Who gets to decide what your work means?

A system that describes your work can eventually decide which work you are offered. That is a substantial change in the relationship between an employee and an organisation, even when the description begins as a helpful skills recommendation.

SAP's proposed acquisition of TechWolf moves an evolving model of tasks and skills towards the centre of a major HCM portfolio. On the same day, cybersecurity training and contracting research offered a different kind of evidence: demonstrations of whether agents and the people directing them can actually perform bounded work. The opportunity is to connect these developments without confusing their authority. An inferred skill, a successful demonstration and permission to act are three different things.

The executive brief

• SAP has agreed to acquire TechWolf. Its context graph for work could become an important grounding layer for SuccessFactors workforce decisions. The deal has not closed and the planned products are not completed integrations.
• Hack The Box is appraising both sides of cyber work. Its new enterprise range assesses agents against roles, while operator exercises examine human judgement about when to question, intervene or let an agent continue.
• OpenAI and Ironclad have made professional work into an evaluation. Their task-specific research shows progress alongside substantial unmet requirements. A rubric score cannot be translated directly into a percentage of jobs automated.
• Enterprise suppliers are competing to define organisational context. This raises a strategic question for HR and technology: who can correct the model of work before it influences people's opportunities?

Who gets to decide what your work means?

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

Signal one: a context graph is becoming a strategic enterprise asset

REPORTED FACT. On 6 October, SAP and TechWolf announced an acquisition agreement, with closing expected in Q4 2026 subject to customary conditions and regulatory approval. Financial terms were undisclosed. SAP says TechWolf uses connected HR and business systems to model tasks within jobs, applied skills and the external labour market against business strategy. It expects the platform to become an intelligent core of SuccessFactors, supporting workforce planning, skills-based hiring and organisational redesign. New joint products would be developed after closing. Current plans, subject to closing and consultation, retain TechWolf as an independent entity and keep its platform available to non-SAP customers. These are intentions and supplier descriptions, not independently verified decision outcomes. SAP announcement, 6 October 2026

ANALYSIS. The strategic purchase is a representation of work, alongside technology and expertise. A continuously refreshed account of tasks and skills can make internal talent visible where a job title cannot. It can also become the lens through which leaders see the organisation. Once that lens feeds decisions, errors in the representation can have consequences well beyond an inaccurate profile.

Signal two: cybersecurity is testing the agent and the operator

REPORTED FACT. Hack The Box launched AI Range Enterprise Edition on 6 October. It describes repeatable, role-based appraisal of security agents, with environment-level pass/fail results and performance tracked over time. Separately, its AI-augmented penetration-testing and SOC analyst roles assess practitioners' judgement: recognising weak reasoning, understanding recommendations, intervening when necessary and allowing sound work to proceed. The methodology and scoring are proprietary supplier offerings; the announcement does not establish predictive validity across enterprise operations. Hack The Box announcement, 6 October 2026

ANALYSIS. This changes the shape of competence evidence. A person may be capable of the underlying task yet poorly prepared to supervise an agent doing it. An agent may perform well alone but fail when its output becomes a recommendation that a busy person must challenge. Workforce design needs evidence about the relationship, as well as each participant.

Signal three: a professional task exposes the distance between progress and readiness

REPORTED FACT. OpenAI described its Ironclad collaboration on 6 October. Practitioners helped select 11 legal, commercial and procurement tasks, each judged against 8–50 criteria. Hosted software environments supported practice and evaluation. GPT-6 Astra achieved a mean rubric score of 55.0%, compared with 41.6% for GPT-5.6 Sol. Estimated time per attempt fell from 37.0 to 19.2 minutes. These are research tasks and simulated timing estimates, not measured customer savings or whole-workflow success rates. The comparison also used different reasoning settings. OpenAI research account, 6 October 2026

ANALYSIS. Faster partial completion may be useful. It is not a release decision. A missing approval condition can matter more than several correctly configured fields. The operating owner must decide which requirements are critical and what residual work is safe to delegate. Averaging them together can conceal the condition that makes the process unacceptable.

The unexpected connection

SAP is bringing a model of work closer to workforce decisions. Cybersecurity is building appraisals of mixed human–agent capability. Contracting research is defining success through practitioner-authored criteria. Together they suggest a new organisational boundary: the boundary between a claim about capability and a decision based on it.

The Workforce Context Architecture put Workforce Knowledge & Semantics at its centre. Yesterday's capability corridor created room for a person to learn an adjacent task through real work. Today's evidence adds a discipline between those ideas: the context graph must remember how a capability claim was earned, challenged and authorised.

