Saving an Interior Design Studio $69,984 a Year on AI

Most studios looking at AI costs focus on the model. Switch to a cheaper one, save money. The instinct isn't wrong, but it's usually not where the real savings are. In a recent build for a small interior design firm, we dropped an AI workflow from $5,800 a month to $0.44 a month. Same workflow. Same accuracy. Very different cost.

01The problem

Receipts landing in a shared inbox

The firm receives receipts by email in a shared inbox. For each receipt, an AI system needs to read it, identify which project it belongs to, and create a transaction in their custom project management software. The software was built to talk to AI directly, which made it a good candidate for automation.

The first version used a Gmail MCP connector, a bridge between AI and Gmail. It worked for some receipts, but MCP can't read HTML email bodies, and a lot of receipts come in that format. So we rebuilt it.

02The first design

$5,800 a month

The straightforward rebuild was a cloud worker. A small program running in the cloud that checks Gmail every 15 minutes, pulls in emails, runs them through Anthropic's flagship Opus model, classifies which are receipts, and creates transactions for those. Clean logic. Sensible architecture. Completely unaffordable.

Running that workflow in the cloud means billing at API rates instead of drawing from a subscription. With Opus processing every email in the inbox every 15 minutes, the estimated annual cost was $69,984. Close to an operator's salary.

Same workflow. Same accuracy. $5,800 a month to $0.44.

03The five changes

How we cut it to $0.44 a month

Only process new emails

The first version re-analyzed every email in the inbox on every run. We fixed that by storing a timestamp of the last processed email and telling the system to ignore anything older. That single change made the workflow about 100 times cheaper.

Filter out promotions

More than half of the inbox was promotional email. Never receipts. We added a filter to skip those before any AI processing. Another 2x reduction, putting us at roughly 200 times cheaper than the original design.

Split classification from processing

Instead of running every email through the full receipt-extraction workflow, we added a lightweight first pass: of these emails, which ones are receipts? That cut out more than half the remaining emails before the expensive processing step. Another 1.3x in savings.

Use a cheaper model, with tests

With the volume already down, we switched from Opus to Sonnet. We tested first. We took 23 examples, hand-graded the expected output, and confirmed Sonnet hit 100% on the test set before making the switch.

Run from the subscription, not the API

The final change was the biggest. Instead of running in the cloud at API rates, we scheduled the task to run on the designer's computer using Claude Desktop. That draws from the subscription the firm was already paying for. At API-equivalent rates, that's another 10x savings. In practice it's effectively free, because the subscription is a fixed expense.

04The takeaway

Architecture is the lever, not the model

The workflow now runs as a scheduled task in Cowork on the designer's computer. When the computer is on, Claude Desktop checks Gmail, compares against the last-processed timestamp, filters promotions, classifies the remaining emails, and creates transactions for receipts. Cost: $0.44 a month. The drawback is that it requires the computer to be on. For a $70,000 annual saving, that's a fine tradeoff. Worst case, a $600 Mac Mini left running covers it.

The lesson isn't specific to receipt processing. It applies to any AI workflow that runs on a schedule or at volume. The biggest savings come from thinking carefully about what the system needs to do and where waste can be designed out. Model choice matters, but it's rarely the largest lever. Architecture usually is.