TLDR
Executive briefThousands trained, one consequential workflow still unable to operate
- Plan workforce capability around a business outcome and work event, not a catalogue of future skills or the number of assistant licences activated.
- Combine six layers: literacy, occupational craft, human-AI practice, team design, product and assurance, and leadership and institutional capability.
- Represent operating capability as connected cells with authority, evidence, capacity, support, sourcing and refresh, including shared-service dependencies.
- Progress proof from awareness through protected practice and supervised production to sustained outcome, risk, worker and learning evidence across change.
- Stop scaling when reviewer, risk, incident or specialist capacity cannot support the proposed volume, even when users are enthusiastic and pilots look strong.
An enterprise has trained several thousand employees, created prompt libraries and reported high adoption of an approved assistant. Yet its first consequential customer workflow cannot scale. Novices improve quickly, but complex cases require experts whose queue is already full. Knowledge articles decay without an owner. Product staff can demonstrate an agent but cannot run incidents. Risk challenge arrives after release. Supervisors cannot tell which decisions remain human. The organisation has individual tool exposure without the capability to operate the outcome.
An AI-augmented workforce is not a workforce in which everyone uses an assistant. It is a system of people, roles, teams, workflows, specialist services, controls and learning loops that can repeatedly produce a defined result as tools and tasks change. Individual literacy is necessary, and Paper 131 specified it by role and consequence. It cannot substitute for domain craft, coordination, workflow ownership, engineering, assurance, capacity or leadership judgment.
The unit of planning should therefore be a Capability Cell: a bounded outcome and work event, the human and AI work required, authority, proof, enabling environment, capacity, sourcing choice and refresh trigger. A portfolio of cells reveals what can be built in roles, supplied by a guild, bought as a commodity service, embedded temporarily, co-developed with a partner, automated after redesign or retired. A graph then exposes shared evaluators, data stewards, supervisors, platform teams and risk challengers that several cells depend on.
Official and empirical evidence supports task-sensitive workforce design. UK sources distinguish role levels and connect upskilling to organisational adoption, practical work and workforce planning.123 ILO analysis emphasises occupational variation, job transformation, human input, job quality and social dialogue rather than a simple replacement story.56 OECD evidence links adoption barriers to technical, cognitive, social, managerial and organisational capability.79 Field research shows heterogeneous gains and altered expertise or work patterns, so one experiment should inform a mechanism, not become a forecast for every occupation.131415
The Six-Layer Augmented Workforce Capability Model, Capability Cell, Workforce Capability Graph, Capability Sourcing Portfolio, Team Composition Test, Capability Half-Life Calendar and Augmented Workforce Capability Studio are this article's synthesis. They differ from Paper 027's adoption-adjusted value, Paper 040's change-bearing architecture, Paper 044's product ownership and Paper 131's literacy qualification. This paper builds the enterprise capability portfolio that makes those mechanisms operable.
Six-Layer Augmented Workforce Capability ModelExpand a licence count into the capability that carries the outcome
The first layer is foundation and literacy: recognising AI, judging the task, checking evidence, protecting data, using approved channels and escalating. It establishes a safe common language and the role-specific permission described in Paper 131. The second is domain and occupational craft: the customer, financial, clinical, legal, engineering or operational knowledge needed to frame the problem and judge whether an answer is useful.
The third layer is human-AI work practice: specification, iteration, verification, exception handling, handoffs, human review and appropriate refusal inside the configured task. The fourth is team and workflow design: complementary roles, queue design, authority, workload, end-to-end ownership, worker voice, adoption and learning. A proficient person can still fail inside a workflow whose target punishes checking or whose exception queue has no capacity.
The fifth layer is product, engineering and assurance: data, model, evaluation, platform, reliability, security, risk, audit, incident, change and retirement capability. It turns a demonstration into a persistent service. The sixth is leadership and institutional capability: strategy, portfolio choice, workforce transition, investment, governance, labour dialogue, benefit capture and organisational learning. It decides which outcomes deserve augmentation and which capabilities contain strategic context.
