AI automation is increasingly moving beyond isolated productivity tools and into the operating systems of businesses. AI is being embedded in customer service, logistics, finance, HR, software development, enterprise applications and decision-support workflows.
The productivity case is strong. Research from Harvard Business School and collaborators has found meaningful performance gains from generative AI on a range of knowledge-work tasks, while Stanford’s AI Index has documented the rapid growth of organizational AI adoption.
But there is another question that receives less attention:
What happens to organizational expertise when people stop doing the work that taught them how to do it?
This is not an argument against automation. It is an argument for designing automation more carefully.
The goal should not be to preserve every manual task. Many manual tasks are precisely what organizations should eliminate. The challenge is distinguishing between work that creates unnecessary process burden and work that, when removed entirely, also removes opportunities to maintain judgment, diagnosis and operational competence.
Automation Changes More Than Workload
A common way to evaluate automation is to ask whether a task can be performed faster, more cheaply or more consistently by software.
Those are useful questions, but they are incomplete.
A task can have two outputs:
- The immediate business result.
- The expertise employees develop by performing or reviewing the task.
Suppose a company automates the classification of customer complaints. Previously, employees read a broad range of complaints, identified patterns and decided which cases required escalation.
After automation, employees may see only unusual cases.
That can be beneficial. They spend less time on repetitive work and more time on difficult cases.
But there is a condition: they must continue receiving enough exposure to the underlying domain to recognize unusual cases correctly.
If the automation becomes so dominant that employees no longer understand the normal process, the organization can develop a strange dependency. It has fewer people doing routine work, but also fewer people capable of reconstructing the process when the system fails.
This is the core distinction between task automation and capability preservation.
Expertise Is More Than Documentation
Organizations often respond to knowledge-retention concerns by documenting procedures.
Documentation matters, but it does not fully solve the problem.
Some knowledge is explicit: process instructions, policies, formulas, system configurations and decision trees.
Other knowledge is contextual.
An experienced employee may recognize that a particular transaction “doesn’t look right” before being able to explain exactly why. A logistics manager may know that a routing recommendation is technically valid but operationally unrealistic because of a recurring local constraint. A senior engineer may recognize that an apparently harmless code change creates an architectural problem.
This is often called tacit or experiential knowledge.
Research into organizational knowledge creation increasingly examines how AI can help convert tacit knowledge into more accessible organizational knowledge. A 2026 systematic review of 82 academic articles, for example, found that AI can participate in processes involving the conversion and combination of organizational knowledge.
That is promising, but there is an important limitation:
Capturing expertise is not the same as maintaining expertise.
A company can document how experts think while simultaneously allowing the people who possess that expertise to stop practicing it.
The Skill-Erosion Mechanism
Skill erosion usually does not happen because management explicitly tells employees to stop learning.
It happens through workflow design.
Consider a simplified sequence:
Manual work → AI assistance → AI recommendation → automatic execution
At each stage, less human effort may be required.
The first stage can be highly beneficial. An employee spends less time searching for information.
The second can also be beneficial. AI prepares a draft or recommendation that the employee reviews.
The third creates more dependence. The employee increasingly evaluates an answer rather than developing one.
The fourth creates the greatest capability question: if the system is wrong, who still knows how the underlying work operates?
Recent research on technology-driven skill degradation describes this as a form of skill depreciation caused by sustained reliance on technology. The research highlights mechanisms including substitution and automation bias.
The practical lesson is not “keep everything manual.”
It is:
Do not confuse removing a task with removing the capability behind the task.
A Better Automation Framework
A useful way to assess an AI-enabled workflow is to classify each activity according to four dimensions:
1. Repeatability
Does the task follow a predictable pattern?
Highly repetitive activities are strong candidates for automation.
Examples include:
- Data formatting
- Routine classification
- Scheduling
- Status updates
- Standard document generation
- Basic information retrieval
2. Judgment
How much contextual reasoning is required?
A task can be repetitive but still contain important judgment. For example, reviewing invoices may be routine until a transaction violates several unusual business rules.
