Human–agent co-work is neither a chatbot beside the process nor complete autonomy. It is an operating model in which people and agents share the same case state, available sources, executed actions and open decisions. The goal is not to remove people from every loop but to focus their attention in the right places.
Create one shared case state
Email, chat and tickets often create several versions of truth. A co-work system needs one status, owner, priority, next action and deadline per case. Agent activity becomes part of that state rather than disappearing into a separate conversation.
Every action shows whether it was proposed, approved, executed or failed. People must be able to understand what the system has already done.
Expose sources and uncertainty
Answers point to the approved documents or system data from which they were created. When sources conflict or are missing, the system does not produce a confident-looking conclusion; it routes the case for clarification with the relevant context.
The People + AI Guidebook recommends designing mental models, feedback and failure behavior deliberately. Users need a realistic expectation of what an agent can do and where it needs help.
Scale approval with risk
Reading and summarizing within bounded sources may run with little intervention. An internal draft has different consequences from an external message, price change or booking. Define permissions by tool and escalation thresholds by impact and uncertainty.
An approval presents the input, planned action, affected data and possible consequence. Only then can a person make a meaningful decision.
Capture correction as a learning signal
Free-text feedback is difficult to analyze. Classify corrections: wrong source, missing information, rule violation, unsuitable wording or required exception. Feed those signals into knowledge, tools, prompts and evaluations.
Measure the complete system
Do not measure automated volume alone. Lead time, rework, escalation quality, critical failures and business outcome matter. An agent that closes many easy cases but escalates difficult ones too late can make the process worse.
A sustainable co-work model makes accountability clearer: agents perform bounded work; people own goals, rules and consequential decisions.