The layers are not maturity levels. Each cell selects a combination. A commodity drafting assistant may rely on strong foundations and centrally supplied security. A consequential finance cell needs deep domain craft, independent challenge and product assurance. A team may use a managed external model while retaining evaluation, data stewardship, workflow ownership and benefit capture internally. Missing capability in one layer can invalidate strength in the others.
| Layer | Operating question | Representative proof |
|---|---|---|
| Foundation and literacy | Can people recognise, judge, check, protect and escalate within authority? | Role-specific scenario evidence and approved-use behaviour |
| Domain and occupational craft | Can the work be framed and judged using real context and consequence? | Observed domain decisions, exceptions and affected-party reasoning |
| Human-AI work practice | Can people specify, verify, refuse and hand off under work conditions? | Protected and supervised performance on representative cases |
| Team and workflow design | Do roles, queues, authority, workload and learning combine into an outcome? | Flow, capacity, handoff, worker-experience and outcome evidence |
| Product, engineering and assurance | Can the service be evaluated, secured, operated, changed and retired? | Production controls, incidents, evaluation and recovery receipts |
| Leadership and institutional capability | Can the enterprise choose, fund, govern and learn across a portfolio? | Investment, transition, labour dialogue and benefit decisions |
Capability CellBind skills, capacity and evidence to a decision about real work
A Capability Cell is this article's synthesis. It names an outcome and work event; users and affected population; human tasks and AI tasks; roles and decision authority; required six-layer capabilities; volume, scarcity and capacity; proficiency evidence and assessor; tools, data, workflow and specialist support; sourcing choice; protected learning path; performance, quality, risk and worker evidence; half-life, change trigger and accountable owner.
The cell forces three distinctions. A skill is an ability, such as checking evidence. Capacity is enough qualified time to perform that checking at the expected volume. Capability is the whole configuration that produces the outcome, including access, authority, evidence, support and recovery. A team may contain skilled reviewers but lack capacity. It may have capacity but no right to inspect supplier evidence. It may have both but no incident route. The cell is incomplete until the configuration works.
Cells should be small enough to change an investment or design decision. “Marketing AI” is too broad. “Prepare a regulated campaign claim for legal approval using the approved evidence assistant” can specify population, authority, workload, reviewers, proof and stopping conditions. Cells can be composed into a larger service, but each retains an owner and a measurable work event.
A cell also makes residual human work visible. Automation can reduce first-pass production while increasing exceptions, review time, escalation, coordination or evidence custody. Capacity planning uses the residual queue and its variation, not the percentage of tasks labelled automated. Worker experience covers autonomy, workload, clarity, voice, learning and the ability to challenge the system.
| Cell object | Required declaration | Decision changed |
|---|---|---|
| Outcome and work | Event, users, affected population, human and AI tasks, exceptions and handoffs | Scope and redesign |
| Authority and layers | Decision rights and the six-layer capability depth required | Team composition and control |
| Evidence and capacity | Proof level, assessor, volume, queue, scarcity, concentration and succession | Scale, staffing and stop gate |
| Environment and sourcing | Tools, data, workflow, specialist support and build, borrow, buy or partner route | Investment and contracting |
| Learning and refresh | Practice, performance, risk, worker evidence, half-life, triggers and owner | Renew, narrow, transfer or retire |
Workforce Capability GraphExpose the missing shared service behind an apparently trained role
The Workforce Capability Graph is this article's synthesis. Nodes are capability cells, shared specialist services and enabling assets. Edges show prerequisite knowledge, workflow handoffs, evidence dependencies, common tools, shared reviewers and concentration. A graph reveals why qualifying a frontline role does not create an operating capability when evaluation, data stewardship, expert escalation, platform support or independent challenge remains absent.
Start from a cell and trace every dependency required at expected volume. A customer recommendation cell may depend on knowledge stewardship, multilingual expertise, outcome evaluation, complaint handling, platform monitoring and a supervisor queue. Several cells may share the same evaluators or incident team. The risk is not visible in an individual skills inventory: one scarce node constrains the portfolio and produces correlated failure.
