Purpose Description: A New Role Definition for the AI Era

An Organizational Theory of Dynamic Role Design and Future Value Creation in the Age of AI

​

Job descriptions (JDs) go stale faster than they used to. Most companies respond by rewriting them more often.

This paper makes a different claim. The problem is not the frequency of revision. It is the starting point of the definition. As long as a role is written from "what do you do," every shift in what AI can take over makes it obsolete again.

So the paper starts elsewhere: from purpose, beneficiary, questions and decision principles. That is the Purpose Description (PD).

VURA Working Paper No.4. This page explains the core argument in plain English; the full paper (English original, 35 pages; Japanese edition, 29 pages) is linked at the end.

This paper in three minutes

• It rejects, at the outset, the idea of "leaving to humans whatever AI cannot do." The reason is structural, not a view of human nature: AI's capability frontier keeps moving, so a role defined as the residual goes stale on its own.

• So it changes the starting point. A role is written not from "what it does" but from why it exists, for whom, what future value it creates, what it keeps asking, and by which principles it decides. That is the PD.

• The PD does not abolish the JD. Role design becomes two-tiered: the PD above provides stability of direction; the JD and its controls below provide agile updating. They complement each other.

• What cascades from the top is questions, not tasks. The procedure is the Query Translation Protocol (QTP). What flows down is a chain of future value hypotheses and questions, not a chain of KPIs.

• AI adoption without a PD lands in "optimization of means without purpose." Technically a success, operationally a modest success, and nothing flows through to future value.

• And the paper spells out the harm of over-use itself. Not every job at the same intensity. Never the sole basis for performance appraisal. Where the conditions are not met, the outcome can be negative, not merely zero — stated explicitly as Proposition P9.

The paper's strongest claim

A position people often take at the start of this discussion: "Make whatever AI cannot do the human role."

The paper rejects it. Not on ethical grounds, not from a view of human nature. The reason is structural.

AI's capability in knowledge work does not expand evenly. Task by task, strengths and weaknesses interlock in what the research literature calls a jagged frontier — a boundary that zigzags rather than running in one clean line.

And the frontier keeps moving. The moment you define the human role as "the work AI cannot yet do," the frontier's next move invalidates it. Definition by residual is definition with permanent obsolescence built in.

An analogy: building a house on a beach as the tide comes in

This is the hinge of the argument, so here is an analogy.

Think of AI's capability as a rising tide. Defining the human role by "what AI cannot do" is like picking a patch of sand that is not yet under water and building a house on it.

The tide keeps rising, and not in a straight line: one hollow floods early while the low ground beside it stays dry. That is the jagged frontier. Choose the site by watching the waterline and the rebuilding never ends. Revising the JD more often is rebuilding faster.

What the PD changes is how you choose the site. Not "where the water does not reach today," but "why do we live here (purpose)," "with whom (beneficiary)," "what life are we trying to make possible (future value)." When the tide changes, the house may move — it should. What does not move is the reason for living there.

The conclusion is not "revise the JD more often." It is this:

To define human roles sustainably, one must start not from tasks but from purpose, questions, judgment, responsibility, and the value to be created.

In plain terms — Do not build the definition on what moves (tasks). Build it on what does not move (purpose), and let the tasks move on top.

As a single contrast: the JD allocates present tasks; the PD distributes the creation of future value.

What the job description assumed

The JD is not purpose-free

A caveat first. The paper does not claim that JDs say nothing about purpose; JDs in practice have a "purpose of the position" field.

The problem is not the document format but how the JD is operated as an institution. Operation revolves around managing assigned tasks and their boundaries, and purpose is easily demoted to a short note. The critique is confined to that.

The JD has long supported the division of labor and specialization, hiring, pay and fair evaluation. The paper denies none of those functions.

What broke: the assumption that job boundaries are stable

What generative AI and autonomous AI agents are breaking is an assumption the JD held implicitly: that the composition of tasks and the boundaries of jobs are stable.

The point is not that occupations vanish overnight. It is that the boundaries of the fine-grained tasks that make up a job, and the right way to allocate those tasks among humans, AI and external partners, keep changing continuously and non-linearly.

The empirical evidence is itself uneven. One study finds that generative AI assistance raised customer-support productivity by 15% on average, with the largest gains for the least experienced workers. Another finds that AI's capability boundary forms the jagged frontier described above.

The answer to "which work can be handed to AI" differs finely from task to task, and keeps moving. Together, the two findings undercut the premise of the static JD.