Judgment-heavy tasks require greater care.
3. Consequence
What happens if the system gets the task wrong?
A flawed internal categorization may be easy to correct.
A flawed financial, medical, compliance or customer-impacting decision may not be.
The higher the consequence, the stronger the case for meaningful human involvement.
4. Expertise Dependency
Does performing this task help employees maintain an important organizational capability?
This is the dimension organizations often omit.
A task may be inefficient but still useful because it keeps employees familiar with the domain.
That does not mean the task should remain manual forever. It means the organization should replace the learning opportunity if it removes the task.
What Should Be Automated First?
The safest early candidates are usually activities that are:
- Highly repetitive
- Well-defined
- Easy to verify
- Low consequence when incorrect
- Poor sources of differentiated human expertise
Administrative process load is often a good target.
For example, AI can prepare routine meeting summaries, populate structured records, classify incoming requests or identify missing information.
The human can then spend more time on activities that require interpretation, negotiation, prioritization or relationship management.
The distinction becomes especially important when the organization considers automating an entire workflow rather than individual steps.
A process may contain ten activities. Seven may be highly automatable. Two may require expert review. One may require accountable human judgment.
Automating the seven is very different from automating all ten.
Preserve Expertise Through Exception Work
One of the strongest ways to preserve expertise is to make exceptions part of the operating model.
Suppose an AI system successfully handles 95% of incoming service requests.
The remaining 5% should not simply be treated as “the tickets the AI couldn’t answer.”
They are a strategic learning pool.
Those cases can reveal:
- New customer needs
- Missing business rules
- Model limitations
- Process weaknesses
- New edge cases
- Training opportunities
- Changes in the operating environment
The expert team should therefore have a defined mechanism for reviewing difficult cases and feeding the lessons back into the workflow.
This creates a loop:
Automation → exception → expert judgment → documented lesson → workflow improvement
That is very different from:
Automation → exception → manual cleanup
The first creates organizational learning. The second merely creates residual work.
Keep a Real Manual Fallback
Business continuity plans frequently assume that critical systems can fail.
AI introduces another version of the same problem: employees may technically have a fallback process but lack sufficient familiarity to execute it.
A documented manual procedure is not necessarily a usable manual capability.
Employees need periodic exposure to critical fallback processes, particularly when automation is embedded deeply into operations.
This does not mean conducting full manual operations every week.
It may mean periodic exercises.
For example:
- Temporarily disable an automated workflow.
- Give the responsible team a representative case.
- Observe how quickly the process can be completed.
- Identify where employees depend on information available only through the AI system.
- Update documentation and training.
The objective is not to prove that people can work without technology indefinitely.
It is to discover whether the organization can recover when technology becomes unavailable.
Measure Capability Alongside Productivity
AI programs often use metrics such as:
- Processing time
- Cost per transaction
- Automation rate
- Error rate
- Throughput
- Employee time saved
These are valuable.
But organizations should also track capability.
Useful questions include:
- Can employees explain why an AI recommendation is incorrect?
- Can they identify cases that require escalation?
- Can they complete critical processes without the AI system?
- Can new employees learn the underlying process rather than simply learning the interface?
- Are experienced employees still exposed to difficult cases?
- Is expertise becoming concentrated in a small number of people?
These questions matter because automation can improve average performance while increasing the consequences of rare failures.
NBER research on expertise and automation reinforces the importance of this distinction: the effect of automation depends partly on whether removing tasks reduces or increases the expertise required for the tasks that remain.
A Practical Example: AI in Customer Operations
Imagine a B2B company using AI to manage incoming support requests.
Before automation, support specialists handled every ticket.
After implementation:
- AI classifies requests.
- AI retrieves relevant documentation.
- AI drafts responses.
- Routine requests are resolved automatically.
- Complex cases go to specialists.
The first four changes can be highly beneficial.
The potential problem is the final step.
If specialists receive only the most unusual cases, they may lose familiarity with common customer workflows. New employees may never learn the basic operating model because the AI handles it for them.
