Utility budgeting and forecasting from bill data

Most energy budgets are built the same lazy way: take last year's spend and add a percentage. It survives because nobody checks it closely until a variance report lands on a CFO's desk. Given that energy is the single largest operating expense in commercial office buildings, at roughly one-third of a typical operating budget, a budget built on a guess is a large exposure to carry into the year.

A defensible budget separates the two things that actually move your spend: how much energy you use and what each unit costs. Both can be forecast from data you already hold, if that data is clean.

Start with a clean usage and cost history

A useful budget model needs a consistent series of usage and cost by account, ideally 24 to 36 months so seasonality is visible. The trouble is that raw bills do not give you that cleanly. Billing periods vary in length, some reads are estimated, sites open and close, and rate changes land mid-year. Before you forecast anything, that history has to be normalized so each month is comparable to the last.

Separate the two drivers: usage and unit price

Forecast usage and price separately, then combine them. Usage is driven mainly by weather and operations, so a weather-normalized baseline plus known changes (a new tenant, a closed floor, added equipment) gets you close. Unit price comes from your tariff and supply contract, and it is trending up: EIA forecast wholesale power to average about $40 per MWh in 2025, up roughly 7 percent from 2024, driven by gas costs that rose about 56 percent year over year. A flat percentage bump hides both drivers; splitting them makes the budget explainable.

Accruals: book the cost in the month it was incurred

Accruals are where utility budgeting quietly goes wrong. You cannot simply book bills as they are paid, because billing cycles vary between utilities and rarely line up with your accounting month. Resource budgets must be aligned to the month the energy was used, not the month it was billed. That means estimating cost for partial and unbilled periods every close, then truing up when the actual bill arrives.

• Prorate any billing period that straddles your accounting month boundary
• Estimate an accrual for sites whose bill has not yet arrived, using recent usage and the current rate
• Adjust for seasonal load so a summer accrual is not built off a spring average
• True up the accrual against the actual bill and carry the difference forward

Explain variance by its root cause

When actuals miss budget, the useful question is not how much but why. A variance is only actionable once it is decomposed into its drivers, so you can tell a genuine cost problem from a warm winter. Trace each variance to usage, unit price, weather, or a contract change, at the site, region, or portfolio level, and most of the argument with stakeholders disappears.

Why this is hard by hand, and what removes the work

Every step above assumes a clean, consistent dataset across all accounts. In practice teams rebuild that dataset by hand each cycle, keying figures off PDFs into spreadsheets that drift the moment one person is away. The labor is real: benchmarking and data entry for just 20 buildings can run more than $40,000 a year, most of it spent assembling data rather than analyzing it.

MartinAI removes that assembly step. We read every bill, structure usage and cost by account and by period, flag estimated reads and gaps, and keep one clean series ready for your budget model. The forecast, the accrual, and the variance explanation all draw from the same validated data, so the numbers reconcile and you can show your work.

A practical build sequence

1. Assemble 24 to 36 months of usage and cost per account and normalize the periods
2. Weather-normalize usage to separate demand from temperature
3. Layer in known operational changes and the current tariff and contract terms
4. Project usage and unit price separately, then combine for the spend forecast
5. Set an accrual method and true it up every close
6. Report variance decomposed by driver, not as a single number