Don't Automate the Workflow. Redesign It.
We need a mental re-frame as a entire industry.
Imagine a process with twelve steps.
Three approvals.
Four system handoffs.
Two spreadsheets.
An email chain.
And one employee whose job seems to consist primarily of asking everyone else whether they're finished yet.
Then AI arrives.
The company proudly makes Step 7 seventy percent faster.
Everyone celebrates.
But there are still eleven other steps.
This is one of the biggest mistakes companies are making with AI.
They're automating the business they already have instead of asking whether that business should still work the same way.
Deloitte reports that 48% of organizations have introduced AI without redesigning the workflows or roles around it, while only 12% report doing that redesign at scale. McKinsey has found workflow redesign to be one of the strongest factors associated with meaningful financial impact from AI. BCG's latest research says high performers are roughly seven times more likely to redesign workflows and reshape the business end to end.
There's a better approach.
Put one workflow on the wall
Choose a workflow tied to an important outcome.
Not “finance.”
Not “customer service.”
Something concrete:
Customer requests quote → customer receives approved quote.
Or:
Claim submitted → claim resolved.
Then map what actually happens today.
Not what the policy manual says happens.
Follow the work.
Where does information enter?
Who touches it?
Where does somebody make a decision?
Where does it wait?
Where is data re-entered?
Where does somebody leave the primary system and open Excel?
Where do exceptions occur?
Where does somebody ask another person for information that theoretically already exists?
Those are the interesting parts.
Now challenge every step
Run each step through four questions.
Eliminate: Does this need to happen at all?
AI's biggest contribution to a step may be making the step unnecessary.
Automate: Can the system reliably perform this with clear rules and appropriate controls?
Think information gathering, classification, reconciliation, routing and repetitive execution.
Augment: Should AI prepare the work while a human exercises judgment?
This is often where the highest-value applications live.
AI gathers the evidence. AI identifies anomalies. AI models alternatives.
The person decides.
Escalate: When should the machine stop and bring in a human?
Define this before deployment.
Low-confidence output.
High financial exposure.
Unusual customer circumstances.
Regulatory implications.
Material exceptions.
This is where good AI design becomes good business design.
Design backward from the outcome
PwC describes the difference well: automation improves a step; orchestration changes how the business performs. Its current guidance recommends starting with the required outcome and working backward through the signals, decisions and actions necessary to create it.
That's the exercise.
Instead of asking:
How can AI make claims processing faster?
Ask:
If we were designing claims processing today, with modern AI available from day one, what would the process look like?
Maybe twelve steps become five.
Maybe AI continuously gathers documentation instead of waiting for employees to request it.
Maybe straightforward cases flow automatically while unusual cases immediately reach experienced adjusters.
Maybe managers stop reviewing every case and instead manage exceptions.
Now you've changed the economics of the workflow.
Measure the whole process
This matters.
If AI makes one analyst twice as productive while creating another review layer downstream, you may have created zero enterprise value.
Measure from trigger to outcome.
Cycle time.
Cost per transaction.
Conversion.
Throughput.
Error rate.
Customer outcome.
Working capital.
Whatever matters.
Don't optimize the task and lose sight of the system.
AI transformation is not putting intelligence inside every step of the company we built yesterday.
It's asking:
If we could build this workflow again today, would we build it the same way?
Sometimes the most valuable automation is the step you never automate—
because you realize it shouldn't exist anymore.
