AIYING · V5.0 · AI IN ACTION
When Expertise Reaches the Newcomer
See differently. Generate possibilities. Accomplish what matters.
By Ravindra Dastikop
An experienced customer-support professional often knows which question to ask before suggesting a solution.
A newcomer may have access to the same documentation but struggle to recognise what matters in the customer’s description.
What if AI could make some of that practical experience available during the conversation?
A documented workplace deployment offers evidence—and a reason to avoid giving everyone the same kind of assistance.
The real-world deployment
In Generative AI at Work, researchers Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied a staggered AI rollout using data from 5,172 customer-support agents.
The assistant suggested responses and relevant documentation during customer chats. Human agents could accept, edit, or ignore its suggestions. The published study reported 15% more issues resolved per hour on average. [1]
The average, however, concealed an important difference.
Less experienced and lower-skilled workers benefited substantially. The most experienced and highest-skilled workers saw small speed gains and small quality declines. This was evidence from one firm, not a universal productivity guarantee. [1]
See differently: where is expertise available?
Consider a familiar workplace situation.
A new employee encounters an unfamiliar problem. The organisation has manuals, training sessions, and experienced colleagues. Yet the employee still needs help deciding:
- What should I investigate first?
- Which detail changes the diagnosis?
- When should I stop and escalate?
- How should I explain the next step?
Our interpretation: The opportunity is to bring relevant experience closer to the moment when a person needs to act.
That changes the design question from “What can this AI produce?” to:
“Whose difficult next step could become easier with the right assistance?”
The assumption to challenge
Hidden assumption: If an AI assistant improves average performance, everyone should use it in the same way.
The movement is to separate different needs.
A newcomer might benefit from structured diagnostic questions. An experienced worker might prefer documentation on request. An expert handling an unusual case might need space to investigate independently.
The practical possibility is assistance that varies with the person and the task.
Turn that possibility into a pilot
Imagine a support team considering an AI assistant. The following is a proposed application, not part of the reported deployment.
| Worker’s situation | Assistance to test | Evidence to examine |
|---|---|---|
| New to a recurring task | Suggested diagnostic questions and relevant guidance | Correct resolution and appropriate escalation |
| Familiar with the task | Documentation and optional suggestions | Resolution time alongside quality |
| Experienced, facing an unusual case | Search support and alternative explanations on request | Diagnostic quality and useful new findings |
The aim is to discover where assistance helps, where it adds little, and where it interferes.
Measure accomplishment, not activity
Counting generated suggestions tells you how much the system produces.
It does not tell you whether the customer’s problem was solved.
For a pilot, choose one outcome and a small set of checks:
- Outcome: Correctly resolved issues per working hour.
- Quality check: Reopened cases or independently reviewed errors.
- Customer check: Whether the customer understood the next step.
- Learning check: Whether workers can explain their decisions.
These are proposed measures. Select definitions and review periods that fit your workflow before starting.
Examine results separately for newcomers and experienced staff. A positive team average can coexist with a poor result for one group.
Where could the situation generate its own next possibilities?
A support conversation can reveal more than a resolved ticket.
It may expose a confusing instruction, an undocumented exception, or a question that newcomers repeatedly miss.
A possible extension: Review selected cases and turn those discoveries into improvements:
| Discovery from work | Possible next accomplishment |
|---|---|
| Customers repeatedly misunderstand an instruction | Rewrite and test that instruction |
| Newcomers miss a diagnostic question | Add a worked example to training |
| Experts repeatedly reject a suggestion | Investigate the suggestion and its context |
| A case falls outside existing guidance | Document a reviewed exception |
This improvement cycle is our proposed extension. It should not be confused with an outcome established by the study.
Accomplish what matters
Start with one recurring task where newcomers need frequent help.
Define what successful work looks like. Test assistance on a limited scope. Review quality and results by experience level. Use observed failures to decide what to change next.
The accomplishment is concrete:
A newcomer resolves a real problem correctly, can explain the decision, and knows when to seek help.
Source and evidence note
[1] Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942.
This post uses the published journal version. The reported findings concern a historical workplace deployment; this is not a new product announcement. Interpretations, the PO provocation, and proposed pilot designs are identified separately from the research findings.

