# A job can be decomposed. Responsibility cannot.

Workday is breaking jobs into tasks to show what AI should automate, augment or leave to people. New enterprise-agent research shows the limit: local actions can obey their rules while the whole decision fails. Organisations need a decision spine that keeps purpose, evidence, authority, consequence and remedy connected across people and agents.

**AyEye — Workforce Management · 2026-10-15**

A job can be broken into tasks. A task can be matched to a skill, assigned to a person, handed to an agent or removed altogether.

That decomposition is useful. It is also incomplete. The organisation still has to hold together the purpose of the work, the evidence considered, the policy boundary, the authority to act and the remedy when several individually sensible steps produce one harmful outcome.

As HCM platforms make work more divisible, the scarce capability may become the one thing a task model cannot safely distribute: responsibility for the whole decision.

## The executive brief

- Workday has introduced Work Intelligence. The early-access product breaks jobs into tasks, connects tasks to skills and is intended to show where AI should automate, augment or leave work to people. The supplier says its Skills Cloud is used by more than 3,450 customers and contains more than 75,000 canonical skills.

- A new enterprise-policy study finds that boundary cases need more than a fast classification. In one sealed benchmark, allowing a fixed model to reason before choosing whether to answer, ask, escalate, use a tool or refuse improved exact action selection by 17.3 percentage points. The work is narrow, model-specific and not yet field evidence.

- Multi-agent research shows how locally reasonable actions can violate an organisation-wide constraint. Separate agents may each stay within their own permissions while collectively exceeding a shared limit.

- Another study shows why retrospective logs may not be enough. In a sandbox, 7,200 logged episodes could not determine the effect of two proposed retry gates. The missing evidence had to be designed into saved states and targeted reruns.

- The CHRO/CIO implication is decision continuity. Decompose activity as far as useful, but keep an unbroken line from purpose through evidence and authority to consequence and remedy.

## A job can be decomposed. Responsibility cannot.

ORIGINAL SYNTHESIS · Confidence: medium-high · Horizon: now to 24 months.

### Signal one: workforce systems are moving from job architecture to task architecture

REPORTED FACT. Workday announced Work Intelligence on 13 October. It says the early-access product breaks each job into tasks, links those tasks to required skills and helps leaders decide whether to hire, build capability, automate or redesign roles.

Workday says the system can show how an employee's time shifts, which skills become more important and whether people move into new work. It connects with Decision Intelligence and Workday Learning, while the company says people rather than the tools make employment decisions. Workday also says Skills Cloud is used by more than 3,450 customers and is based on more than 75,000 canonical skills.

These are supplier descriptions of an early-access product, not independent evidence that its task inferences are accurate or that customers will improve work, careers or productivity. [Workday, 13 October 2026](https://newsroom.workday.com/2026-10-13-Workday-Introduces-Work-Intelligence-to-Help-Customers-Design-Work-for-AI)

ANALYSIS. The significant change is representational. A job title is a relatively stable container. A task model turns work into smaller units that can be observed, compared and reallocated continuously.

That can expose hidden work and make job redesign far more precise. It can also encourage a subtle category error: because activity can be separated, responsibility can appear separable too.

A recruiter may source candidates, an agent may screen them, a manager may select a shortlist and finance may approve the position. Yet the consequential decision is not any single task. It is whether this organisation should add this kind of capacity, under these constraints, with these effects on current people and future capability.

### Signal two: policy boundaries are not ordinary tasks

REPORTED FACT. Researchers submitted a 43-page preprint on 13 October asking when enterprise policy decisions should be learned and when they should be reasoned through.

An enterprise agent first has to choose an action such as answer, request a missing fact, ask for confirmation, use tools, escalate or refuse. Holding one 27-billion-parameter Korean model fixed, the researchers compared deployment paths and reasoning conditions across 2,800 in-distribution items and a sealed 300-item benchmark derived from a hidden executable policy.

On the sealed benchmark, reasoning before action improved exact selection by 17.3 percentage points, with a reported interval of 12.3 to 23.0 points. Boundary-targeted supervision improved two covered boundaries but did not improve a held-out one. Some interaction results did not replicate under a later open-model rendering, and the authors do not claim stable cross-language transfer for all effects. [Han and colleagues, 13 October 2026](https://arxiv.org/abs/2610.17643)

ANALYSIS. The study does not establish that every HR agent should expose a chain of thought, nor that one model and benchmark generalise to employment decisions.