Edges need operating terms. A handoff includes trigger, information, service time, capacity and return path. An evidence dependency includes artefact, owner, quality and expiry. A shared reviewer edge records demand variability and priority rules. A partner edge identifies knowledge and learning rights. The graph becomes useful when it changes funding, staffing, sequencing or scope.
Concentration is both a person and system issue. One expert may become the only bridge between a vendor model and a regulated process. One guild may support too many launches. One platform team may own every action boundary. Succession, rotation and reusable artefacts reduce fragility, but only if replacements can perform the work under observed conditions.
Capability Sourcing PortfolioChoose where context and learning should live
The Capability Sourcing Portfolio is this article's synthesis. Build enduring capability where proprietary context, repeated learning or decision authority creates advantage. Borrow temporary internal expertise through rotations or guilds while a domain forms its own practice. Buy a commodity service when boundaries, evidence, substitution and operating responsibility remain controllable. Embed and transfer scarce specialists inside a workflow with explicit knowledge-transfer and expiry terms.
Partner when both parties contribute scarce assets and learning rights can be shared. Automate or retire only after the outcome, exceptions, affected-party rights and residual human load are redesigned. The sourcing decision applies to a cell or dependency, not to “AI capability” as a whole. An enterprise can buy model hosting while building evaluation and domain stewardship, borrow an incident specialist and partner on a novel data asset.
Strategic context should not be outsourced by accident. If competitive learning depends on how the workflow frames exceptions, measures outcomes or adapts to customers, retain those feedback loops. Conversely, building a commodity model runtime can consume scarce engineering capacity without differentiating the outcome. The cell declares which knowledge must persist after a contract ends.
Embed arrangements need an exit. Name the artefacts, paired staff, observed tasks, learning rights, milestone for reduced involvement and capability owner. A consultant delivering strong output without transfer may improve this quarter's work while weakening next year's capability. The portfolio records that liability rather than calling it capacity.
| Route | Use when | Control term |
|---|---|---|
| Build | Context, authority and repeated learning create durable advantage | Protected practice, succession and outcome ownership |
| Borrow | A temporary internal bridge can seed repeatable practice | Rotation, artefacts, paired work and return date |
| Buy | The service is commodity and its boundary remains replaceable | Evidence, portability, operating responsibility and exit |
| Embed and transfer | Scarce expertise must work inside the domain while capacity forms | Learning rights, named recipients, milestones and expiry |
| Partner | Both parties contribute scarce assets and learn together | Decision rights, joint evidence and ownership of reusable learning |
| Automate or retire | Work can be removed after exceptions and residual load are redesigned | Outcome, rights, recovery and human-capacity evidence |
Team Composition TestDo not put every missing capability into one new AI role
The Team Composition Test is this article's synthesis. For each workflow, ask whether the operating group contains enough domain judgment, AI work practice, evidence evaluation, product or engineering support, risk challenge, operational ownership and change capacity. Then test whether those capabilities can meet actual volume, exceptions and service times. A list of names does not prove a team.
Complementarity matters more than novelty. A domain expert does not automatically provide evaluation or security. An engineer does not own the affected-party outcome. A risk specialist should not silently become the product manager. A new “AI lead” cannot absorb platform, data, workflow, assurance, workforce transition and benefit ownership without becoming a fragile junction. Existing roles, shared services and targeted new roles can be recomposed around the cell.
Coordination itself is a capability. Field evidence suggests AI can change expertise boundaries and team performance, while individually controllable work may shift more readily than meetings and approval-heavy patterns.1415 That distinction cautions against expecting assistant access to repair cross-functional handoffs. Shared evidence, queue rules, escalation and decision ownership must be redesigned.
Worker experience belongs in the test. Track autonomy, workload, role clarity, access, learning, speaking up and the distribution of exception work. Psychological safety supports error reporting, experimentation and learning, which become essential when capability boundaries move.18 High output paired with exhausted reviewers or suppressed challenge is not sustained capability.
