.webp)
Ottimate builds AI-powered accounts payable automation that catches overpayments, stops fraud, and helps finance teams close their books faster. Its platform digitizes invoices and pulls out the data finance teams depend on.
Automation handles clean, repeatable invoice formats well. But a real share of documents still defeats it: handwritten invoices, low-light and poor-quality scans, irregular vendor-specific layouts. Those are the cases where accurate human-labeled data matters most. It clears the work automation cannot yet do, and it becomes the ground truth that trains and validates the next generation of models.
Careerflow built the human data team behind that work. In under six months, a team that started at five people and grew to fifteen helped move Ottimate's model accuracy on complex invoices from 60% to 85%, while processing more than 50,000 invoices a month.
Ottimate's automation already performed well on standard, repeatable invoices. The hard problem was the long tail: handwritten documents, degraded and low-light scans, and irregular layouts. Many of these invoices are genuinely complicated, involving complex calculations, detailed annotations, and multi-step reasoning to read correctly. And the accuracy bar is unforgiving, because a wrong number on an invoice has real financial consequences. These cases drove model accuracy down, and more automation alone could not fix them. They needed large volumes of accurately labeled examples, plus disciplined human review to benchmark what the model produced.
Volume made it harder. Invoice flow was unpredictable, with sharp monthly spikes. Ottimate needed a partner that could do several things at once:
.webp)
The work was not simple data entry. Invoices arrive in varied and sometimes unusual layouts, so contributors had to learn Ottimate's process, tools, and quality bar. That made trainability and consistency matter as much as raw speed.
The engagement was built around a closed loop, with Careerflow as the human validation layer inside it. This was not a one-time labeling project. It was a continuous cycle that compounds:
.webp)
Each pass raises accuracy and shifts more work to automation, while human validation keeps the model honest on the cases that matter most. This is the mechanism behind the move from 60% to 85% on complex invoices.
The goal was never just to clear invoices. It was to build a workforce that produces reliable, model-ready data: accurate labels for training and disciplined review for validation, at a volume and consistency that actually moves model accuracy.
Careerflow pulled the strongest contributors from a large applicant pool through a multi-step vetting process that combined structured assessments, AI-assisted interviews, telephonic screening, and human evaluation.
Candidates were judged on more than technical aptitude. Communication, attention to detail, learning agility, and problem-solving all counted, because labeling quality on complex documents depends as much on the ability to learn and adapt as on prior knowledge.
The team's output roughly tripled as it matured, while quality held steady. Per-contributor productivity grew about 3x from go-live, and the pace stood out to Ottimate. When the team crossed 200 invoices a day, Akshay, Ottimate's trainer, noted:
"We have many people in our current team who are doing 200+ every day, but they have many years of experience compared to this team, which is just a few months old."
A core requirement was handling unpredictable volume surges without losing quality or speed. As the team matured, peak-period capacity climbed sharply.
Peak throughput rose 100.9% between Mont 1 and Mont 3. In 3rd Month alone, peak periods accounted for nearly 20,000 invoices, about 35.5% of the month's output.
The team also flexes with demand. When Ottimate needs more hands, Careerflow vets, trains, and onboards new contributors in under 24 hours, then scales back down once the spike passes, so Ottimate gets surge support without carrying a permanent surplus. During one unexpected high-volume weekend, Careerflow recruited, onboarded, and prepared ten additional contributors in under a day. Before touching live invoices, each completed training invoices, got mentorship from experienced team members, and passed a quality and readiness review, so the added capacity never came at the cost of data quality.
In under six months, Careerflow turned a data requirement into a partnership that measurably improved the product.
Model accuracy. Ottimate's internal accuracy on complex invoices improved from 60% to 85% over the engagement, driven by the team's human-labeled training and validation data. The gains came specifically on the hard cases: handwritten invoices, low-light and degraded scans, and irregular layouts that automation alone could not handle.
Operations. The team scaled from 5 to 15 dedicated contributors while holding quality steady, roughly tripled output, grew monthly volume from 41,222 to a projected 60,000, and doubled peak-demand throughput. It also became self-sustaining, training, mentoring, and running QA internally, which reduced Ottimate's dependence on client-side resources.
This engagement grew beyond staffing into a data partnership that moved a real product metric. Four things drove the outcome:
What started as a single invoice-processing engagement is now a long-term partnership. Careerflow and Ottimate are planning new work together, including new dataset types and workflows beyond invoice processing.