Three states of a capability claim

ORIGINAL SYNTHESIS. Keep three states separate in the Workforce Context Graph. These are proposed management distinctions, not an industry standard or a claim about any supplier's implementation.

Claim and evidence	What remains open
Inferred Available traces suggest a person or agent may have a capability.	Whether the traces represent their contribution, proficiency and current conditions.
Demonstrated A bounded task was performed against explicit criteria in recorded conditions.	Whether the result transfers to different stakes, systems, collaborators or constraints.
Authorised A named owner permits defined work under a mandate and control conditions.	Whether that permission remains justified after the work or system changes.

These states can coexist. A person may have demonstrated a skill but lack permission to exercise it in a regulated process. An agent may be authorised for a reversible update while its ability to handle an unusual exception remains uncertain. Someone with little digital evidence may have deep competence that the graph has not yet recognised.

A capability record should therefore retain its source, date, task conditions, contributor and reviewer, together with disagreement or limitations. Confidence is helpful but cannot replace those particulars. A confident inference about writing a document does not establish who resolved its hardest judgement.

For an employee, the record needs a route to correction. For an agent, it needs a configuration and version. For the mixed workflow, it needs the allocation of judgement: who set the objective, who produced the result, who challenged it and who accepted the remaining risk.

PROVOCATION: A work graph must not become a verdict on a person.

Making work legible can expand opportunity. Treating the machine's account as complete can close it. The employee least visible in connected systems may be the person who prevents failure, teaches others or negotiates an exception before it ever becomes a ticket. Conversely, a prolific record may reflect delegation, copied work or a role with unusually good data capture.

HYPOTHESIS. As inferred capability enters allocation and development decisions, the right to correct a work representation will become as practically important as the right to correct a personnel record. This is an emerging operating-model issue, not a statement that a particular law creates that right everywhere.

What if we are right?

Opportunity. People with overlooked skills could become visible without writing a new CV every quarter. Demonstrations could give an employee a fair route beyond their title. Agents could earn wider mandates through evidence rather than supplier reputation. HR could connect learning, redeployment and work design through a shared account that remains open to challenge.

Organisational consequence. Work intelligence becomes a jointly stewarded representation. HR owns fair use and access to opportunity; process owners define meaningful task criteria; technology preserves provenance and configuration; workers contribute missing context. None can establish competence alone.

Horizon. The three distinctions can be used in one workflow now. System-wide integration is plausibly a 12–24 month effort, dependent on data quality, worker participation and suppliers' implementation. The acquisition does not establish that this model will be built.

What would prove us wrong?

The argument weakens if work graphs remain advisory discovery tools and never meaningfully affect allocation, hiring or promotion; if existing manager review reliably corrects inference errors before decisions; or if employees find the representations accurate, contestable and useful without additional structure.

It also fails if the proposed evidence states add administration without changing a decision. A local test should ask whether exposing provenance and disagreement changes who gets an opportunity, what an agent may do or how a task is redesigned. If nothing changes, simplify the record rather than mandate more documentation.

A constructive possibility: recognition becomes a conversation

The best work graph could give people a better way to say, “That is only part of what I contribute.” A manager can ask for an example; a colleague can identify invisible work; a task demonstration can resolve uncertainty. The representation becomes useful because it can improve, rather than because it pretends to be final.

This could also reduce defensive appraisal. Instead of arguing about a broad label such as “AI-ready”, a team can discuss a specific task, the conditions under which it was done and the next permission or learning opportunity. Better evidence creates more precise encouragement.

Context is becoming an enterprise product category

OUTSIDE-IN SIGNAL. Infor announced its expanded Industry AI architecture on 6 October, including a semantic layer, coordinated agents and governance controls. It also published commissioned research conducted by YouGov in August among 2,111 decision-makers across seven markets. Infor reports that 68% of businesses across six surveyed markets considered generic AI inadequate for industry needs and that 54% of leaders globally were comfortable with agents executing critical processes without human input at every step. These are respondent attitudes and a supplier's interpretation, not independent evidence of safe autonomy. Infor announcement and methodology, 6 October 2026

ANALYSIS. A market is forming around who can supply the organisation's meaning to its models. That can reduce integration effort, but it also makes semantic ownership a procurement issue. A buyer should be able to understand and change task definitions, mappings and exceptions, and retain a usable account when the platform changes. Portability means preserving relationships and provenance, not simply exporting a spreadsheet of skill names.

