Generative AI adoption has moved faster than almost any workplace technology in modern history. But while headlines focus on disruption and job loss, a more practical question matters for Learning and Development leaders:
Where exactly is AI changing work—and what should we do about it?
A large-scale study from Microsoft Research offers an unusually concrete answer.
In Working with AI: Measuring the Applicability of Generative AI to Occupations (Tomlinson, Jaffe, Wang, Counts, & Suri, 2025), researchers analyzed 200,000 anonymized conversations with Microsoft Copilot and mapped them to O*NET work activities. Rather than forecasting theoretical impact, they examined where AI is already being used successfully in real work tasks.
The findings have direct implications for every L&D function.
AI Is Most Effective in Information Work
The study found that the most common and most successful AI-assisted work activities involve:
- Writing and editing
- Explaining policies and procedures
- Gathering information
- Communicating technical details
- Responding to customer inquiries
- Preparing instructional materials
In short: creation, processing, and communication of information.
This matters because most jobs—even those not considered “knowledge work”—contain some information component. Documentation, reporting, scheduling, compliance communication, and customer interaction are embedded across roles.
The researchers conclude that AI applicability is widespread because most occupations include information work elements. AI is not limited to engineers or analysts. It overlaps with everyday tasks across sectors.
For L&D, this reframes AI from a specialized technical skill to a baseline productivity capability.
The Automation vs. Augmentation Distinction
One of the study’s most useful contributions is its separation of two types of AI applicability:
- User goal applicability: Employees use AI to assist their work.
- AI action applicability: AI performs parts of the work activity itself.
This distinction changes how training should be designed.
In some roles, AI acts as a collaborator—supporting drafting, analysis, or explanation. In others, it performs discrete components of the job, shifting what humans focus on.
L&D programs often treat AI capability as a single category—usually framed as “AI literacy” or tool proficiency. The research suggests that approach is too simplistic.
Instead, organizations should ask:
- Which tasks in this role should be delegated to AI?
- Which require human judgment?
- Where does AI increase speed but introduce risk?
- What decision-making capability becomes more important as automation increases?
The skill shift is not just technical. It is cognitive.
The Measurement Problem
Perhaps the most important lesson for L&D lies in how the researchers measured AI impact.
They evaluated:
- Task completion success
- Scope of AI’s capability within work activities
- Frequency of AI use in specific tasks
- Applicability at the occupation level
They did not measure engagement, attendance, or usage time.
This exposes a gap in many AI training initiatives.
If your organization is measuring:
- Percentage of employees who completed AI training
- Learner satisfaction scores
- Hours of content consumed
You are not measuring performance impact.
AI capability initiatives should track:
- Decision quality before and after intervention
- Reduction in rework or error rates
- Escalation accuracy
- Speed-to-output while maintaining quality
- Risk calibration in high-stakes decisions
AI changes how work is performed. Training must therefore measure how work outcomes change.
AI May Democratise Expertise—But Only If Employees Can Evaluate It
The authors suggest that generative AI may reduce performance gaps by broadening access to expertise. If employees can apply AI-generated insights effectively, previously specialized capabilities may become more widely distributed.
However, there is a crucial caveat: employees must have sufficient foundational knowledge to evaluate AI output.
Without domain knowledge, AI amplifies error.
For L&D, this creates a paradox. AI does not eliminate the need for expertise. It increases the need for:
- Foundational domain understanding
- Critical thinking
- Output validation skills
- Ethical and risk judgment
Training programs that focus exclusively on tool usage will underperform. Programs that strengthen evaluation and judgment will create durable capability.
Where AI Currently Struggles
The study also identifies areas where AI is less effective:
- Physical or manual tasks
- Highly contextual decision-making
- Certain forms of data analysis
- Tasks requiring nuanced interpretation beyond provided information
This is equally important for training design.
Employees must learn:
- When not to rely on AI
- How to detect hallucinations or incomplete reasoning
- How to validate outputs
- When to escalate to human expertise
Responsible AI capability is not about compliance modules. It is about calibrated judgment.
What This Means for L&D Strategy
The findings point toward several structural shifts for L&D leaders.
-
Redesign Training Around Work Activities, Not Tools
Instead of building general AI courses, map AI integration to:
- Specific work activities
- High-frequency information tasks
- High-risk decision points
Ask: Where does AI overlap with daily work? Where does it meaningfully change task execution?
Training becomes role-specific and workflow-aligned.
-
Replace Awareness Sessions with Simulation
Employees do not need more slide decks explaining AI concepts. They need structured practice.
Effective interventions include:
- Scenario-based decision simulations
- Trade-off exercises under time pressure
- Delegation boundary decisions
- Risk and escalation judgment practice
Measure decision quality movement before and after intervention.
Simulation accelerates capability far more effectively than passive instruction.
-
Update Competency Frameworks
Traditional competency models rarely include:
- Human–AI collaboration skill
- Prompt clarity and refinement
- Output validation
- Escalation judgement
- Risk calibration
These must now be embedded within role profiles and leadership models.
AI capability is not an elective. It is becoming part of performance expectations.
-
Shift L&D’s Own Value Proposition
There is another uncomfortable reality embedded in the research.
AI itself performs tasks such as:
- Explaining procedures
- Drafting materials
- Creating instructional content
- Teaching foundational concepts
If L&D defines its value primarily as content production, that value will compress.
The strategic opportunity lies upstream:
- Diagnosing performance gaps
- Designing practice environments
- Engineering decision simulations
- Building measurement systems
- Aligning AI integration with workforce transformation
L&D’s differentiator becomes system design, not slide development.
From Training Delivery to Performance Engineering
The Microsoft Research study does not predict mass job loss. It measures overlap between AI capability and work activities.
But overlap is enough.
If AI meaningfully assists or performs portions of information work across occupations, then job tasks will shift—even if roles remain intact.
History suggests that technology often changes task composition more than job count. The organizations that benefit most are those that redesign workflows early.
For L&D leaders, the mandate is clear:
- Stop treating AI as a technology topic.
- Start treating it as a work redesign catalyst.
- Move measurement from participation to performance.
- Design for judgment, not just knowledge.
AI is not replacing learning.
It is raising the standard of what effective learning must accomplish.
The future of L&D will not be defined by how quickly we deploy AI courses.
It will be defined by how effectively we help people think, decide, and perform in AI-augmented environments.
And that is work worth leading.
