Role Responsibilities, Processes, and Human–AI Collaboration
Role Responsibilities, Processes, and Human–AI Collaboration
The same process definition must support role responsibilities, work instructions, performance criteria and AI Skills. This is not a writing requirement for aligning documents, but a theoretical prerequisite for human–AI collaboration to become repeatable operating capability.
The process definition sits at an intermediate scale. It specifies a process's local purpose, inputs and outputs, acceptance criteria, interface conditions and takeover requirements when exceptions occur. Once established, this definition becomes the common basis for accountability, execution and evaluation.
Role responsibilities use it to allocate accountability and authority: who owns the deliverable, who can accept it, and who takes over when conditions are insufficient. Work instructions use it to explain how people make judgements and perform actions. Performance criteria use it to define evaluation standards, keeping what performance measures aligned with what the process requires. AI Skills use it to package executable rules, permitted inputs and constraints. If the four are written independently, they impose different requirements on the same work: responsibilities emphasise results, instructions omit review, evaluation counts only volume, and the Skill generates according to another set of rules. Accountability, execution and evaluation then become disconnected; problems cannot be located and reuse cannot be achieved.
A common standard is necessary because accountability does not change according to whether the executor is a person or AI. Acceptance criteria define whether a result is acceptable, not who produces it. Authority and takeover points define when human intervention is required, rather than the limits of a tool's capabilities. If people and AI use different standards, it becomes impossible to tell whether acceptance came from the method itself or a particular executor's special handling. Nor can another person repeat the work using only the instructions or Skill and achieve the same result. Evidence consequently loses its basis for comparison, and the formation of capability cannot be tested.
The process definition must also remain constrained by the scales above and below it. It serves operating results above and constrains specific actions below. Only when responsibilities, instructions, evaluation criteria and Skills are derived from this same definition can the four scales validate one another: action outputs can be accepted as process deliverables, process deliverables can be taken up as operating results, and accountability and evaluation can point to the same deliverable. Any use of different standards reveals a break during handoff or reuse.
When designing human–AI collaboration, therefore, establish one process definition first, then make role responsibilities, work instructions, performance criteria and AI Skills depend on it together. Only then can people and AI divide the work within the same acceptance criteria and authority boundaries, and a method potentially be confirmed through evidence and reuse as lasting operating capability rather than temporary coordination dependent on particular people.
Mapping with a Value-Stream Diagram
A value-stream map can unfold a business workflow at the process scale, making operating results, process interfaces and handoffs observable. It is not a standalone management tool here, but a mapping method that helps the same result impose mutual constraints across different scales.
The map starts with the final result stakeholders need and shows, in their actual sequence, the processes through which that result is produced. Each process retains only the deliverables directly related to the result: what it receives, what it transforms or judges, and what it passes to the next process. Every handoff explicitly marks two judgements that must remain separate: whether the previous process's result meets the acceptance criteria, and whether the next process has the authority, resources and dependencies needed to start. Waiting, rework, missing authority and repeated clarification, which do not directly produce acceptable results, are also recorded faithfully so evidence can later distinguish the nature of each problem.
Once mapped this way, the constraints across four scales can be tested on the same diagram. The operating result sits at the end of the value stream and becomes the objective against which every process must be validated. The business workflow consists of the process sequence in the stream, whose interfaces must support the two handoff checks. Functional responsibilities map to delivery and handoff accountability in each process, assigning it to specific roles. Specific actions sit inside each process and explain how people and AI perform local tasks using the same acceptance criteria. If a process cannot explain how it contributes an acceptable deliverable to the final result, or a handoff cannot record result acceptance separately from start conditions, the scales are not constraining one another at that point.
The value-stream map also gives evidence collection a place. Execution evidence records whether processes run as mapped. Acceptance evidence corresponds to the first judgement at a handoff. Downstream-use evidence corresponds to the second judgement and actual use by the next process. Objective-achievement evidence records whether the final result improves operating metrics. In a reuse test, another person can follow the same map and its associated process definitions, acceptance criteria and Skills, run the workflow again and compare results.
The purpose of introducing a value-stream map here is therefore not to optimize speed or eliminate all waiting, but to clarify a workflow's processes and interfaces so that value-defined results, the two handoff judgements, assigned responsibilities and four kinds of evidence can be tested together in one structure. Only after this mapping can the internal consistency of the full logic be assessed within the workflow, informing whether the method can be retained or expanded.