Training volume is not an evidence state

IMPLEMENTATION SIGNAL. Tech Mahindra announced on 6 October that its Google Cloud partnership would roll out hands-on Build with Gemini workshops to more than 12,500 associates. Participants practise creating, securing and governing agents and earn a skill badge. The announcement describes a rollout and learning design, not demonstrated outcomes for every participant. Tech Mahindra announcement, 6 October 2026

ANALYSIS. Practical learning is promising. A badge can record an achievement while leaving its scope explicit. A workshop completion should not silently become permission to alter a critical process. The useful bridge is a subsequent bounded demonstration in the work that a person will actually inherit.

The work graph could become a self-fulfilling labour market

TENUOUS BUT PLAUSIBLE · Confidence: medium-low · Horizon: 12–36 months. This is a risk hypothesis, not a reported behaviour of the products discussed.

If inferred skills determine who receives interesting tasks, those people generate the next round of evidence. Their profiles strengthen; colleagues overlooked initially have fewer chances to demonstrate an alternative. A seemingly neutral recommendation can gradually manufacture the pattern it claims merely to describe.

The causal chain is: incomplete work traces shape an inference; the inference shapes allocation; allocation changes the evidence available; the next inference reinforces the first. It depends on recommendations materially influencing opportunity and on weak correction routes. Neither is inevitable.

Watch for repeated concentration of developmental tasks, profiles that never change despite employee challenge and apparent capability gaps that disappear when people receive comparable opportunities. The optimistic alternative is to use uncertainty to open a trial assignment rather than exclude someone. A graph can expose an evidence gap without treating it as a human deficit.

Operating-model implication

Put the evidence boundary inside the architecture

The eight-layer Workforce Context Architecture does not need rebuilding. It needs a clear separation between its knowledge layer and its decision rails.

Architecture responsibility	Required behaviour
Workforce Knowledge & Semantics	Link the capability claim to the task, outcome, evidence and unresolved challenge.
Data & Context	Preserve origin, freshness, coverage and permitted use of the underlying trace.
Work & Workflow Orchestration	Carry the evidence and conditions into allocation, development or escalation.
Human Accountability & Decision Rights	Make permission an explicit decision by an owner with actual authority.
Trust, Security & Governance	Limit access and secondary use; provide an intelligible correction route.
Quality, Observability & Improvement	Reassess after material changes and learn from disagreements and failed demonstrations.

An employee-built agent inherited by a new role holder needs the same discipline. A successful demonstration belongs to a particular configuration and operating context. Reassignment should preserve the instructions, integrations, evidence and mandate while testing whether the new incumbent can challenge the workflow. The role inherits a digital capability, together with its limitations, rather than a mysterious personal tool.

Human control watch

Assistance. Software suggests a capability or learning path. The person can inspect the basis, add context and reject a poor description.

Delegated execution. A demonstrated person–agent workflow receives bounded permission. Review covers critical conditions and exception handling, not only average task quality.

Autonomous control. A system allocates work or changes consequential records. The institution must preserve challenge, intervention and remedy before inferred capability becomes a controlling decision.

Today's shift: control includes the ability to challenge what the system believes about work, as well as the ability to stop what it does.

Capability-model update

Gaining value	Under pressure
Work-evidence steward	Skill labels detached from source and conditions
Task-criteria designer	Average scores used as release decisions
Human–agent operator judgement	AI literacy treated as sufficient supervision
Semantic portability and correction	A supplier's representation treated as organisational truth
Opportunity-aware workforce planning	Missing evidence treated as missing ability

Mental-model update

The permeable enterprise has acquired authority, continuity, consequence limits and routes for people to grow through changing work. Today it acquires a way to distinguish knowledge from permission.

A Workforce Context Graph can become the organisation's working memory, but its claims must remain revisable. The North Star is an enterprise that sees more capability while preserving the freedom to challenge its account of a person, an agent or a role.

Questions for the executive table

1. Which inferred skill can already influence someone's opportunity, and can that person correct its basis?
2. Where are demonstrated competence and permission to act currently collapsed into one label?
3. Which indispensable contribution leaves the weakest trace in our systems?
4. Can we move our work definitions, evidence and exceptions to another platform without losing their meaning?
5. When a role inherits an agent, who establishes that the new incumbent can supervise it?

Evidence note. The acquisition is pending. All five current signals come from organisations describing their own products, research or programmes; none establishes enterprise-wide outcomes. OpenAI's evaluation covers 11 research tasks and simulated time estimates. Infor's commissioned survey records attitudes, not operational safety. The three capability states and the self-reinforcing allocation hypothesis are original AyEye analysis. No supplier is alleged to be misusing employee data or making unchallengeable employment decisions.