Reads the incoming chat, searches available knowledge, and drafts a response. → Resolves the issue or escalates when managerial intervention is needed.
After
Observes the live chat and proposes response text based on patterns from prior interactions. → Accepts, edits, or ignores the suggestion and owns the sent response and escalation.
Human boundary
Agent owns every sent response; managers own escalation.
Why it matters
AI can diffuse expert practice, with value concentrated among novices.
How the work changed
Before
How the work ran before the change.
Step 1 of 2
Customer-support agent
Reads the incoming chat, searches available knowledge, and drafts a response.
ControlControl group had no generative assistant; agents could handle concurrent chats.
Step 2 of 2
Support agent or manager
Resolves the issue or escalates when managerial intervention is needed.
ControlHuman resolution and escalation authority.
What changed
AI can diffuse expert practice, with value concentrated among novices.
Decision rightHuman moves from creator to judge
After
How the same work runs now.
Step 1 of 2
Generative conversational assistant
Observes the live chat and proposes response text based on patterns from prior interactions.
ControlSuggestion only; the working paper does not disclose the company or vendor.
Step 2 of 2
Support agent
Accepts, edits, or ignores the suggestion and owns the sent response and escalation.
ControlHuman acceptance control; effects must be monitored by agent skill because top-agent gains were minimal.
Process model built from the published workflow evidence for Anonymous Fortune 500 software company. Every step, actor, and control appears in full below.Every step, actor, and control
Exception path
Incorrect suggestions are ignored or escalated.
Work removed
Some manual search and drafting
Decision authority
Agent owns every sent response; managers own escalation.
Before
01
Customer-support agent
Reads the incoming chat, searches available knowledge, and drafts a response.
Control: Control group had no generative assistant; agents could handle concurrent chats.
02
Support agent or manager
Resolves the issue or escalates when managerial intervention is needed.
Control: Human resolution and escalation authority.
After
01
Generative conversational assistant
Observes the live chat and proposes response text based on patterns from prior interactions.
Control: Suggestion only; the working paper does not disclose the company or vendor.
02
Support agent
Accepts, edits, or ignores the suggestion and owns the sent response and escalation.
Control: Human acceptance control; effects must be monitored by agent skill because top-agent gains were minimal.
Work that left the path
Some manual search and drafting
Human role before
Customer-support agents read incoming chats, searched available knowledge, drafted replies, and escalated cases requiring managerial intervention.
Human role after
Agents accept, edit or ignore suggestions and own resolution.
AI roleObserves chats and suggests responses in real time.
Outcomes
Issues resolved per hour
Verified
Agents without access→About 14% average increase; 35% for novice/low-skill agents