Defining the Purpose Description

Definition | Purpose Description (PD) A description of why an individual or team exists, for whom it creates what future value, what questions it asks to that end, and by which principles it decides — written in light of the expectations of the future shared by markets and society, and of the future the enterprise seeks to realize.

In plain terms — Not "what is your job," but "whom does your role exist for, and what is it there to bring about."

A PD has six components.

TierComponentWhat it states

InputPurposeWhy this role exists; which part of the enterprise's reason for being it carries

InputBeneficiaryWhose problem it solves; to whom it delivers value

MediatingDefining QuestionsThe core questions it must keep asking to track change and update the role

MediatingDecision PrinciplesThe standards of ethics, quality, safety, customer priority and responsibility that persist even when the means change

OutputDesired FutureThe change of state it commits to bringing about

OutputFuture ValueThe capabilities, options, expectations and structural changes it creates for beneficiaries, markets and society

Specific tasks, technologies, staffing and the human–AI division of labor are all variables subordinate to the six components. Purpose stays stable; means move. That allocation is the heart of the design.

Guarding against misreading (1): not OKR or MBO in new clothes

OKRs (Objectives and Key Results) and MBO (Management by Objectives) are goal-management techniques that set what is to be achieved within a period.

The PD sets something else: why this unit exists, and for whom it creates what future value. It is the smallest unit of dynamic organization design that governs the division of labor and the responsibility boundaries between humans and AI. A different level.

Under high uncertainty, where goals themselves go stale quickly, the anchor is questions and principles, not numerical targets. Where OKRs run alongside a PD, their objectives and key results are subordinate to the PD's purpose, questions and principles, and are updated as learning progresses.

Guarding against misreading (2): the PD does not abolish the JD

The PD is not a proposal to replace the JD. It splits role design into two tiers.

• Upper tier (PD) — purpose, beneficiary, questions, decision principles, desired future, future value. Provides stability of direction.

• Lower tier (JD and controls) — current tasks, the human–AI division of labor, authority, legal obligations, safety standards, reporting lines, minimum performance standards. Updated nimbly as technology advances.

With a PD alone, the locus of final responsibility, the things that must not be done, and the applicable regulations turn vague. With a JD alone, the organization cannot respond to unforeseen change. They are not substitutes but a complementary pair: higher purpose, lower control.

Left alone, the gap between the tiers widens, so a synchronization check runs at regular intervals. Four triggers: introduction of a new AI, automation of a major process, a large change in regulation or the market, and the quarterly or half-yearly review.

The check asks one question only. "Do the current tasks and the current human–AI allocation contribute most to exploring the questions the PD defines and to creating the future value it commits to?"

The Query Translation Protocol: deploy questions, not tasks

Ask individuals and teams to write PDs, and a translation rupture occurs every time.

Distribute a highly abstract corporate purpose as it is, and it ends as a slogan. Rush to concretize it into existing tasks and KPIs, and the vision of future value regresses into a static JD.

The mechanism that closes this rupture is the Query Translation Protocol (QTP) — the procedure by which intent at one level is translated into the language of the level below. Its principle is simple.

Deploy questions, not tasks.

In plain terms — What comes down from above is not "do this" but "ask this." The front line supplies the answer.

The upper levels do not fix answers or means. They provide the future value hypothesis, the strategic intent, the decision principles, and the boundaries that must not be crossed. The lower levels use customer contact, expertise, frontline data and AI capabilities to translate those into more concrete, testable Defining Questions. What cascades is a chain of future value hypotheses and questions, not a chain of KPIs.

H3 | Three levels

1. Company level — "In the age of AI, what essential value and what desirable future does our company create for the market and for society?"

2. Business and strategy level — "To realize that future, which constraints — of customers, of society, of the organization — should this business or function remove?"

3. Team and individual level — "To remove that constraint, whom do we serve as beneficiaries, under what decision principles do we deploy humans and AI, and what do we keep asking?"

This last translation is what gets written into the six components of the PD.

Each level is not a shrunken copy of the wording above it. Keeping the intent and constraints of the higher question, each level concretizes it, in its own context, into beneficiaries, constraints to remove, and testable questions.

Not one-way

Management provides the future value hypotheses, priorities, decision principles, the boundaries of ethics, safety and law, and resource allocation. The front line concretizes the higher questions from frontline evidence and, where warranted, sends back refutations of the higher hypotheses themselves, with proposed revisions.

Approval is not formal assent to wording. It endorses the logical coherence of questions, decision principles, beneficiaries and future value.

Translation quality can be judged by five criteria. Can it explain the causal connection to the higher-level future value? Does it identify concrete beneficiaries and the constraints to remove? Is it an explorable question that fixes neither the answer nor the tasks? Does it preserve the decision principles and responsibility boundaries? Can it be updated in light of results and learning?

