Why AI High Performers Redesign Workflows Instead of Just Automating Them
Most companies start AI adoption with a reasonable question:
Where can we add AI to the work we already do?
- Find a repetitive task.
- Add a copilot.
- Automate a step.
- Deploy an agent.
- Measure the time saved.
That can create value.
But McKinsey’s 2026 State of AI survey suggests the organizations getting the strongest results are doing something more ambitious.
Nearly three-quarters of AI high performers say they are fundamentally redesigning workflows because of AI. Among everyone else? Only about one-quarter.
That is a big gap.
And it raises a more interesting question:
What if the real value of AI is not making existing work faster—but reconsidering how the work should happen in the first place?
A faster bad workflow is still a bad workflow
Imagine an approval process with twelve steps.
Someone gathers data. Someone else prepares a report. Another person checks it. A manager reviews it. Then come the emails. And more emails.
Now add AI.
The data arrives faster. The report gets drafted automatically. The emails write themselves.
Wonderful.
There are still twelve steps.
That is the difference between automation and transformation.
Automation makes the existing process faster. Transformation asks whether the existing process still makes sense.
AI can absolutely make inefficient work more efficient. It can also make unnecessary work happen at impressive speed.
High performers appear to start somewhere else
McKinsey defines AI high performers as organizations attributing at least 5% of EBIT to AI while also reporting significant value. They make up only about 6% of respondents.
What separates them is not just heavier AI use. They are more likely to:
- redesign workflows
- pursue growth and innovation alongside efficiency
- measure AI impact
- manage costs
- involve senior leadership
- use human-in-the-loop structures
- actively manage AI risk
That combination matters.
They are not treating AI as a pile of disconnected tools. They are changing how the organization works.
Start with the decision, not the model
This is where I think the standard AI sequence gets flipped.
Instead of asking: Where can we insert an AI agent?
Ask: What decision is this workflow actually trying to produce?
Then work backward:
- What data does the decision require?
- Which steps actually add value?
- Which steps exist because of historical process baggage?
- What needs human judgment?
- What requires approval?
- What should trigger escalation?
Only after those questions are answered should the organization ask: Where does AI help?
A useful sequence looks more like:
↓
Data quality
↓
Value-adding steps
↓
Remove unnecessary friction
↓
Approval and escalation boundaries
↓
AI capability
That is a very different starting point.
Re-architect before you automate
This has become one of the principles I find most useful in enterprise AI:
Re-architect before you automate.
Because automation tends to amplify whatever is already there.
A poor workflow becomes a faster poor workflow. An unclear decision becomes an automated unclear decision. Bad data moves through the company faster. Weak governance becomes harder to contain.
Speed is valuable. But only if the thing being accelerated deserves to exist.
The productivity gain has to go somewhere
McKinsey found that 80% of respondents report individual productivity improvements from AI, while only 37% report positive EBIT contribution at the enterprise level.
That gap makes more sense when you think about workflow design.
Suppose AI turns a five-hour task into a two-hour task. Great. What happens to the other three hours?
If nothing else changes, the gain may simply disappear into the working day.
But redesign the workflow, and that capacity can potentially be redirected toward:
- higher throughput
- faster customer response
- fewer handoffs
- less software spend
- more analysis
- faster revenue generation
The productivity gain did not change. The management of the gain did.
Agents raise the stakes
This becomes even more important with agentic AI.
A copilot helps a person perform a task. An agent may execute several steps on its own.
That means workflow design now has to answer questions like:
- What can the agent access?
- What can it change?
- How much can it spend?
- What requires approval?
- When must it stop?
- When does a human take over?
At that point, governance cannot live in a policy document sitting somewhere on SharePoint. It has to exist inside the workflow.
For example:
↓
Validation
↓
AI analysis
↓
Rule + risk check
↓
Threshold evaluation
↓
Human review if needed
↓
Action
↓
Outcome tracking
The AI is part of the workflow. It is not the workflow.
This is where Decision Systems fit
At Decision Systems AI, I have been approaching this problem by defining the decision architecture before the model.
That means making explicit:
- rules
- thresholds
- ownership
- risk
- cost
- confidence
- escalation
- human review
Then AI can be inserted where it genuinely adds value.
The architecture becomes:
The goal is not simply to automate more work. It is to make the workflow more intelligent about what should happen next.
The question leaders should ask
McKinsey’s high performers scale more AI than their peers. But their advantage appears to come from combining scale with redesign, measurement, governance, and leadership involvement.
That suggests the real constraint may not be access to AI. It may be the organization’s ability to change around it.
So when evaluating an existing process, I think leaders should ask:
If we were designing this workflow today—with AI available from the beginning—would we build it the same way?
If the answer is no, automating the old workflow may be the wrong starting point.
The organizations that capture the most value from AI may not be the ones that automate the most work. They may be the ones that become best at deciding how the work itself should change.