The claims team got a copilot in March. By June, the head of operations was staring at a productivity report that had barely moved and trying to understand why. The tool was good. The adoption was real - people used it every day. But the work had not changed. A claim still entered through the same queue, waited for the same three approvals, bounced to the same specialist for the same manual check, and left through the same bottleneck it always had. The AI made each step a little faster. The process was still the process. They had automated a task and left the workflow exactly where it was - and a faster step inside a broken workflow is a rounding error, not a return.
This is the most expensive mistake in enterprise AI, and it is almost invisible while you are making it. The tool works. People adopt it. And the value does not arrive, because value does not live in the task. It lives in the workflow - the end-to-end flow of work, decisions, and handoffs that actually produces the outcome. McKinsey's 2025 State of AI survey is unambiguous on this point: out of 25 attributes tested, workflow redesign had the biggest effect on an organisation's ability to see EBIT impact from generative AI. Not model quality. Not adoption rates. Redesign. The 94% layer AI onto the work they already do. The 6% redesign the work first, then deploy AI into the workflow they have rebuilt.
A faster step inside a broken workflow is a rounding error, not a return.
Quick answer
AI pays back when leaders redesign the workflow, not when they simply add a tool to the old process. The work has to be reimagined end-to-end, tasks must be redistributed between people and AI, and the return must be measured against the original workflow baseline.
Why adoption is not enough
Morgan Stanley shows what the redesign - not the tool - actually buys you. The firm did not hand its wealth advisers a chatbot and move on. It rebuilt the knowledge workflow around the AI: the way advisers find, trust, and use the firm's vast research library. The result was adoption above 98% in the adviser teams, and the share of the knowledge base that advisers could actually reach in the flow of work rising from roughly 20% to 80%. The number that matters is not the adoption rate. It is the 20-to-80 shift - because that is a workflow being redesigned, not a task being sped up.
Klarna shows both the promise and the caution in one story. Its AI customer-service agent handled 2.3 million conversations, took on around two-thirds of the company's customer chats, and cut average resolution time from eleven minutes to under two. Spectacular - and a genuine workflow redesign, not a bolt-on. But Klarna later moved back toward human access for customer service, acknowledging that speed alone was not enough where empathy, nuance, and trust mattered. That is not a failure of the technology. It is the lesson written in real time: redesigning the work is not the same as removing the human. It is deciding, deliberately, what the human does now - and where the human must stay in the loop.
The 3R sequence
That decision is a discipline, and it has a shape. At Green Everest we redesign workflows through a sequence we call the 3R: Reimagine the work from first principles, with no legacy assumptions; Redistribute the tasks between human and machine; and calculate the Return against the workflow you started with. Reimagine is where the operators' expertise becomes the raw material - they know how the work really happens - even as the output challenges the process they have mastered. Redistribute is where you decide, task by task, what the AI does and what the human does. Return is where you prove it was worth doing.
The human-AI loop
Running through every redesigned workflow is the loop that governs the handoff between human judgment and machine capability. It has six moves: Delegate the task to the AI; let it Generate the output; the human Reviews it; the human Overrides where judgment demands; the workflow Escalates the cases that exceed the agent's boundary; and the system Learns from every intervention so the next pass is better.
Designed well, the loop builds trust progressively. The human starts by reviewing everything and widens the agent's autonomy only as the evidence earns it. Designed badly, it does one of two things: it erodes trust, because the AI makes calls nobody can verify; or it wastes trust, because the human rubber-stamps everything and the review becomes a reflex. High acceptance rates can mean the AI is excellent or that oversight has gone slack - and only the quality of the overrides tells you which.
POPIA makes this structural
In South Africa this loop is not only good practice; it is a compliance design issue. Section 71 of POPIA protects data subjects from decisions with legal or substantial consequences that are based solely on automated processing of personal information, subject to defined exceptions and safeguards. The human-in-the-loop is therefore not a decorative design preference. Organisations that design review, override, and escalation properly are addressing compliance structurally, while those that automate without it may be exposed from day one.
Three questions for leaders
For the leader whose productivity report has not moved, the questions are sharper than "is the tool working?"
-
Have we redesigned the workflow?
Take the highest-volume workflow. Has it been rebuilt end-to-end around AI, or have the same steps been made faster while the bottleneck stays in place?
-
Can we point to the human judgment design?
For every place AI now makes or shapes a decision, where are review, override, and escalation built into the work?
-
What does the human do next?
When AI enters the workflow, have we deliberately changed the role around it, or are we hoping the headcount question answers itself?
The tool is the easy part. The workflow is the work.
The teams that win the next phase of AI will be the ones willing to take one workflow apart and put it back together around what humans and machines now each do best.
The teams that win the next phase of AI will not be the ones with the best tools. They will be the ones willing to take a workflow apart and put it back together around what humans and machines now each do best. The tool is the easy part. The workflow is the work.
The workflow is where AI ambition becomes business value.
FAQ
Why does AI adoption fail to improve productivity?
AI adoption fails when the tool is added to the old process without changing handoffs, decision rights, roles, measurement, or governance. The task may get faster, but the workflow bottleneck remains.
What is workflow redesign in AI?
Workflow redesign means rebuilding the end-to-end flow of work around what AI can automate or augment, what people should still own, and how the result will be measured against the original baseline.
What is a human-AI loop?
A human-AI loop is the operating design that lets AI generate outputs while people review, override, escalate, and teach the system through structured feedback.
How does POPIA affect AI workflow design?
POPIA section 71 makes solely automated decisions with legal or substantial effects a serious compliance question. South African teams should design meaningful review, override, escalation, and transparency into AI-enabled workflows.
Which single workflow, redesigned end-to-end, would change a number on your board report? The AIO Build engagement redesigns one value stream over 90 days - the 3R sequence, the human-AI loop, and your first AI Lead embedded in the work. Start with one workflow. Prove it. Then scale.
Sources: McKinsey, The State of AI: How Organizations Are Rewiring to Capture Value; OpenAI, Morgan Stanley AI case study; Morgan Stanley, AI @ Morgan Stanley Debrief launch; Klarna AI assistant announcement; CX Dive on Klarna's human-service shift; Protection of Personal Information Act, section 71.