Perspective · Essay
The AI transformation problem is usually a workflow problem
Organizations often start AI transformation by asking where AI can be inserted. A better starting point is understanding how the work, decisions, information, and handoffs actually operate today.
Organizations understandably want to know where they should use AI.
The technology is moving quickly.
Executives see competitors announcing initiatives.
Employees are already experimenting with tools.
Vendors arrive with impressive demonstrations.
Someone creates a list of possible use cases.
The question becomes:
"Where can we put AI?"
I think that is often the wrong starting point.
The more useful question is:
"How does this work actually happen?"
Because enterprise work is rarely one task.
It is a system of people, information, decisions, tools, permissions, incentives, handoffs, exceptions, and history.
If that system is poorly understood, adding AI may simply automate one visible fragment while leaving the underlying problem intact.
Sometimes it makes the workflow worse.
Start with the work, not the technology
Imagine a process that takes three weeks.
A team identifies one task that consumes twenty minutes and automates it with AI.
The task now takes two minutes.
The process still takes three weeks.
Why?
Because the real delay may live elsewhere:
an approval waits four days,
critical information arrives late,
a specialist must interpret an exception,
two systems contain conflicting data,
a customer repeats information at every handoff,
ownership becomes unclear halfway through the process,
or a decision cannot be made until six people coordinate.
The automated task was real.
It just was not the system constraint.
This is why workflow understanding matters.
Before asking what AI should do, understand:
what triggers the work,
who participates,
what each person is trying to accomplish,
what information they need,
what decisions occur,
what systems they use,
where work waits,
where information is re-entered,
where judgment matters,
where exceptions happen,
where risk accumulates,
and what outcome the workflow is actually supposed to create.
Only then does the AI question become interesting.
A workflow is more than a process map
Traditional process mapping can become an exercise in documenting steps.
Step one.
Step two.
Step three.
That is useful, but insufficient for AI transformation.
AI changes what is possible at the level of responsibility.
So the workflow needs to reveal more than sequence.
For each meaningful activity, ask:
Who owns the decision?
What information informs it?
How predictable is the task?
How variable is the context?
What happens if the decision is wrong?
Is the action reversible?
Which systems are involved?
Does the person create value through judgment or through mechanical execution?
What information exists only in someone's head?
What causes escalation?
Which steps exist because of an old technological limitation rather than a real business requirement?
Those questions reveal where intelligence can create leverage.
They also reveal where automation may be dangerous.
Do not automate the artifact and preserve the dysfunction
Knowledge work often produces visible artifacts.
Reports.
Emails.
Presentations.
Tickets.
Forms.
Briefs.
Summaries.
Plans.
AI can generate many of these remarkably well.
But the artifact may be the residue of a workflow rather than its purpose.
If a team spends six hours producing a weekly report, the obvious AI opportunity is generating the report faster.
But why does the report exist?
Who reads it?
What decision does it enable?
Which parts matter?
Could the underlying information reach the decision-maker directly?
Could the system identify only the exceptions requiring attention?
Could the report disappear entirely?
Automating the artifact can preserve an obsolete workflow.
Transformation asks whether the artifact is still necessary.
Find the decision points
One of the most useful ways to analyze a workflow is to identify its consequential decisions.
What gets prioritized?
What gets approved?
What requires escalation?
Which customer receives attention?
Which risk receives investigation?
Which recommendation becomes action?
Which exception stops the process?
Which opportunity is ignored?
Many AI use cases become clearer when framed around decisions rather than tasks.
An AI system may:
gather evidence for a decision,
identify relevant context,
recommend an option,
predict likely outcomes,
detect anomalies,
surface uncertainty,
execute an approved action,
or monitor what happened afterward.
The product question becomes:
Which portion of the decision system should intelligence support?
That is much more precise than "add AI."
Separate assistance from autonomy
Once the workflow is visible, teams can make better choices about autonomy.
Some activities are good candidates for assistance.
Draft this.
Summarize this.
Find the relevant policy.
Compare these options.
Prepare this analysis.
Others may support bounded autonomy.
Monitor these conditions.
Take this low-risk action when defined criteria are met.
Update this system if the result is reversible.
Escalate when uncertainty crosses a threshold.
And some decisions should remain explicitly human.