Proof levels and Capability Half-Life CalendarMove from awareness to sustained evidence, then refresh on change
Capability proof has four levels. Trained and aware records exposure to concepts and boundaries. Demonstrated in protected practice shows performance on representative work without live consequence. Performed in supervised production proves that people, tools, handoffs and support function under real workload. Sustained across change requires outcome, quality, risk, worker-experience and learning evidence through material updates or incidents.
The levels prevent attendance from being mistaken for operating capacity. A cell can contain qualified individuals yet remain at protected practice because its supervisor queue is untested. A pilot can operate in production yet lack sustained evidence because demand, staff turnover and model change have not been absorbed. Investment and scale should follow the weakest evidence needed for the consequence.
The Capability Half-Life Calendar is this article's synthesis. Durable capabilities such as domain judgment, causal reasoning, evidence discipline and collaboration may persist across tools. Volatile product operation, interface navigation, prompt syntax and model-specific limits can decay quickly. Refresh is triggered by a material change in model, workflow, authority, population, policy, incident pattern, tool or evidence, not by the annual learning calendar alone.
Each cell lists its durable foundation, volatile elements, observation date, change triggers, owner and renewal action. A changed button may need guidance. A new action tool, protected population or decision authority can require supervised re-performance. A severe incident can suspend scale and reopen the team test. The calendar allocates learning capacity to real decay rather than repeating stable material for everyone.
| Level | Evidence | Scale implication | Refresh trigger |
|---|---|---|---|
| 1. Trained and aware | Concepts, boundaries and help routes understood | No consequential operating claim | Changed policy or common foundation |
| 2. Protected practice | Representative work and failure scenarios observed | Prepare supervised use | Tool, task or scenario delta |
| 3. Supervised production | Real workload, handoffs, capacity and outcomes observed | Bounded production expansion | Volume, role, authority or population change |
| 4. Sustained across change | Outcome, risk, worker and learning evidence survives change | Portfolio-scale operation with monitoring | Model, workflow, policy, incident or evidence shift |
Ten capability-cell casesLet the outcome, dependency and capacity determine the investment
Each case declares a Capability Cell, the six-layer requirement, graph dependencies, proof, sourcing, capacity, worker evidence, half-life trigger, owner and investment decision. The cases show why training or one new job title rarely closes the whole gap.
1. Customer support gains depend on the expert system
Cell: resolve standard customer enquiries with assistant suggestions while preserving expert escalation. Layers and graph: foundation, customer craft and checking sit with advisers; knowledge stewardship, multilingual evaluation, supervisors and feedback ownership are shared nodes. Proof and sourcing: supervised production shows novice gains, but sustained proof depends on complex-case outcomes. Build adviser practice and knowledge ownership; borrow evaluation support while transferring methods. Capacity and worker evidence: expert queue age, cognitive load and speaking-up routes determine scale. Half-life and owner: knowledge or model change triggers refresh; the service owner funds expert capacity and feedback operations. A new “AI adviser” role is not the answer.
2. Finance requires more than broad prompt fluency
Cell: prepare a material forecast with traceable assumptions and independent challenge. Layers and graph: finance craft, evidence evaluation, data stewardship, controlled work practice and risk independence dominate. Dependencies include source owners, reconciliation tooling and a challenge panel. Proof and sourcing: protected cases test stale inputs and uncertainty before supervised production. Build domain and evaluation capability; borrow specialist model review for the first cycle. Capacity and worker evidence: challenge time and deadline pressure are measured alongside accuracy. Half-life and owner: data, policy or model change reopens evidence; the finance director owns investment in provenance and reviewer capacity.
3. Legal buys model expertise but leaves matter checking hollow
Cell: use AI to prepare privileged matter analysis for accountable lawyer review. Layers and graph: legal craft, privilege handling, evidence checking, secure product controls and partner review are required. The purchased modeller does not supply matter judgment or review capacity. Proof and sourcing: buy technical advice temporarily, but build matter-specific checking and secure workflow ownership. Capacity and worker evidence: partner review time, junior learning and rework are explicit. Half-life and owner: new data access or legal duty triggers refresh; the practice leader funds secure evidence handling and protected learning rather than another generic course.
