Expense entry is a good example of work that is simple in theory and tedious in practice.
For each invoice, someone needs to capture the expense name, vendor, purchase date, payment source, currency, unit cost, pre-tax and post-tax totals, classification, expense type, expense category, and CAD conversion where required. If an invoice contains multiple items, those also need to be split into separate line items.
Manually, it took about 1 minute and 45 seconds per expense.
We reduced that to 9 seconds per expense. A 91.4% reduction in handling time.
Our workflow
User invoice upload → Airtable Field Agent → JSON extraction → Airtable automation → populated line-item records → human review if flagged.
We have agents that read the uploaded invoice, parses the relevant information into structured JSON, and passes it into an Airtable automation. That automation creates the line-item records and populates the required fields automatically.
If something looks unusual, the agent flags the submission for review instead of quietly guessing.
The accuracy test
Finance isn’t a place where seemingly small errors have small consequences. To ensure accuracy, we manually reviewed 344 submissions.
Of those:
- 306 were processed without issue, which is an 89% success rate
- 38 were flagged or incorrect, and the 38 problem cases represented only 13 unique error types
That last point matters.
The system was not failing in 38 completely different ways. Most of the exceptions are repeated across a relatively small set of edge cases, which makes the workflow easier to refine over time.
More importantly, those exceptions are designed to be surfaced for human review.
The goal was never to remove humans from the process entirely. It was to stop requiring human attention for every routine transaction.
The 11% error rate was on the first pass of building this automation. Since then, we’ve refined it and cut the error rate down to ~3%.
The result
So what’s the ROI?
The build took around 36 hours and produced something we fully own, can modify as our accounting process changes, and can continue improving as new exception patterns appear.
It reduced active processing time by 1 minute and 36 seconds.
It might not seem like much, but think about how much an Accounting tool’s subscription costs month over month, and how many hours get spent per month manually keying in invoices.
For us, the value is less about how many minutes are saved; it’s more about how we can spend those extra minutes on higher-value work.
Back