Its organisational lesson is narrower and more useful. The difficult part of policy is often not the rule at the centre. It is the boundary where incomplete facts, competing duties or an unusual consequence determine whether the work should continue, pause or move to someone with different authority.

Those boundaries are not simply more tasks. They are moments when the institution decides what kind of case it is dealing with.

### Signal three: local compliance can still create global failure

REPORTED FACT. The EAIR research project, submitted on 10 October, studies interaction risk in enterprise multi-agent systems. Its core problem is that agents operate with different goals, permissions and local information. Several service agents might each issue compensation within their individual limit while collectively exceeding a shared organisational cap.

The researchers created executable micro-worlds with roles, policies, permissions, facts and traces, then used those traces to diagnose interaction boundaries, repair workflows and retest them. They report recurring interaction failures across the organisational properties studied. The evidence comes from controlled environments, not production enterprises, and the phrase “first framework” is the authors' claim. [Ma and colleagues, 10 October 2026](https://arxiv.org/abs/2610.14800)

ANALYSIS. This is the machine equivalent of a familiar organisation-design problem. Every team can meet its target while the customer journey fails. Every function can follow its process while the employee receives no remedy. Every task can be green while the decision is wrong.

Task decomposition increases the number of boundaries where context can be lost. Agentic execution increases the speed at which those locally plausible actions can accumulate.

### Signal four: evidence must be designed before the change

REPORTED FACT. A separate preprint, agent-risk, asks how an organisation can assess a change to a deployed agent using its existing logs.

In the authors' bash-sandbox experiment, a 7,200-episode log alone could decide neither of two retry-gate changes because the proposed gates concerned actions the log had never observed. Targeted reruns from saved states reached both decisions using half the agent calls of a post-hoc live test sized for the comparison. When per-step harm was absent from the record, the certified harm ceiling for a stopping gate rose by as much as 2.5 times.

The setting is synthetic and depends on assumptions including independent episodes and limited behavioural drift. It nevertheless demonstrates a hard information problem: a log cannot answer a question it was never designed to preserve. [Zhao and colleagues, 12 October 2026](https://arxiv.org/abs/2610.16280)

ANALYSIS. Workforce systems face the same retrospective weakness. A record may say who approved a requisition, which skills were matched and which policy fired. It may not preserve which alternatives were considered, which uncertainty changed the judgement, whose workload was transferred or what harm occurred before a stop.

Observability is therefore not the same as memory. A task trace records movement. Decision evidence preserves why the movement remained legitimate.

### The unexpected connection

Task-level workforce intelligence + policy-boundary reasoning + multi-agent interaction risk + evidence designed before change

These developments come from HCM product design, model evaluation, multi-agent systems and safety engineering. Together they expose a limit to the task as the new basic unit of work.

ORIGINAL SYNTHESIS. A task is a unit of activity. A decision is a relationship among purpose, evidence, policy, authority, consequence and remedy. When work is decomposed, that relationship needs a decision spine: an unbroken organisational thread that keeps the parts answerable to the whole.

The decision spine does not require one person to perform every step. It requires the organisation to preserve:

- the outcome the work is meant to serve;

- the evidence and uncertainty that shaped the choice;

- the policy boundary and the person authorised to interpret it;

- the cumulative effect of several local actions;

- the person who can stop, reconsider and remedy the outcome.

Without that spine, task intelligence can produce organisational amnesia at high resolution. The enterprise knows precisely what happened in each step and cannot explain why the whole result was acceptable.

PROVOCATION: Do not automate a task until you know what responsibility was hiding between the tasks.

Not every hand-off contains profound judgement. Many are delay, duplication or clerical friction and should disappear.

But some apparent friction is where conflicting goals are reconciled, weak evidence is challenged or a person's circumstances become visible. Remove the hand-off without relocating that integrative work and the organisation has not simplified the decision. It has deleted part of it.