How "optimization of means without purpose" happens

Success in AI adoption comes in more than one kind. The paper distinguishes three.

1. Technical success — the model or system works as specified and meets standards of accuracy, speed and safety.

2. Operational success — time, cost, quality, throughput and employee experience improve.

3. Success in future value — new value is created for customers, the company and society, and the expectations of, and capabilities for, the future the company seeks to realize grow stronger.

AI adoption without a PD usually begins with "which of our current operations can we automate?" A useful question for finding candidates.

But it does not ask what matters most. Is the operation necessary at all? Whose problem does it solve? Which judgments and responsibilities should humans retain?

The likely results: faster unnecessary processes, departmental KPIs that improve while customer value does not, non-use on the front line, blurred responsibility boundaries, and a bias toward easily measured workload reduction. Technically a success, operationally a modest success, and nothing flows through to future value. That is the mechanism.

With a PD, the order reverses.

First, define the desired future, beneficiaries, future value, questions and decision principles. Next, make explicit the judgments, actions and capabilities that purpose requires. Then allocate: to humans, setting purpose, judging exceptions, ethics, oversight, halting and accountability; to AI, exploration, generation, prediction and automated execution. Only then design processes and systems. AI is not the starting point. It is a means chosen to realize the purpose.

Reducing workload is not the destination but an intermediate result — a way of moving resources toward the exploration of new value. That is why the QTP precedes the allocation of roles between humans and AI.

Once management has identified the routine execution to automate, the next question is: "Into which unexplored future value do we reinvest the time and cognitive resources freed up?"

Value creation can be shared. Accountability cannot be transferred

The paper does not assume that a human approves every step. It is equally consistent with AI executing autonomously within a defined scope while humans hold purpose, boundaries, evaluation criteria, exceptions and redesign. AI is an execution partner in value creation, and can be a co-creator.

But execution and responsibility are distinct.

The determination of purpose and decision principles, the handling of exceptions, the authority to halt AI, and legal and ethical accountability remain with humans.

In plain terms — Work can be shared. Responsibility cannot.

The further humans step back from direct execution, the more clearly their role must be defined: setter of purpose, overseer of the system, final judge in exceptions, holder of halt authority, and ultimate bearer of responsibility for outcomes.

This boundary matters more as AI grows more capable. Once autonomous AI agents explore, judge and execute across multiple processes, the purpose, beneficiaries, questions, decision principles and halt conditions must be made explicit to the AI as well.

An AI agent may hold all or part of a PD as a "delegated purpose specification." That does not make it an independent purpose-setting agent or a bearer of responsibility. The purposes and principles given to AI are set and approved by humans or a legitimate organizational body. They must be auditable, and revocable at any time.

What the paper admits it cannot yet show

The paper is conceptual; it has not yet demonstrated that PD and QTP work (the next section says so in full). What it does set out at length is where the PD costs, and where it can do harm.

The PD has costs, so do not over-apply it

Co-drafting PDs, updating them regularly, and synchronizing the two tiers carry non-trivial costs of dialogue, information gathering and consensus building. And translating questions is not neutral language processing. It is also a negotiation over resources, authority and evaluation criteria.

So the PD must not be applied to every job at the same frequency and granularity. Intensity varies with the speed of change in AI capabilities, markets, regulation and risk.

• Exploratory, fast-moving domains (R&D, new ventures) — detailed PDs and short review cycles.

• Safety-critical domains (healthcare, manufacturing) — the PD states only, concisely and stably, why the process exists and what must never be crossed. Then the lower tier is made thick: detailed JDs, procedures, qualification requirements, halt authority.

• Standardized work — a simple team PD is the default. Individual PDs are, as a rule, not created.

The PD is never permitted to thin out discipline. Operating costs get a ceiling: measure the share of time spent on synchronization, approval wait times and the volume of duplicated text, and when a set ceiling is exceeded repeatedly, simplify the PD's granularity, frequency and number of items by one notch. Do not copy the same information into the JD, the OKR and the PD.

Dangers specific to the PD

The paper lists the dangers the PD itself creates. Blurring of responsibility through abstraction. Demands for excessive devotion in the organization's name. Forced identification of personal purpose with organizational purpose. Subjectivization of evaluation. Weakening of regulatory, safety and professional responsibility. These are not operational trivia; they are boundary conditions of PD design.

The guardrails come with them. Base judgments of responsibility and discipline on the objective standards of the lower tier. Prohibit unbounded overwork and self-sacrifice justified by the PD. Establish procedures for individual appeal and role adjustment. Codify the front line's authority to halt and refuse AI outputs that violate ethics, safety or customer protection.