The correct allocation depends on risk, context, value, reversibility, trust, and regulation.
The workflow gives those decisions somewhere concrete to live.
Without it, discussions about agents often become abstract debates about whether AI "should be autonomous."
Autonomy is not a philosophical setting.
It is a workflow design decision.
Enterprise AI is an integration problem
A useful enterprise agent rarely operates in isolation.
It may need:
customer information,
product data,
documents,
policies,
transaction history,
CRM,
ticketing systems,
communication tools,
analytics,
permissions,
and operational systems that actually perform work.
The quality of the model matters.
So does whether the system can access the right information safely and act inside the right boundaries.
This makes AI transformation partly an integration and operating-model challenge.
Who owns access?
Which source is authoritative?
How are permissions inherited?
What happens when systems disagree?
Who monitors the automated action?
What does the user see?
What does Audit see?
What happens when the underlying process changes?
A compelling prototype can conceal all of these questions.
Production cannot.
The process may need to change before AI can help
Some workflows are difficult to automate because they are poorly designed.
Inputs are inconsistent.
Ownership is unclear.
Policies conflict.
Data is inaccessible.
Exceptions dominate.
Every team uses different terminology.
A human survives the workflow through institutional knowledge and improvisation.
An AI system exposes the disorder.
That can be frustrating.
It can also be useful.
AI readiness can become a diagnostic tool.
If the organization cannot explain the process well enough to determine what an agent should do, the organization may not understand the process as well as it thought.
The transformation opportunity may therefore begin with:
simplifying the workflow,
clarifying decision rights,
standardizing inputs,
improving data,
removing obsolete steps,
and defining success.
Only then does automation become durable.
Measure the workflow outcome, not the AI activity
AI initiatives can easily measure the wrong thing.
Prompts sent.
Summaries generated.
Employees trained.
Agents created.
Hours theoretically saved.
Those may be useful operational measures.
They are not necessarily transformation outcomes.
The real measures often live in the workflow:
cycle time,
conversion,
resolution rate,
quality,
error,
customer effort,
employee effort,
risk,
cost,
revenue,
retention,
or time to decision.
If AI produces more activity but the workflow outcome does not improve, the transformation has not worked.
This is another reason to understand the process before introducing the technology.
You need to know what good looks like before deciding whether AI made it better.
The organizational model changes too
A successful AI implementation can alter responsibilities.
People may move from creating first drafts to reviewing exceptions.
Managers may supervise outcomes rather than tasks.
Specialists may handle fewer routine cases and more ambiguous ones.
Product teams may receive new behavioral data.
Operations teams may need new monitoring capabilities.
Risk teams may need visibility into automated decisions.
The workflow after AI may therefore require different skills, roles, metrics, and escalation paths.
Transformation is not merely:
Old workflow + AI.
It may be:
Different workflow, different responsibilities, different information flows, different controls, and different expectations of people.
That is organizational design.
The best use case may be removing work
AI transformation programs can become biased toward adding intelligence.
But one of the most valuable outcomes of workflow analysis is discovering work that should disappear.
A report nobody needs.
A duplicate approval.
A handoff created by organizational history.
Information repeatedly copied between systems.
A meeting whose only purpose is synchronizing data.
A form whose fields already exist elsewhere.
The best AI solution to unnecessary work is sometimes not an agent.
It is deletion.
Technology leaders should be equally excited about eliminating complexity as automating it.
A better transformation sequence
Instead of:
Find AI use case → build prototype → search for adoption
I prefer:
Understand the workflow → identify the constraint → define the desired outcome → locate consequential decisions → redesign responsibility → determine where AI creates leverage → establish controls and evaluation → deploy → measure workflow impact → learn and redesign again
That is slower at the beginning.
It is often faster overall.
Because the organization is solving the problem rather than merely implementing the technology.
AI transformation is operational transformation
The most durable AI advantage will not come from how many employees have access to a model.
Access will become ordinary.
The advantage will come from redesigning important work around new capabilities while preserving judgment, trust, accountability, and business value.
That requires people who can move between:
customer experience,
product,
technology,
operations,
organization,
risk,
and economics.
AI transformation is therefore not a technology-installation problem.
It is a workflow and operating-system problem.
And the workflow is where I would start.