### A worked example: “I need another person”

A manager tells a conversational assistant, “I need another person.” The request can be decomposed into useful tasks: identify whether the need is replacement or growth, forecast demand, check budget, find adjacent skills, examine internal capacity, test automation potential, define the role, open a requisition and route approvals.

Each component may be correct in isolation:

- finance confirms that a budget line exists;

- workforce intelligence identifies a skill shortage;

- an agent predicts that part of the workload can be automated;

- talent systems find no immediately available internal candidate;

- the recruitment workflow opens a vacancy.

The whole decision can still be poor. The forecast may ignore a product closure. The internal candidate may need one protected assignment rather than already possess the full skill. Automating the routine work may remove the experience through which the next person learns. A neighbouring team may have spare capacity that no task model can see. The manager may be responding to a temporary surge or an unresolved performance problem.

The decision spine retains the original business outcome, the options rejected, the capability consequences, the budget holder's authority and a named human who decides whether recruitment is the right intervention. Only then does orchestration create the requisition.

This is not a defence of slow approval chains. It is a way to remove unnecessary hand-offs while preserving the one accountable judgement the organisation still needs.

### What if we are right?

Opportunity. Task intelligence could make work more adaptable without turning organisations into collections of micro-transactions. Leaders could move routine activity between people and agents while protecting the few decision relationships that carry purpose, fairness and consequence.

Organisational consequence. Job architecture would become more dynamic, but every consequential workflow would also have a named whole-decision owner. HCM, workflow, policy and agent telemetry would connect through the Workforce Context Graph rather than remain separate records.

Likely horizon. One workflow can adopt a decision spine now. Shared data structures that join task, policy, evidence, authority and remedy are plausible within 12–24 months, particularly where HCM and enterprise-agent platforms already share identity and process context.

### What would prove us wrong?

The thesis weakens if most task-level reallocations remain low consequence, reversible and fully governed by simple rules. It also weakens if modern orchestration reliably preserves global constraints and produces evidence that people can interpret without reconstructing the whole case.

The idea fails if “decision spine” becomes a reason to centralise everything, retain pointless approvals or appoint an executive owner who never sees the evidence. A visible spine without usable authority is governance theatre.

The operational test is comparative. If decomposed workflows produce equal or better decisions, fewer unowned exceptions, faster remedy and no loss of contextual understanding, existing controls may already preserve the whole adequately.

### A constructive possibility: work becomes modular without becoming meaningless

Good decomposition can liberate people from inherited job containers. Someone could contribute a specialist skill across several teams, learn through progressively harder cases and work with agents that handle preparation and administration.

The human contribution would not be “everything the machine cannot do”. It would be explicit: frame the outcome, interpret contested evidence, exercise authorised judgement, connect local actions and repair consequences.

The prize is not a perfectly atomised workforce. It is an organisation that can move activity freely while keeping meaning and responsibility intact.

## Revisited: the Context Graph needs to remember decisions, not only entities

On [30 September, Workforce introduced the Workforce Context Architecture](https://www.lecxie.com/publications/ayeye-workforce/2026-09-30.html): eight capability layers and three control rails centred on a graph linking purpose, work, tasks, skills, people, agents, roles, organisation, policy, systems and data.

Today's evidence strengthens that model and adds a missing relationship. The graph should link not only entities and activities but also the decision that justified a consequential transition: which evidence was considered, which policy boundary applied, who possessed authority and who owns remedy.

On [7 October, Workforce argued that inferred, demonstrated and authorised capability must remain distinct](https://www.lecxie.com/publications/ayeye-workforce/2026-10-07.html). Work Intelligence makes the opportunity and risk more concrete. A task model can reveal hidden adjacency, but it must not silently turn an inference about work into an authorised decision about a person.

The earlier judgements are confirmed, not overturned. The update is that context needs a temporal spine: not just what is connected, but how one authorised choice became the next.

## A market for tasks could hide the cost of integration

TENUOUS BUT PLAUSIBLE · Confidence: medium-low · Horizon: 18–48 months.

If work systems can represent tasks continuously, organisations may begin reallocating them more like a live internal market. Agents perform some tasks, employees bid or are matched to others, and managers buy temporary capability instead of filling stable roles.

