Sending an invoice doesn’t end the work of collecting payment. Customers may still ask for another copy or a statement of account, leaving finance staff to take on extra work before collection can move forward, which ends up consuming time and can hold up payment when replies are delayed or missed.
Nupur Mittal, co-founder and COO of OCTA
, a company that builds agents for accounting and bookkeeping firms, saw this firsthand while working with the finance team at Careem, a ride-hailing and delivery platform. While developing OCTA, she visited the team each week to watch how they handled invoicing, customer requests, and payment reconciliation. Seeing where their time went was crucial for her to see which specific tasks could be automated, building her approach to building AI around the work accountants actually do.
Realizing the automation opportunities firsthand
Drawing on her experience in payments and B2B, Nupur Mittal began looking into the problems OCTA would tackle through conversations with founders and finance professionals. Those discussions kept bringing her back to the work of getting paid. Even after recording revenue, teams were spending time chasing invoices, checking which customers had paid, and sorting out partial payments.
The problem isn’t exclusive to the businesses she spoke with. In a2025 Intuit QuickBooks survey
, 47% of U.S. small businesses reported having some invoices overdue by more than 30 days, showing how widespread this problem can be.
For Nupur, understanding what held up a collection meant spending time with the people handling it. She personally recruited design partners and visited their finance teams each week. One was Careem, where Nupur had previously served as Senior Strategy Manager and Business Owner for Careem Pay. The company later became a paid OCTA customer, but those early visits gave her a chance to sit alongside the team and see what handling invoices and collections involved.
“The finance teams showed me how they were switching systems, how they were logging in here and doing it,” says Nupur. “Those few months of sitting personally with the clients to understand their pain points and just experiencing it, if they say it takes five hours, literally seeing that it’s taking five hours, opens up a lot of things.”
Sitting with the teams, Nupur could watch the extra logins and movements between systems that a conversation about collection delays might leave out. She also visited every week, which was crucial to establishing trust, as the teams could see she was taking the time to properly get their work before trying to change it.
The recurring tasks OCTA could take on
At Careem, Nupur’s analysis revealed a recurring reason for delayed payments: customers were asking for another copy of an invoice or a statement of account showing their invoices over several months. Responding meant more work for the finance team, and delays in sending those documents could hold up payment.
“Most of the time,” Nupur says, “what we analyzed in the data was that people were not paying because they ended up asking questions again: ‘Can you send me a copy of the invoice? Can you send me the full statement of account?’ Basically, the full list of invoices from the past few months to now.”
Those recurring requests gave OCTA’s team a clear place to start.
The team built automation to respond with the relevant documents, taking a repeated administrative task off the finance team’s hands and helping customers get the information they needed to pay. That work became part of a receivables workflow connected to Oracle Fusion, the enterprise software Careem was using for its financial processes. OCTA helped automate invoice creation, customer follow-ups and replies, and matching incoming payments to invoices.
As a result, these tasks addressed the work involved in moving an invoice from creation through to payment. OCTA’sCareem case study
reports that the implementation saved more than 110 hours per month and reduced days sales outstanding, a measure of how long it takes to collect payment, by 24%. For Careem’s finance team, the benefits also extended to how they spent their time.
“OCTA transformed Careem’s AR process from a bottleneck into a well-oiled machine. By automating follow-ups and optimizing collections, we cut through inefficiencies, improved cash flow, and put our focus back on what truly matters—growth,” says Padma Rajagopalan, Ex-Director of Strategic Finance at Careem.
Keeping humans in the loop
While implementing OCTA helped streamline Careem’s internal workflows, complex cases and missing data in Oracle Fusion still needed the finance team’s involvement. For Nupur, designing useful automation required making sure the system properly reflected how routine work and those exceptions fit together, so the system could bring in an accountant when needed.
OCTA tests certain complex tasks with accountants, using their evaluations to determine how much autonomy to give the program. That turns firsthand knowledge of the work into more specific and context-reliant decisions about what the system can execute on its own.
“The client has the option to add an approval layer,” she says. “We’ve made sure internally that we do rigorous testing for every new feature and every new task given to AI. If the confidence is below 90%, that’s where human intervention comes in. We raise it as an alert and stop that work: ‘For this, the human needs to get involved or execute it.’”
Nupur says customers often need help deciding which work to give OCTA, and testing specific tasks makes that conversation practical. The aim is to keep routine work moving while directing people’s attention to the cases that need their expertise.
Expanding to accounting firms
Nupur encountered similar manual work inside OCTA when a bookkeeping agency sent spreadsheets and asked for documents, surprising her with how much administration even a small startup required. That experience made her realize another opportunity in front of her: helping the firms doing that work for multiple businesses.
Customer conversations pointed in the same direction. Accounts receivable customers wanted help with vendor payments and accounts payable. Businesses would consult with their accountants for advice on software, and accounting firms began asking whether OCTA could take on more of their bookkeeping. Tasks Nupur had observed within individual finance teams were repeated across an accounting firm’s client base.
That feedback also led Nupur to rethink how OCTA approached the market. In March, the company was selling an AI finance product to mid-market businesses in the Gulf region. By May, she had rebuilt its U.S. approach around a service-led offer for small accounting firms, changing how OCTA positioned itself in the market.
Nupur also designed a 90-day customer journey to set expectations for the service. Her role extended beyond deciding which market to pursue: she worked through how to introduce the offer and help firms understand what to expect as they began using it.
Making that offer work also meant understanding what could discourage accountants from adopting it. For those managing an already full workload, the time needed to learn a new system is part of the buying decision.
“Accounting professionals also want a product that’s easy to use because, with the amount of work they have, they don’t have time to learn something new. If I tell them, ‘This is a new tool, but you’ll need two weeks to learn it,’ they’ll be like, ‘No, leave it. I’m happy with how I’m doing it,’” Nupur Mittal says.
That makes ease of use central to OCTA’s value. Automation needs to reduce the work on an accountant’s desk without creating a lengthy learning exercise. Accountants already know their clients and can draw on those relationships to resolve unusual cases. Nupur Mittal’s customer conversations help identify where OCTA can take over repetitive administration so that expertise can be put to better use.
The accuracy problem in front of everyone
Accounting is less forgiving of error than most places AI has been applied. A drafting assistant that is occasionally wrong is still useful. A reconciliation agent that is occasionally wrong creates work instead of removing it, and the person who has to find the error is the accountant who was supposed to be saved the time.
Nupur Mittal says that objection comes up in most conversations with firms, and that she treats it as reasonable rather than as resistance to overcome. OCTA’s answer is testing rather than reassurance. New tasks are evaluated with accountants before the system is allowed to run them unsupervised, and the confidence threshold determines what happens next.
Some customers verify independently, which Nupur encourages. One finance director in the region hired an outside AI consultant to run OCTA against a general-purpose model on his own data before committing to implementation.
For Nupur Mittal, those questions about accuracy belong alongside the ones that first took her into customers’ finance teams: what takes up their time, and what would genuinely make the work easier?
Watching those tasks firsthand helps OCTA identify what to automate; testing with accountants helps determine whether it is ready for everyday use. The aim is to give accountants less work to chase and correct, and more time for the decisions that need their expertise.
This story was distributed as a release by Jon Stojan under HackerNoon’s Business Blogging Program.