And institutionalize intervention protections: an employee who halts AI in good faith on grounds of safety or ethics suffers no disadvantage in evaluation, compensation or assignment. A high count of halts must never itself justify a low rating.

Without this protection, humans become the moral crumple zone. The term borrows from a car's crumple zone, built to absorb the force of a collision. It names the position of a human made to absorb, alone, the impact of a system's failure.

Keeping responsibility with humans does not mean sacrificing humans. Only when a person can grasp the situation, can access the AI's state and reasoning, and is given effective authority, capability and time to halt, correct and escalate, can that person be a bearer of responsibility.

In plain terms — Handing over responsibility without authority, information and time is not responsibility design. It is blame-shifting.

Linking the PD to performance appraisal also needs care. The PD must never be the sole basis for changing working conditions, job grades, wages, direction and supervision, or disciplinary standards.

Evaluation splits into three layers.

1. The contract-and-compliance layer — fulfillment of the JD, work rules, quality, safety.

2. The purpose-contribution layer — formation of questions, adherence to decision principles, understanding of beneficiaries, and learning, judged through behavioral evidence.

3. The results layer — OKRs, KPIs, customer outcomes, future value indicators.

What is evaluated is behavioral evidence. Not enthusiasm, not loyalty, not abstract "sympathy with the purpose." An employee who properly fulfills JD obligations must never be steered toward low ratings, adverse changes, transfer or exit on the sole ground of a vague "misfit with the PD."

Nine propositions and their falsification conditions

This is a conceptual paper. The effectiveness of PD and QTP has not yet been empirically demonstrated.

What it offers is nine testable propositions, each with a falsification condition: "if this result appears, the theory is wrong," written down together with the theory. That is the discipline of this series.

The propositions cover environmental fit (P1), dynamic reallocation between humans and AI (P2), vertical alignment (P3), horizontal alignment (P4), self-adjustment capacity (P5), psychological ownership (P6), the shift in evaluation criteria (P7), suppression of low-value automation (P8), and institutional boundary conditions (P9).

The weightiest is the last.

P9 (Institutional boundary conditions) The positive relationship between PD use and organizational outcomes holds when beneficiaries, decision principles, decision rights, JD-level responsibility standards, intervention protections and appeal procedures are made explicit. The relationship is stronger the higher the psychological safety, procedural justice and non-retaliation norms; where excessive demerit-based evaluation, steep authority gradients, or misalignment of responsibility and authority dominate, the relationship is nullified or turns negative through role ambiguity, silence and unbounded role expansion.

In plain terms — The PD does not work in an organization with bad soil. Worse than not working, it does harm.

The paper itself states, in other words, that introducing a PD where the conditions are not met can produce a negative outcome, not merely zero. It recommends adoption and, in the same breath, writes down the conditions under which adoption must not happen. That is its stance.

Implications for executives, HR and the front line: what changes tomorrow

For executives. For your key roles, is there documentation — separate from the task list — of purpose, beneficiary, questions and decision principles?

If not, your AI adoption decisions almost certainly begin with "which operations can we automate?" That order yields technical and operational success at best; nothing connects to future value. What changes is the order. Before listing automation candidates, decide: "Into which unexplored future value do we reinvest the time freed up?"

For HR. The PD is not a wholesale overhaul of the HR system. It is a two-tier design that leaves the contract-and-control layer (the JD) in place and sets a purpose layer above it.

What can start tomorrow comes down to three things.

1. Leave existing evaluation and compensation unchanged, and start with a learning-oriented pilot of one to three teams. That is the paper's recommendation.

2. Vary the intensity by job. In standardized work, stop at a simple team PD; do not create individual PDs. In safety-critical domains, keep the PD concise and make the lower tier — JDs, procedures, halt authority — thick.

3. Retrain middle managers from transmitters into translators. Not "transmitters of policy from above," but translators who surface contradictions between KPIs and PDs and connect them to those responsible for the decision.

When KPIs and PDs collide, the burden of resolving it must not fall on individual managers. Providing an institutional place to escalate the collision is HR's job.

For the front line. The PD does not conflict with job crafting — workers reshaping the scope and meaning of their own work at their own discretion.

But individuals cannot rewrite the PD unilaterally. They submit change proposals grounded in frontline evidence; the effects are made visible; and the role's owner decides whether to approve, run a time-boxed experiment, modify or reject. The design preserves frontline autonomy while preventing unauthorized abandonment of roles, shifting of responsibility and local optimization.