Naol Basaye Is Tracking the Money That Disappears Inside Routine Operations

A company can approve its invoices, pay its suppliers, and close its books without realizing how much money has slipped through the spaces between those activities. Naol Mekuriya Bassaye has built his career around finding those losses, assigning them a dollar value, and creating technology capable of addressing them.

The losses rarely arrive as one dramatic mistake. They accumulate through duplicate payments, missed early-payment discounts, invoice-match exceptions, payment-term errors, and work completed manually because critical information sits in disconnected systems.

"Most operational losses do not announce themselves," Naol said. "They hide inside processes that appear to be functioning because the work eventually gets done. The real question is how much time and money it took to get there."

Naol is the co-founder and chief technology officer of Seeft, an operational-AI company focused initially on procure-to-pay workflows for mid-market manufacturers, distributors, and consumer packaged goods businesses. Its technology connects to enterprise resource planning data on a read-only basis, reconstructs how procurement activity moves through the company, and ranks inefficiencies by financial impact.

The problem Seeft is solving is acute in the mid-market, where companies run enterprise-grade procurement complexity on margins that leave no room to absorb duplicate payments, off-contract spend, and manual invoice exceptions, and without the analysts or software budgets that would surface them.

Naol believes many businesses have approached artificial intelligence in reverse. They start with a model, search for places to deploy it, and expect useful outcomes to follow.

"Companies keep asking what an AI system can do before they establish what the business needs it to accomplish," he said. "That produces features. It does not necessarily produce returns."

Procurement data is particularly resistant to a simple technology overlay. Purchase orders may live in an ERP platform, while invoices arrive as PDFs. Accounts payable teams maintain separate records. Employees resolve exceptions through email, spreadsheets, or undocumented workarounds.

A general-purpose model placed above that environment lacks a reliable understanding of how the business operates. It may retrieve information, but retrieval alone does not make it safe to approve payments, identify discrepancies, or initiate financial actions.

"The difficult part is creating a truthful operational picture," Naol explained. "You have to connect an observation to the underlying record, understand the process variation, and know where human approval is required. Without that foundation, confidence from the model can become a liability."

Seeft deploys agents only after the financial opportunity has been identified. The agents can draft, match, flag, and recommend actions inside controlled workflows, with approval gates designed to keep people responsible for consequential decisions.

Early work with a distributor design partner identified more than $200,000 in invoice workflow inefficiencies being addressed, along with an annual recovery opportunity exceeding $1 million.

Naol views those figures as more significant than the number of AI features a product contains.

"The unit of success should be recovered value," he said. "A finance leader should be able to see what was found, where it came from, and what happened after the system intervened."

His perspective was shaped during two years at Celonis, where he worked with process intelligence for multinational enterprises in the Fortune Global 500 across energy, aviation, automotive, pharmaceuticals, medical devices, and industrial manufacturing. The work involved reconstructing enterprise processes from system data and identifying where delays, deviations, or repeated actions created waste.

Naol says his work contributed to more than £2 million in measurable value. The experience also taught him to translate technical analysis into language that operations and finance executives could use.

"A model score means very little to the person responsible for a budget," he said. "They need to know whether a change reduces cost, releases working capital, or improves the speed of a process. Technical work becomes commercially useful when the impact is clear."

That standard is becoming more important as companies scrutinize their AI spending. Early enthusiasm encouraged organizations to fund experimental tools without consistently measuring the expense associated with each result. Large-language-model deployments can consume substantial resources through unnecessary calls, excessive context, and systems designed without cost discipline.

Naol welcomes the scrutiny.

"The market needed to move past novelty," he said. "Every token has a cost, and every automated step should justify itself. Cost-per-outcome belongs in the design conversation from the beginning."

He expects the next phase of operational AI to be shaped less by highly visible chat interfaces and more by systems embedded within repetitive, expensive processes. Procurement presents an early opportunity because the relevant failures are frequent, quantifiable, and often dispersed across departments.

Seeft plans to extend its approach into collections, inventory, and fulfillment. Each area contains its own forms of trapped information, delayed decisions, and preventable financial drag.

That roadmap runs against the grain of how enterprise AI has mostly been packaged. Since 2023, the dominant form factor has been the assistant: a chat window bolted onto an existing system, answering questions about data the user could, in principle, have looked up themselves. The appeal is obvious: chat interfaces are quick to build, they demonstrate well in a sales meeting, and they require no change to the underlying process. But they also leave the work where it started. A finance lead who asks an assistant which invoices are at risk still has to open the ERP, verify the exceptions, and decide what to do about them. Usefulness gets measured in questions answered rather than in cash recovered or hours removed.

Naol does not dismiss conversational AI. He questions whether conversation should remain the dominant way businesses judge the technology's usefulness.

"A chatbot can make a system easier to explore," he said. "The larger opportunity is software that understands enough of the process to help complete the work responsibly."

That shift requires patience with the less visible parts of product development. Data must be cleaned, events must be placed in sequence, and controls must reflect the consequences of an incorrect action. None of that creates the instant reaction generated by a polished demonstration.

It does, however, determine whether the product can survive contact with a live financial operation.

"The impressive layer depends on the disciplined layer beneath it," Naol said. "Weak groundwork allows AI to magnify confusion. Strong operational context gives it the ability to return money that the company did not know it was losing."

For Naol, that is the standard separating an experiment from a business tool. The technology does not need to appear extraordinary on a screen. It needs to make waste visible, support a safe response, and leave the company with more value than it had before.