AI keeps pushing the automation boundary. Where each workflow settles is a design decision.

The automation line in any workflow sits where manual work ends and automated work begins. That line has always existed. Spreadsheets moved it. ERPs moved it again.

AI is moving it further than either, and faster, into complex work that crosses system boundaries and requires interpretation.

The line does not land in the same place for every workflow. Where it settles is a design decision.

Automation is a spectrum, not a switch

The standard view treats automation as a destination. Either you have automated a process or you have not. The conversation jumps from 'we need a human for this' to 'AI can do this now', as though the switch flips once and stays.

In practice, it is a spectrum, and AI is pushing workflows further along it.

AI can now handle use cases that previously needed human involvement at every step. The shift is not about speed. AI handles greater complexity, works across systems, reasons with context and carries actions through to completion.

That is a different kind of shift. Rules-based systems handled repetitive, well-defined tasks. AI handles tasks that cross system boundaries, involve ambiguity and require interpretation of context.

AI automates more because it connects, reasons and acts

In the first article in this series, I described AI as a loudspeaker for the platform it is plugged into. The same capability that amplifies your platform also lets AI work across it.

An agent can read from a pricing system, a customer record and an order system in a single conversation. It can reason across all three and recommend an action. Then it can carry that action out.

That combination is what pushes more complex use cases past the point where they needed a human at every step.

I built a KYC/KYB onboarding platform for a global B2B business. Customer onboarding involved document collection, identity verification and regulatory checks across 100+ countries. It took four days. With automated verification and AI-powered checks, it came down to under four hours.

The workflow moved dramatically towards full automation. But a human still reviews the exceptions: documents that do not match, flagged entities, edge cases where the regulation is ambiguous.

The settling point for that workflow sits near full automation, with human oversight on the exceptions.

Different workflows, different settling points

AI moves the settling point towards full automation by automating a greater share of complex, multi-system use cases. But the settling point differs per workflow, and it is set by the business, not the model.

On the algorithmic pricing platform I built for the same business, AI integrated catalogue, master data, routing, capacity and order inputs to recommend prices across 20+ countries. Commercial teams accepted or overrode those recommendations.

The settling point sat in the middle of the spectrum: AI recommends, the human decides, the system acts. That was by design. The cost of a wrong price on a strategic account outweighed the efficiency of full automation. This is a product and business decision. Not a technology one.

Klarna learned what happens when you set one settling point for an entire function. In early 2024, the company handed roughly three-quarters of customer service conversations to AI. By May 2025, CEO Sebastian Siemiatkowski said they had 'gone too far' and 'focused too much on cost'. The result was lower quality. Klarna reversed course and moved to a hybrid model.

The lesson is not that AI failed. It is that one settling point does not fit an entire function. Routine queries sat comfortably at full automation. Complex complaints did not.

Design the roles around the decisions AI hands back

As AI takes on more of the routine, the human role does not shrink. It changes.

The people who used to handle the full workflow now concentrate on the cases AI cannot close: the override, the escalation, the judgement call. On the pricing platform, commercial teams stopped spending time on routine quotes and focused on strategic accounts where the context was too nuanced for an algorithm. On the onboarding platform, compliance analysts stopped chasing documents and focused on flagged cases that required investigation.

Aviation went through this shift decades ago. When autopilot took over routine flight, the industry did not cut pilots. It retrained them. Crew Resource Management shifted pilot training from manual flying skills to exception handling, situation awareness and team decision-making. The responsibilities of the role changed to match what automation could not do.

Design the roles around the decisions AI hands back, not the tasks it takes over.

The settling point will keep moving as AI improves. The workflows that sit near full automation today will push further. New workflows will cross the line. But the human does not disappear from the spectrum. The human moves to where the stakes are highest and the context is most nuanced.

In the first article, I argued that AI amplifies whatever platform it is plugged into. This article asks how far that amplification runs before a human needs to step in and decide. The answer is different for every workflow. Set the settling point deliberately.

The next article turns to what happens when proof is cheap. When AI can automate more of the build, the constraint moves from development effort to evaluation capacity. The bottleneck shifts.

This is the second of three articles on AI as an amplifier. The first, on why AI amplifies your platform rather than fixing it, is [here](/blog/ai-is-a-loudspeaker). The next asks what happens when the cost of proving an idea drops to near zero.