Insight

AI in established work processes:

A critical perspective

Every organization has a growing list of AI use cases.
But research suggests the real impact comes from somewhere else: redesigning how work gets done.

Don't ask if AI can do the step faster. Ask if the step should exist.

Organizations keep finding new AI use cases: summaries, reports, customer communication, research, forecasting. The list grows every month. But are these use cases solving the right problem?

Faster tasks, same way of working

Many use cases are useful. But most of them speed up single steps in processes nobody questions. The result is acceleration, not redesign.

The research points the same way. In McKinsey’s State of AI 2025, redesigning workflows showed the strongest link to measurable impact from generative AI. Yet only about 21 % of organizations using it said they had fundamentally redesigned their workflows. A follow-up survey in July 2026 confirmed the pattern persisted. Around 79 percent of organizations continued to skip workflow redesign entirely.

The contradiction is remarkable. It speaks for itself: Organizations can execute individual tasks faster. But the underlying work practice remains untouched. And the research says this is precisely where the impact lives.

Example: the customer proposal

A typical proposal follows a familiar path: research, review of past projects, team briefing, storyline, draft, circulation, numbers check, leadership comments, revisions, delivery.

With AI, single steps get faster: research, meeting notes, text, formatting. But the workflow stays the same: same people, same handovers, same approvals, same sequence.

It’s like fitting a high-performance engine into a system built for far lower capacity.

Start from the goal instead

What if you started with the real goal: the best possible answer to the customer’s problem? Then work backwards:

Internal knowledge is always available instead of researched by hand.
External research runs in parallel and gets questioned critically.
Hypotheses are tested before anyone commits to them.
Experts spend their time on the customer’s actual problem.
Quality checks happen continuously, not only at the end.
Leaders ask “Is the recommendation sound?” instead of editing slides.

The final presentation becomes a byproduct of a better way of working.

The same question across functions

Customer service: Not “Can AI answer faster?” but “Why did the customer have to contact us at all?” Can problems be anticipated and solved before the customer notices?
Management reporting: Not “Can AI write the monthly report faster?” but “Why wait for month-end if processes can be monitored continuously and escalated when needed?”
Talent acquisition: Not “Can AI write better job ads?” but “If AI changes the work, are we hiring for the right roles?”

The harder questions that come first

Workflow redesign is fundamentally more disruptive than isolated use-case optimization. It forces organizations to confront questions they often prefer to avoid:

Which activities should disappear entirely? Not everything should be automated. Some activities should simply stop existing.
Which decisions can AI make without repeated approval? If every execution requires sign-off, most of the benefit evaporates.
Where is human judgment genuinely valuable? In ambiguity. Empathy. Responsibility. Problem definition. Ethics. Trade-offs. Creativity. Context understanding.
Where should humans actually be in the process? This requires intentional design, not default inclusion.

These questions have no comfortable answers. But they are where real transformation begins.

Human AI collaboration: A design problem, not an assumption

“Human in the loop” is often seen as automatically safe. Research is more nuanced. A meta-analysis in Nature Human Behaviour by Vaccaro, Almaatouq and Malone covered 106 experiments. On average, human-AI teams beat humans alone, but not the better of human or AI working independently.

So the question is not “How much human, how much AI?” but which setup fits the task:

Humans decide, AI provides information.
AI decides, humans check exceptions.
Humans and AI work together.
AI executes alone.
Humans do the work themselves to build expertise.

The goal is the best overall system, not maximum automation or maximum human involvement.

Redefining Sucess Metrics

This leads to a complete reset of how success is measured:

Not: Does AI save eight minutes writing an email?
But: Did the overall five-day process compress to two days?

Not: How many AI-generated documents were created?
But: Did a decision that took three weeks become a three-day decision?

Not: How many prompts were used?
But: How did rework change?

Not: How many licenses were purchased?
But: How did customer outcomes improve?

Not: How many hours were saved?
But: What became possible with the freed capacity?

Three takeaways

Question the process before you speed it up. Faster steps don’t equal better work.
Design the human role on purpose. Choose the right setup for each task.
Measure outcomes, not activity. Cycle time, rework and customer results show real impact.

Once enough workflows are redesigned, roles, decision rights and skills change with them. At that point, it is no longer only workflow redesign. It is organizational transformation.

Author’s note: The observations in this article draw on recurring patterns from our work with organizations navigating AI transformation, combined with publicly available research. Client situations have been generalized and anonymized; no individual organization or engagement is being described.

Sources: McKinsey State of AI 2025 and Operating Model Advantage 2026; Vaccaro, Almaatouq & Malone, Nature Human Behaviour, “When combinations of humans and AI are useful.” 