In The News
-Sathish Raman
Every quarter, in finance departments at some of the largest companies in the world, a small group of analysts sits down with a spreadsheet and decides who gets credit for what. A deal closed in one region but was sourced in another. A partner influenced the sale but did not sign it. Two teams both touched the account. Someone has to make the call, and for most organizations that someone is still a person, working line by line. More than 75% of companies manage incentive compensation on spreadsheets, and 80% of spreadsheets contain errors. The math is unforgiving. When the process that determines how salespeople get paid runs on manual judgment at scale, mistakes are not a possibility. They are a certainty.
Enterprises struggle with manual revenue attribution, leading to costly errors and lost productivity. Discover how Nithish Shetty’s human-in-the-loop AI automation approach transforms sales crediting, reducing manual work by 68% and saving hundreds of thousands. Learn why understanding processes before technology is key to successful enterprise AI deployment.
Nithish Shetty has spent 14 years building the systems that replace that spreadsheet. A Lead Business Intelligence Architect whose career spans financial services, retail, and enterprise technology, he specializes in the unglamorous middle ground where analytics meets operational reality. He set out the case for that approach in a piece titled Keeping AI-Powered BI Honest: A Human-in-the-Loop Playbook, which argues that automation earns trust only when a human stays accountable for the decisions a machine cannot safely make alone. Several years ago he was handed a problem that tested the idea directly: a revenue attribution process, running entirely by hand, that nobody was happy with.
The Hidden Tax on Getting Credit Wrong
The Sales crediting design had three layers. First, a real-time ingestion pipeline built on Incorta pulled transactional records from source systems and normalized them, so crediting decisions ran on current data rather than stale extracts. Second, the crediting logic itself: hundreds of business rules encoded as deterministic rule sets and decision trees, applied in a staged sequence so every decision could be traced back to the rule that made it. Third, an exception path. Borderline and high-risk cases were not forced through the automation. They were flagged and routed to analysts for human review.
That last part was deliberate. “Compensation is not the place for a black box,” Shetty says. “If a system decides who gets paid what, you need to be able to open it up and see exactly why. Deterministic rules give you that. A model that can’t explain itself doesn’t.”
The results mapped directly onto the three costs. Manual processing dropped 68 percent, which returned weeks of finance capacity to exception handling and higher-value review. The external consultants who had been retained just to keep the old workflow alive were no longer needed, saving roughly $150,000 a year. And every crediting decision now carried a full audit trail, so disputes could be resolved by looking at the record instead of reconstructing it from memory.
“Nobody sets out to build a process like that,” Nithish Shetty says. “It accumulates. A rule gets added, then an exception to the rule, then an exception to the exception, and 10 years later the whole thing only works because a few people remember why.”
Turning Institutional Memory Into Logic
The instinct in these situations is to buy software. The harder and more useful work happens before that. Complexity in incentive plans correlates directly with higher error rates and higher sales turnover, and the average error rate across compensation programs sits around 3%, which sounds small until it is multiplied across thousands of transactions. An automation layer built on top of rules nobody has actually written down does not reduce that error rate. It industrializes it.
So Shetty started with the rules, not the technology. Working directly with the strategy and planning team that owned the process, he mapped the full decision tree governing attribution, including the exception scenarios and compliance requirements that had never been formally documented. Only then did he architect the platform: a real-time data layer that ingested and surfaced crediting data as it arrived, with automated logic sitting on top to make routine attribution decisions the moment the underlying data landed. He owned the solution end to end, from platform architecture through workflow design and production delivery, and shipped it inside a 5-month engagement. Manual touch-points across the crediting lifecycle dropped by 68%.
“The automation was the easy half,” Shetty explains. “The hard half was getting a room full of people to agree on what the rules actually are. You cannot automate a decision nobody can articulate.”
Most Enterprise AI Stalls Before It Ever Reaches Production
This is where most projects like it die. Across 300 publicly disclosed enterprise AI deployments, 95% delivered no measurable impact on the bottom line, against $30 billion to $40 billion in spending. Only about 5% of pilots made it into production with demonstrable value. The gap was not explained by model quality or by regulation. It came down to approach: tools that could not adapt to the messy reality of an existing workflow, deployed into organizations that had never mapped the workflow in the first place.
Shetty has written about that failure pattern directly, in an article on why most enterprise AI projects never get past the pilot stage. His crediting platform avoided the trap for a specific reason: it was designed around an existing process rather than in spite of one, and it was measured against an operational outcome from day one rather than a technology milestone. The system did not have to be impressive. It had to reduce manual work, produce an auditable trail, and survive contact with the exceptions that had defeated every previous attempt at cleanup.
“A pilot that works in a demo and a system that works on a Tuesday afternoon in a real finance close are different animals,” Shetty notes. “Most AI projects are built for the demo. Then the first ugly exception shows up and the whole thing quietly gets shelved.”
Where the Machine Stops and the Human Starts
The organizations getting real value out of automated decision-making are not the ones removing people from the loop. They are the ones being deliberate about where the loop is. 72% of organizations now plan to expand incentive compensation programs into new departments, and those reviewing performance weekly rather than annually report nearly twice the significant revenue growth. Faster, more frequent oversight is the pattern that works, which requires automation to handle the volume and humans to handle the judgment.
That division is the design principle underneath Shetty’s platform. Routine attribution decisions, the overwhelming majority, are made automatically the moment data arrives. Exceptions, the genuinely ambiguous cases where a rule conflicts with another rule or where the facts do not fit a defined pattern, are escalated for human review rather than resolved by a confident guess. The result was a process the internal team could own outright after launch, with no ongoing dependency on outside resources and six figures in annual savings from eliminating that spend. The people who used to do the clerical work now do the work that actually requires them.
“An automated system that never escalates anything is not confident. It is broken,” Shetty observes. “The exceptions are where the money and the risk live. Those belong to a person, every time.”
The Unfinished Work of Enterprise Automation
The market has caught up to the idea. Intelligent process automation, the category covering systems that combine AI with workflow orchestration to handle judgment-oriented business processes, is projected to grow from $15.42 billion in 2025 to $37.54 billion by 2031. Early adopters in banking, healthcare, and manufacturing report 25% to 35% run-rate savings and cycle-time reductions of 50% to 60% after full deployment. The technology is no longer the constraint. What separates the organizations capturing those returns from the ones still running pilots is whether they did the unglamorous work of understanding their own processes first.
That is the pattern Shetty keeps returning to across financial services and retail, and it points somewhere uncomfortable for a lot of enterprises. The bottleneck in enterprise AI is rarely the model. It is the decade of undocumented institutional knowledge sitting between a business process and anyone’s ability to automate it. The work of writing that knowledge down, arguing about it, and turning it into logic a machine can follow is slow, political, and impossible to outsource. It is also the only thing that reliably separates the systems that ship from the ones that stall.
“Everyone wants to talk about the model. Almost nobody wants to sit in the room for three weeks arguing about what a rule means,” Shetty reflects. “But that room is where the project is actually won or lost. The AI is just what you do afterward.”
