Keep the human, leave the bottleneck

Nearform built an AI agent for a state-owned financial services provider that cleared 847 anti-money laundering cases in under two hours, matching human expert analysts with 100 percent accuracy, work that would have taken the team close to a year and would have cost approximately €2 million a year to handle by hiring more staff. Nearform said the human stays in the loop by signing off every case at the end rather than monitoring each step, and cited McKinsey's 2026 State of AI survey finding only around two in ten organizations have scaled AI agents across the business, with the share seeing bottom-line impact stuck at 37 percent, plus Gartner's prediction that more than 40 percent of agentic AI projects will be scrapped by the end of 2027.

A state-owned financial services provider had an anti-money laundering bottleneck https://nearform.com/work/state-owned-service-provider/ and enlisted Nearform to solve what seemed to be an insurmountable challenge: each flagged case was taking analysts 15 to 25 minutes to clear manually, meaning a single batch of cases could tie up the team for months. How could they keep pace? The immediate internal thought was to hire more people, but to handle the volume of cases, it would have cost approximately €2 million a year. Instead, we found a better way to tackle the problem - by building an agent to take over the end-to-end investigation process. The agent analysed each flagged name by cross-checking identities, ruling matches in or out, and providing recommendations with full reasoning for final human judgement and decision-making. In a single run, it cleared 847 cases in under two hours. This number of cases would have taken the team close to a year to complete. It matched human expert analysts with 100 percent accuracy and cleared the backlog without increasing headcount. What matters for AI governance is that the human never leaves the loop, and in this case they still sign off every case. We just brought them in at the end, rather than wading through the messy process to get there. Is human-in-the-loop the right answer? Ask most enterprises how they govern their agents, and a human-in-the-loop is the logical answer. But dragging a person through the whole process to babysit the agent is an admission that you don’t trust the technology you built. No rule says their judgement beats the agent’s, and plenty of the time it doesn’t. The smart fix is to build a system you trust from the ground up. That’s what we did on the anti-money laundering work, where the human still owns the final call once the agent has done the job. But what’s needed to keep an AI workflow safe? In this case, we ensured everything was baked into the agent before it touched a single case. The industry regulator had already defined what a good decision looks like, so every result the agent produced could be measured against an established standard. Agents aren’t predictable and tend to take a different route each time, so there’s little point in policing every step. But when the outcome is set, you can make the agent testable and check its work against that. The harder part was unlocking the knowledge held in people’s heads. A lot of anti-money laundering work runs on pattern recognition analysts build up over years and never write down. We dug it out, organised it, and wrote it into the system so the agent carried it out too. The way it was built from day one is what made the solution stand up - not someone keeping watch along the way. The governance gap most won’t close The reality is that most organisations won’t do that governance work, and the numbers back it up. McKinsey’s 2026 State of AI survey found only around two in ten have managed to scale AI agents across the business https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai , and the share seeing any bottom-line impact from AI hasn’t shifted since last year, stuck at 37 percent even as more pile in. Gartner https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 expects more than 40 percent of agentic projects to be scrapped by the end of 2027, killed by murky value and weak controls. This won’t be due to a technology problem, but the lack of building robust controls around it. A clever demo gets waved through, while the dull machinery that would have made it safe is overlooked. Reserve the human for the hard cases By now, the question isn’t whether you need a human, but when you need one. Tying them up in work a machine can do is a waste. The smart move is to free them up for the cases that need a person. Aviation worked this out long ago. A modern plane lands itself in fog that no pilot could see through with their own eyes alone. The system handles the approach and lowers the wheels, while the crew sits and watches. On a standard run, they do nothing, and it lands safely. But the second something unforeseen happens, like a bird strike or a landing gear failure, their human judgement is everything. Autopilot didn’t put pilots out of work, it moved their role to the part of the process that genuinely needs a human. That’s the split you want with an agent. In the anti-money laundering build, an analyst still signs off every case, but only once the investigation is complete, the assumptions that lived in their heads are in the system, and they can see exactly how the agent reached each recommendation. Their job is final judgement, not the slog in the middle. There is a place for the human-in-the-loop, but organisations need to be clear about what that role is. If the intent is measured, the context is captured, and the person at the end still owns the decisions, you’ve got a governance model. Anything less is a security blanket without the security. At Nearform, we build agentic systems with governance baked in from the start, not bolted onto a person at the end. If you’d like to see what that looks like across your own workflows, talk to us https://nearform.com/contact/ . But wait - there's more. Nearform publishes real-world learnings on data & AI, engineering, and digital strategy - with more merged in weekly. Insights Perspectives on AI in engineering, product development, and strategy, for enterprise executives. Community Deep dives and tutorials by engineers, for engineers.