The theoretical basis for one acceptance standard for people and AI
People and AI must share the same acceptance standard because accountability does not change when the executor changes. Acceptance standards define whether an outcome reaches the state required by stakeholder value, not who produces that outcome.
Once a process deliverable has been defined, its acceptability becomes an objective judgment: is its scope complete, is its basis traceable, is its essential content correct, and does it meet the conditions for the next process to take it up? This standard comes from the constraints of operating outcomes and is passed down through the process definition. Whether the executor is a person or AI, the deliverable must be tested against the same standard. Using different standards for different executors would make the meaning of a result unstable: the same deliverable would be accepted in one case and rejected in another, so value could not be realised reliably.
Accountability likewise does not change with the executor. The role accountable for a deliverable must ensure that the result meets the standard and must correct it or take over when it does not. AI may perform some actions or generate drafts within its permissions, but this neither assigns accountability to AI nor removes it from people. Accountability remains with the person granted authority over that process. A shared standard lets the accountable person decide whether to intervene using explicit criteria, rather than subjective estimates of a tool’s capabilities. Permissions and takeover points can therefore be defined clearly: a person takes over when inputs are missing, rules do not cover the situation, or the result cannot be confirmed against the standard.
A shared standard is also a prerequisite for evidence and reuse. Only when people and AI undergo the same acceptance process can results from different executors be compared to establish whether the method itself works, rather than whether one executor’s special handling works. In a reuse test, the result achieved by another person using the manual or Skill must also be assessed against the same standard. If standards vary by person, evidence loses its comparability and cannot establish whether the capability has moved beyond its original designer and been transferred.
Within the role–process–human–AI collaboration model, acceptance standards must therefore be established before work is divided and recorded in people’s operating manuals, assessment rules, and AI Skills alike. Accountability, judgment, and evidence are all anchored in that same standard. Only then can the executor change without changing the requirements for the result or the boundaries of accountability, allowing the method to be tested as a repeatable operating capability.
Applied example: reviewing a quote
A sales team is preparing a quote for a customer. The product specifications and price calculations have passed review, but the quote includes payment terms outside standard policy. The reviewer can accept the package as complete. The account lead still needs approval from the person authorized to grant the exception before sending it.
This research note uses a fictional quote review to apply the source model’s approach to role responsibilities, processes and human–AI collaboration.
Assign responsibility around what the next person needs
Start with the people who use the package. The commercial reviewer needs to check pricing, specifications, terms and the quote version. The account lead needs a quote that can be sent to the customer and confirmation that any required approval is in place. Those needs define what the commercial analyst submits and what each recipient has to check.
The commercial owner maintains pricing and the approval policy, including who can handle exceptions. The analyst assembles the package, the reviewer checks it against the agreed criteria, and the designated approver decides on exceptions. The account lead sends the approved version. Each role needs to know what it delivers, who receives it and what it has the authority to decide.
Apply the same acceptance criteria to human and AI work
Make one checklist for scope, specifications, units, pricing references, expiry dates and exceptions. Every item should be traceable to the input used and the version in force. A human playbook and an AI skill can use this checklist as their common standard.
AI can compare documents, flag differences against known rules and point to the relevant entries. An analyst checks unsupported policy terms, conflicting inputs and prices that cannot be verified. The people with the relevant authority record the review outcome and any exception approval.
Distinguish rework from a wait for authorization
If the order is specified in cartons but the quote uses the price per item, the package has failed acceptance. The reviewer identifies the line that needs correcting, and the analyst verifies the unit, updates the package and resubmits it.
If the specifications, amount and references have passed review but an exception to payment terms is pending, record the package as accepted and hold it from sending. Assign the approval to the authorized person and agree when a decision is needed. If prices or terms change while the quote is on hold, check whether the earlier review still applies.
Use handoff records to improve the process
For the pilot, record the quote ID and version, who submitted it, what the reviewer decided, which conditions are pending, who owns them and when the quote was sent. At review, look separately at problems found in the package, time spent waiting for approval, and feedback after the customer received the quote.
If unit errors keep recurring, improve the input check and test it on a small set of similar quotes. If exceptions repeatedly miss the response deadline, examine approval ownership and timing. Change one part of the process at a time, observe what happens to accuracy, delivery time and reviewer effort, then decide whether to update the playbook and AI skill.
Source
Read the original research · Andrew David W.Y.