalize stops it from flagging every cold month as a problem.

Step 3: Read the anomalies

Analyst worked through it in 11 reasoning steps, created a results workbook, and came back with seven anomalies:

Table of seven utility bill anomalies with variance from expected, dollar impact and likely cause

We had planted three real problems in the data, and it found all three, with no false alarms:

  • Harbourfront electricity, March to May 2025: two estimated reads that were 231,311 kWh too low, followed by a catch-up bill 248,464 kWh too high. Analyst noticed that the two numbers nearly cancel out and called it what it is: a true-up after estimated reads, not a change in how the building uses power.

  • Lakeview water, July to September 2025: 862, 807 and 457 m³ above a normal summer, about $11,300 in total. Three months in a row points to a physical problem such as a leak or an irrigation controller left running, not a billing error.

  • Northgate gas, January 2026: 17,251 m³ more than the weather explains, worth $8,132.91. Analyst gave a checklist rather than a guess: meter accuracy, boiler controls, door schedules and temporary loads.

Just as useful is what it didn't flag. January to March 2026 was colder than the previous winters, and gas bills were higher everywhere. Lakeview uses more water every summer for irrigation. Analyst flagged neither, which is exactly what weather normalization is for.

Step 4: The winter gas forecast

For the forecast it worked out a normal winter from the average degree days in the workbook, fitted gas use against heating degree days for each property, and used a robust regression so the Northgate spike didn't distort the result:

Forecast method: normal heating degree day profile, regression on heating degree days and robust regression

We checked the usage forecast against our own regression. All three properties were within 1%: 322,235 m³ for the portfolio from November to March.

Where it slipped

The usage was right, but the cost was not. Analyst's first answer put the winter gas budget at $145,576. To get there, it modelled cost from the historical trend, which mixes the old rate with the new one. A rate increase in January 2026 is right there in the bills. So we asked:

Check the cost forecast. What price per m3 and fixed monthly charge do the 2026 gas bills imply? Redo the winter cost forecast using the current rate, not the historical average, and show the difference.

This time it worked out the actual tariff from the 2026 bills, $85 a month plus $0.47 per m³, and confirmed that every 2026 bill fits the formula exactly. Then it redid the budget:

Revised winter gas cost forecast at the current rate, $152,725, compared with the original forecast

The revised budget is $152,725, which is $7,149 or 4.9% more than the first answer. We checked it by hand, and it's right to the dollar. Budget with the first number, and you'd be explaining a shortfall in April.

The lesson: when you ask Analyst for a forecast, ask what rate it used. Usage follows the weather, but cost follows the tariff, and tariffs change on a date.

A quick game: anomaly or not?

Three bills. Decide before reading ours.

  • A gas bill 80% higher than last month, in January.

  • A water bill 25% higher than the spring months, in July, at a building with a lawn.

  • An electricity bill marked "Estimated" that's 35% lower than the same month last year.

Ours: no, no, and yes, but not because of the usage. January is cold, and lawns get watered. The third one is the real problem: the meter wasn't read, and the next bill will catch up with interest. Normalize first, then judge.

Lessons from the build

  • Give it the weather. Degree days turn "this bill is high" into "this bill is higher than it should be."

  • Tell it how to normalize. One sentence in the prompt prevents a page of false alarms.

  • Check the rate, not just the model. Analyst forecast usage almost perfectly and still missed the tariff change until we asked.

  • Keep the Read Type column. "Estimated" is often the first clue that a catch-up bill is coming.

Ideas to take it further

  • Run it every quarter: add the new bills, ask the same two questions, and compare with last quarter's answer.

  • Feed the forecast into your budget: paste the revised table into the budget variance agent so next month's comments compare actuals to a weather-adjusted budget.

  • Chase the Northgate spike: ask Copilot to search your work orders for January 2026 at Northgate. A stuck door or heater usually leaves a trail.

  • Add a fourth utility: steam, chilled water or waste hauling all work the same way.

Why this matters

Utility data lives in the AP system, the weather lives on a government website, and the reason a bill is wrong lives in a work order or an engineer's memory. Bring your data from different systems into one place and put AI in front of it, and checking three years of bills takes five minutes instead of an afternoon nobody ever schedules.

Need help?

Smart Solutions builds Microsoft 365 Copilot, Power Platform and reporting solutions for Canadian property and facilities teams. If you'd rather have purchase orders, invoices and vendor spend in one product, take a look at ProcuraCloud, our procurement platform for small and mid-sized businesses. Contact us to talk about your utility data.


Comments


bottom of page