One month is a lie
Ask an agent what you spend on electricity and it'll usually reach for the most convenient window: last month. In April, a Texas electric bill might be $285. In August, the same house pulls $600 or more with the AC running around the clock. Multiply April by 12 and you've budgeted about $3,400 for a line that costs closer to $5,000.
Gas runs the other direction. Summer bills of $45 look adorable right up until February lands at $268.
The number the agent gave you wasn't wrong. Last month really was $285. The projection built on top of it is what's wrong, and unless you hand the agent a methodology, it defaults to the naive one.
The fix: 12 months, annualized
The fix is specific and a little boring, which is exactly what you want in a budget. For every utility, pull a full year of history instead of a month.
Three steps. First, ask for the last 12 months of merchant history for each utility you pay. Second, total it and divide by 12: that's the annualized monthly average, and it's the only monthly number that's safe to project from. Third, compare the highest month to the lowest. If the max is at least 2x the min, mark that bill seasonal and never quote a single month for it again.
Why 12 and not 6 or 24? Twelve captures each season exactly once. Six double-counts whichever half of the year you're standing in. Twenty-four works too, but it weights two-year-old rates the same as current ones, and utility rates mostly move up.
The prompt, ready to lift
Here's the whole methodology as one prompt. Paste it into Claude, ChatGPT, Cursor, or any MCP host with your bank connected:
For each utility I pay (electric, gas, water, trash, internet), pull the last 12 months of merchant history. For each one, report the 12-month total, the annualized monthly average (total divided by 12), the highest month, and the lowest month. Flag any bill where the highest month is at least 2x the lowest and label it seasonal. Use the annualized average as the budget line, never last month.
Under the hood, the agent will typically call get_recurring_charges once to find your utilities, then get_merchant_history for each one. Five or six tool calls, a few seconds, and you get a table you can drop straight into a budget.
Why 2x is the right threshold
Every bill wiggles. Water moves with how often you ran the sprinklers, electric moves with a heat wave, and none of that noise matters much for budgeting. A highest-to-lowest ratio of 2 or more isn't noise. It's structure.
Run the check on real accounts and the pattern is clean. Internet sits at a ratio near 1.0: flat, project from any month you like. Electric in a hot climate comes in around 2.2. Gas in a place with real winters can hit 5x or more.
The threshold also tells you where you can relax. Anything under 2x, last month's bill is a fine stand-in. Anything over, you need the average. That split is the whole point of the flag.
A worked example with real-shaped numbers
Take a Texas house. Electricity peaks around $612 in late summer and bottoms out near $278 in early spring, with an annualized average of $412. Project from a March bill of $285 and your yearly electric budget comes out around $3,420. The real figure is roughly $4,950. You're short about $1,500 on one line.
Gas inverts. August is $48, February is $268, and the average is about $98. Annualize an August bill and you've budgeted $576 for a $1,175 year.
Notice the errors point in different directions depending on when you ask. Ask in spring and electric looks cheap. Ask in late summer and gas does. That's the tell that one month can't be trusted: the answer changes based on the calendar, and a budget shouldn't.
It's not just utilities
The same trap hides everywhere spending has a seasonal shape. December gifts, summer flights, annual insurance premiums, back-to-school, the domain renewals that all land in the same month. Any of these will poison a projection built from the wrong month.
Break down my spending by category for each of the last 12 months. For every category, give me the monthly average and flag any where the biggest month is more than double the smallest. Note which categories spike in the same month every year.
Annual charges deserve special suspicion. Eleven months of zero and one month of $600 gives a ratio of infinity, and an agent projecting from a zero month will drop the expense entirely. Ask it to spread annual charges across all 12 months as a monthly reserve.
When the average itself is wrong
A trailing average assumes next year looks like last year. Sometimes it doesn't. Utility rates jump, and a 30% rate increase means the older months in your window are quietly underweighting the future.
For each utility, compare the most recent bill to the same month one year ago. If it's up more than 15%, call it out and re-estimate my annual average assuming the new rate holds going forward.
Same-month-year-ago is the honest comparison because it holds the season constant. If February gas is up 34% over last February, that's a rate change, not a cold snap, and your average needs adjusting upward.
Life changes break the average too. An EV, a pool, a new work-from-home setup. When you know the future differs from the past, say so in the prompt and let the agent model it instead of assuming continuity.
Set it up once
All of this depends on the agent seeing a real year of your transactions, not a screenshot or a number you half-remember. BankBridge is a hosted MCP server that gives Claude, ChatGPT, Cursor, Gemini, and two dozen other hosts read-only access to your accounts. Every question is fetched live from the bank; nothing is stored on our servers.
It's $5 a month per connected bank, cancel anytime. Connect once and every prompt in this guide works as written. If you want the broader pattern library, start with our guide to prompt patterns for financial questions.