4. Prototype strength hides the production cell
Cell: operate a tool-using agent for a bounded internal service. Layers and graph: the product team has interaction and domain design, but depends on reliability, security, identity, incident, monitoring and retirement nodes. Proof and sourcing: the prototype remains at protected practice until supervised production exercises failure and rollback. Build product stewardship; borrow incident expertise and buy commodity runtime services under controllable boundaries. Capacity and worker evidence: on-call load and manual fallback are measured. Half-life and owner: tool or autonomy change invalidates prior proof; the product owner funds operating capability before scale.
5. An evaluation guild lends, codifies and withdraws
Cell: evaluate domain assistants against real failures before release. Layers and graph: the guild supplies evaluation craft, tooling and assessor calibration while domain teams retain outcome judgment. Proof and sourcing: borrow scarce expertise through paired studies, shared artefacts and rotations; prove domain teams can repeat the method before reducing guild involvement. Capacity and worker evidence: guild demand, domain learning time and independence are visible. Half-life and owner: new model or population triggers a focused refresh; the assurance leader invests in reusable cases and succession rather than permanent central ownership.
6. Embedded delivery without transfer creates a dated gap
Cell: redesign a claims workflow with an embedded partner. Layers and graph: the partner covers engineering and evaluation, while operations owns domain decisions and affected-party recourse. There are no paired staff or reusable artefacts. Proof and sourcing: production performance is strong, but institutional proof is weak. Continue only as embed-and-transfer with learning rights, named recipients and expiry. Capacity and worker evidence: internal shadow time and confidence to handle exceptions are measured. Half-life and owner: contract end is a mandatory trigger; the transformation owner funds transfer milestones or narrows scope.
7. Complementary team design beats siloed expertise
Cell: resolve complex service requests using domain, data and AI support. Layers and graph: a cross-functional group combines craft, interaction, evaluation, engineering, operations and challenge around one queue and shared evidence. Siloed specialists previously handed work across slow boundaries. Proof and sourcing: supervised production tests coordination and escalation; build the team routine while buying a replaceable model service. Capacity and worker evidence: handoff delay, role clarity and learning are tracked. Half-life and owner: workflow or authority change triggers recomposition; the operations owner invests in shared work design, not fewer roles.
8. Individual assistance leaves coordination untouched
Cell: prepare documents that reach timely approval. Layers and graph: employees draft faster, but meeting, review and approval nodes retain old queues, duplicate evidence and ambiguous rights. Human-AI practice improved while team and institutional capability did not. Proof and sourcing: individual use is sustained, yet end-to-end outcome evidence remains flat. Build workflow and decision-right capability rather than another assistant course. Capacity and worker evidence: meeting load, waiting time and rework are measured. Half-life and owner: approval-policy change triggers redesign; the process owner funds shared evidence and fewer handoffs.
9. Buy the commodity model, retain the learning advantage
Cell: classify routine service inputs for a governed workflow. Layers and graph: model hosting is commodity, while workflow design, evaluation, data stewardship, exception policy and benefit capture contain enterprise context. Proof and sourcing: buy the model service with portability and evidence terms; build the retained cells in-house. Capacity and worker evidence: exception reviewers, data stewards and supplier oversight are sized. Half-life and owner: model or price change triggers comparison, not wholesale capability loss; the capability owner invests in evaluation assets and a substitute path. External capability is deliberate here.
10. Enthusiasm cannot create control capacity
Cell: scale an AI-assisted high-consequence review across business units. Layers and graph: users are literate and the tool performs well, but human reviewers, risk challenge and incident response are shared bottlenecks. Proof and sourcing: protected and limited production evidence passes; portfolio-scale proof fails the capacity test. Capacity and worker evidence: queues, concentration, fatigue and escalation delay exceed the authorised envelope. Half-life and owner: volume is itself a trigger. The executive owner stops expansion and funds reviewer, incident and succession capability before another gate. High enthusiasm is not scale evidence.
