The hidden cost of AI slop inside your business operations

This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.

Recently, I’ve been seeing agentic AI use cases that remind me of the RPA (Robotic Process Automation) heyday. For example, a team had built a risk scoring agent for the client onboarding process to accelerate their KYC and AML checks. However, they’d built it in isolation from the Operations team and with a focus on speed to delivery. The result was that this agent wasn’t accessing the data it needed to produce a reliable score. Not only that, but it wasn’t producing a confidence threshold, which meant that it was impossible to see whether the agent output was accurate, and making human-in-the-loop reviews much harder. This agent wasn’t fit for purpose and needed extensive reworking in order to pass testing.

In fact, as AI slop rises, this example is far from uncommon. Merriam-Webster made it their 2025 word of the year, defining it as “low-quality digital content produced in quantity by AI”. The primary reason behind the AI slop is the pressure being applied from the business boardrooms to deploy agents quickly to cut costs and accelerate operational time frames. It’s a speed, cost, quality trade-off. 

We’re currently seeing a focus on speed in the market, which is understandable both from a competitive advantage perspective and for delivering cost savings in this financial year. However, in practice, this corner-cutting results in agents that are designed and built in silo from existing operations. They don’t have access to the right systems, don’t have built-in confidence scores to trigger human-in-the-loop, and, ultimately, when it reaches the operations team for testing, it fails immediately as it’s not fit for purpose. In these scenarios, time or money isn’t saved, and quality is sacrificed.

The problem is visible across the wider market. SSON’s 2026 State of the Shared Services and Outsourcing Industry report found that agentic AI has become the top investment priority for shared services organisations, cited by 65% of respondents, well ahead of traditional RPA at 40%. The appetite is clearly there. The returns are lagging behind it, though, which is why SSON describes 2026 as the year of operationalisation rather than experimentation. As the report puts it, if 2025 was the year of too many tools, 2026 is the year of integration, with multi-agent orchestration as the route to joining everything up. The money is going in, but slop is getting in the way of the return.

Random acts of agentic AI

The flashback to RPA’s heyday is hard to shake wherever you look. At the peak of the RPA hype cycle, plenty of organisations got stuck in a proof-of-concept loop. One banking operation ended up with over 180 RPA proofs of concept it couldn’t get out of testing, because the solutions were designed and built in isolation. They weren’t meeting a need, adding value, or built with existing operations in mind. The industry came to call this “random acts of automation”. The worry now is that the same failure is playing out again, this time as random acts of agentic AI. The only way over the hurdle is to design, build and connect agents within existing operational workflows.

Where AI orchestration comes in

This is where AI orchestration is key to achieving success with agents. Instead of standalone agents operating in isolation, with AI orchestration, all your agents connect to operational workflows through an orchestration layer. That layer routes each piece of work to whichever worker (be it humans or agents) is right for each process step. It coordinates multiple process steps running in parallel, and it provides automatic human-in-the-loop based on agentic confidence thresholds. The point of the orchestration layer is that it reveals and holds the full picture. It tracks which agent is handling what, where each piece of work has got to, and when a person needs to step in. That’s what turns a set of separate agents into an operation that runs end to end.

3 ways to stop the slop

Stopping the slop comes down to 3 key components that set you up for success with AI orchestration.

1. Establish a ground truth for your agents that’s rooted in both regulatory requirements and your standard operating procedures and business rules, to avoid the risk of hallucinations. Agents need something firm to check themselves against, and without it they’ll confidently invent an answer that looks right and isn’t.
2. Store and provide your agents with access to the context from prior interactions. Anyone who’s dealt with an agent that keeps asking the same question knows the frustration, so getting this right improves the experience for both your customers and your team.
3. Build in an agentic second pair of eyes to check every agent output, along with the matching confidence thresholds, against your ground truth and context before it executes any action. That’s what keeps slop from ever reaching a customer.

Go back to the risk scoring example described earlier, and these 3 components combined would have set the agent up for success. The ground truth would keep it working in line with regulatory and company operating procedures. The context and data would let it genuinely accelerate time to resolution. And the automated second pair of eyes would flag any low-confidence output and bring a human into the loop whenever needed, keeping the quality of the agent’s work high.

That’s just one agent operating in isolation. Multiply it across all the agents a business is currently working on, and the risk quickly multiplies.

Ten questions to stop the slop

With all this in mind, here are 10 questions to ask as you design, build and launch agents to stop the slop. If you can’t answer them, treat it as a red flag that you’re delivering random acts of agentic AI, and go back to the drawing board.

• Does this agent solve a real pain point?
• Is it connected to existing operations, not built in isolation?
• Does it only access the data it needs?
• Does it have a defined ground truth?
• Does its output have a named owner?
• Can your human workforce see and intercept it when needed?
• Does an uncertain answer look different from a confident one?
• Does a second pair of eyes check its output, with a clear trigger?
• Would you know if it was wrong, and would it avoid repeating the mistake?
• Are handoffs between agents clearly coordinated?

More from our Opinions section:

• When it comes to keeping up with AI-powered updates, we need to move from human speed to machine speed
• Ransomware is a data theft business now. Most defences still guard the wrong door.
• AI agents can make changes. They cannot own the consequences.
• The EU AI Act could give creative AI a trust upgrade