A better design could include:
- Automatic handling of low-risk requests.
- Human review for ambiguous cases.
- Periodic expert sampling of automated resolutions.
- Rotational exposure to routine cases for newer employees.
- A library of unusual cases and resolutions.
- Regular testing of manual escalation procedures.
The goal is not to force experts back into repetitive work.
It is to make sure the organization continues producing people who understand the system it operates.
Four Common Mistakes
Mistake 1: Maximizing automation percentage
A high automation rate is not automatically a high-quality outcome.
The correct target is the right level of automation for the business risk and capability requirements.
Mistake 2: Treating human approval as expertise
A person clicking “approve” does not create meaningful oversight.
If the reviewer lacks time, context, authority or independent knowledge, the human checkpoint may be mostly cosmetic.
Research on automation bias shows that people can become overly reliant on automated recommendations, with factors including professional expertise, AI literacy and verification demands influencing how that reliance develops.
Mistake 3: Training people only on the AI interface
Employees need to understand what the AI is doing, but they also need enough domain knowledge to evaluate it.
Otherwise, the organization creates tool operators rather than capable decision-makers.
Mistake 4: Assuming documentation preserves everything
Documentation preserves instructions better than judgment.
Organizations should combine documentation with mentoring, case review, simulations, peer learning and deliberate exposure to difficult situations.
The Right Question for Leaders
The question should not be:
“Which jobs can AI replace?”
A more useful operational question is:
“Which tasks can AI perform without weakening capabilities we will still need?”
That reframes automation from a labor-reduction exercise into a system-design problem.
The distinction also fits the broader evidence about AI and work. The ILO’s 2025 analysis finds that GenAI exposure is widespread, but most occupations are more likely to be transformed than eliminated because human input remains necessary across many tasks. The World Economic Forum likewise expects significant skill change through 2030 while highlighting both technological skills and human capabilities such as analytical thinking, resilience and lifelong learning.
The implication is straightforward.
Organizations do not need to preserve every old way of working.
They do need to preserve the ability to understand, challenge, repair and improve the systems they increasingly rely on.
A Five-Step Action Plan
Start with one important workflow.
Step 1: Map the tasks.
Separate repetitive execution, information processing, judgment and accountability.
Step 2: Identify must-retain capabilities.
Ask what employees need to know if the system produces an unexpected result.
Step 3: Automate selectively.
Start with repeatable, verifiable and relatively low-consequence activities.
Step 4: Design exception learning.
Make unusual cases visible to qualified people and use them to improve the system.
Step 5: Test resilience.
Periodically test whether employees can identify errors and operate critical processes when automation is unavailable.
This approach does not reject automation.
It makes automation safer and more durable.
FAQ
Can AI really cause employee skill erosion?
Yes, sustained dependence on technology can contribute to skill degradation when employees lose opportunities to practice capabilities that remain important. The risk varies by task, workflow design and how much human reasoning remains involved.
Should companies keep manual processes just in case AI fails?
Not every manual process needs to be maintained indefinitely. Critical capabilities deserve fallback mechanisms; low-value administrative procedures generally do not.
How can companies preserve tacit knowledge?
Use multiple mechanisms: expert mentoring, case reviews, documented decisions, simulations, communities of practice and structured exposure to difficult cases. AI can help capture and make knowledge searchable, but it should not be treated as a substitute for human expertise.
Is human-in-the-loop enough?
No. Human involvement is meaningful only when people have sufficient knowledge, context, authority and time to challenge the system. Research on automation bias shows why simply placing a person after an automated recommendation is not a complete safeguard.
What is the best metric for AI automation?
There is no single metric. Combine productivity measures with quality, exception rates, human capability, recovery performance and the ability to detect and correct AI failures.
Final Takeaway
The strongest AI-enabled organizations will not be the ones that automate the greatest percentage of their workflows.
They will be the ones that understand which work should disappear, which expertise must remain, and how the two can coexist.
Automation should reduce unnecessary process load.
It should not quietly automate away the organization’s ability to think, diagnose and recover.