The causal chain is plausible:

Task visibility improves → allocation becomes dynamic → jobs lose some of their bundling function → coordination and whole-case judgement move into fewer integrator roles → those roles become a hidden bottleneck.

This could broaden opportunity and make work more flexible. It could also leave individuals with fragmented assignments, weak development narratives and accountability for outcomes assembled elsewhere.

What to watch: whether task marketplaces measure integration load, whether people can refuse incoherent bundles, whether developmental work is funded and whether the owner of the whole decision remains visible when tasks cross organisational boundaries.

## Operating-model implication

### Give consequential workflows a decision spine

ElementQuestionEvidence retainedNamed authority

PurposeWhat outcome are we trying to change?Need, affected groups and alternativesBusiness owner
EvidenceWhat do we know and what remains uncertain?Sources, assumptions, dissent and missing factsCase owner or qualified reviewer
BoundaryWhich policy or ethical threshold changes the route?Applicable rule, interpretation and escalation triggerAuthorised policy owner
Whole effectWhat do the local actions add up to?Cumulative spend, workload, opportunity and affected peopleEnd-to-end process owner
RemedyWho can stop, reconsider and repair?Interventions, reasons, outcomes and unresolved residueNamed human with real decision rights

The spine is not another approval layer. It is the minimum connective evidence required to let distributed work remain one accountable decision.

## Human control watch

Assistance: AI helps a person understand tasks or skills. Control depends on whether the person can inspect the source evidence and reject the inference.

Delegated execution: agents complete bounded tasks. Control depends on preserving the purpose and cumulative effect across hand-offs, not merely reviewing each action.

Autonomous control: systems continuously reallocate work or execute policy. Control depends on a decision spine with global constraints, pre-designed evidence, intervention authority and remedy.

Today's shift: human oversight must move from approving isolated steps to owning the coherence of the whole decision.

## Capability-model update

Gaining valueUnder pressure

Whole-decision ownerTask owner treated as complete accountability
Policy-boundary interpreterRules assumed to cover every edge case
Interaction-risk designerLocal compliance used as proof of system safety
Decision-evidence architectLogs designed only after an incident
Integration-load assessorTask efficiency that ignores coordination work
Context-graph stewardJob and skills taxonomies without temporal decisions

## IF I WERE TO DO ONE THING NOW

### Name one whole decision

TASKS → EVIDENCE → BOUNDARY → WHOLE EFFECT → REMEDY

Return to the workflow where you protected one capability nursery in the previous exercise, or choose one live HR workflow now being decomposed for AI. Ask the process owner to name the single whole decision that the tasks collectively produce, then write beside it one person who can change that decision and one piece of evidence they would need to do so. Test the answer against a real recent exception. Do not create a governance framework or add an approval stage. Find one place where the tasks are visible but the whole decision is not, and make that responsibility explicit this week.

## Mental-model update

The regenerative enterprise has learned to form, practise, preserve and recover capability rather than merely allocate it.

Today it acquires connective tissue.

Work can become modular without responsibility becoming fragmentary. A decision spine lets tasks move among people and agents while purpose, evidence, authority and remedy remain continuous.

The emerging North Star is a regenerative and coherent enterprise: flexible enough to decompose work, disciplined enough to remember what the pieces mean together.

## Questions for the executive table

- Which consequential decision in your organisation is currently represented only as a collection of completed tasks?

- Where can several individually compliant agent actions create an unacceptable cumulative effect?

- Which policy boundary requires interpretation rather than classification, and who is authorised to provide it?

- Could today's logs explain why a decision was legitimate, or only reconstruct which steps occurred?

- When jobs become dynamic bundles of tasks, who protects the integration work and developmental experiences that no task owner sees?

Evidence note. Workday describes an early-access product and its own customer and ontology figures; no independent outcome evidence is available here. The policy-decision, interaction-risk, agent-change and exact-action-control studies are recent preprints using bounded models, synthetic environments or sandbox targets. They do not establish prevalence or production effectiveness. The concepts decision spine, decision continuity and organisational amnesia at high resolution are original AyEye analysis, not claims made by the cited sources.

Canonical: https://www.lecxie.com/publications/ayeye-workforce/2026-10-15.html