Augmented Workforce Capability StudioResolve the cell into one of five workforce actions
The Augmented Workforce Capability Studio is this article's synthesis. It returns exactly five decisions: build foundations across a role family; deepen scarce capability in a specialist guild; redesign the team around complementary work; source temporary capability with transfer; or stop scaling until the workforce control gap closes. It does not create a universal readiness score.
Strong rival and failure boundaryKeep the model small enough to change work
The strongest rival is right that capability models often become large taxonomies that are costly to maintain and disconnected from jobs, hiring and daily work. They can overstate certainty about future skills, encourage credential inflation and treat workers as interchangeable inventory. Formal workforce processes move slowly while tools, tasks and supplier boundaries change quickly.
The response is not a more detailed inventory. Start with a small number of consequential workflows. Define a capability only when it changes team composition, learning, sourcing, authority, capacity or investment. Link every item to a cell, evidence level and owner. Co-design with workers and managers. Preserve uncertain future capability as an option until a real dependency appears. Retire entries that no longer alter a decision.
Reject the easy substitutes. Assistant access is not augmentation. Literacy is not the complete model. A generic future-skills list without a work event does not guide staffing. Self-ratings and attendance do not prove performance. Technical talent cannot replace domain craft, and domain expertise cannot supply engineering or independent assurance. Automating production while ignoring review, escalation and coordination merely moves the load.
Source choices require symmetry. Do not build every commodity service internally, and do not outsource the proprietary context or feedback loop that creates learning advantage. Temporary experts need transfer conditions. Partners need learning rights. Workers need voice, workload protection and accessible practice. Headcount reduction is not the capability objective; a sustained outcome with lawful authority, sufficient capacity and recoverable work is.
Evidence also has boundaries. UK, ILO and OECD materials provide workforce, occupational and organisational context. NIST and EU sources support distributed responsibility and role-sensitive literacy.1012 Field studies show heterogeneous individual and team effects under particular tasks and settings, not universal productivity constants. The named cells, graph, portfolio, tests, calendar and studio remain this paper's synthesis.
The final question is whether the organisation can keep producing the outcome when the model changes, the expert leaves, demand rises or the exception moves outside the tool's capability. If the team can observe, adapt, transfer knowledge, protect workers and affected people, and preserve accountable capacity, augmentation has become an institutional capability rather than a temporary software effect.
Evidence spineSources checked 3 September 2026
- UK Government, AI Skills Tools Package.
- UK Government, What Works for AI Upskilling in the UK: Employer Guide.
- UK Government, Supporting Case Studies: What Works for AI Upskilling in the UK.
- UK Government, Government Digital and Data Profession Capability Framework.
- International Labour Organization, Generative AI and Jobs: A 2025 Update.
- International Labour Organization, Artificial Intelligence Adoption and Its Impact on Jobs.
- OECD, AI and Skills.
- OECD, Artificial Intelligence and the Changing Demand for Skills in the Labour Market.
- OECD, The Adoption of Artificial Intelligence in Firms.
- NIST, Artificial Intelligence Risk Management Framework.
- NIST, AI RMF Playbook.
- European Commission AI Office, AI Literacy: Questions and Answers.
- Brynjolfsson, Li and Raymond, Generative AI at Work.
- Dell'Acqua et al., The Cybernetic Teammate.
- Dillon et al., Shifting Work Patterns with Generative AI.
- Cruces et al., Does Generative AI Narrow Education-Based Productivity Gaps?.
- Dell'Acqua et al., Navigating the Jagged Technological Frontier.
- Edmondson, Psychological Safety and Learning Behavior in Work Teams.
About this paperCapability lives where work, capacity and learning meet
The Six-Layer Augmented Workforce Capability Model, Capability Cell, Workforce Capability Graph, Capability Sourcing Portfolio, Team Composition Test, Capability Half-Life Calendar and Augmented Workforce Capability Studio are this article's synthesis. They turn workforce planning into a portfolio of owned, evidenced and refreshable operating cells.