Live verification inside a production AI pipeline, under 2 minutes, for six months
A human verification layer for AI extracted fields on construction bills, invoices and site notes, running 24/7 until the model no longer needed it.
The challenge
Client A is an AI native construction accounting platform used by more than 1,000 contractors in the US. Its models extract financial data from vendor bills, invoices, receipts, delivery notes and handwritten site notes, which vary wildly in format and quality.
Extraction alone made errors in cost codes, line items, vendor names and amounts. Uncorrected, those errors would flow straight into job costing and billing, so Client A needed humans inside its live pipeline, around the clock.
What we did
A dedicated team was integrated directly into Client A’s ML pipeline and Label Studio instance, across three shifts with weekend and public holiday cover.
When the AI created a task, our team was notified in Slack and responded within two minutes, verifying extracted fields, correcting errors and completing anything the model missed.
Results
Response times stayed under two minutes for the whole six months, with hundreds of tasks a day at peak.
Client A’s model reached the accuracy it needed for production and no longer required the human layer, which was the goal from the start.
“DATACLAP has been great to work with and was very flexible in accommodating our labeling needs. We worked with them to get response times very low so that we could incorporate their HITL team into our live production service.”
