Your workforce is not one population. It is two, and they have to be developed together.
When an enterprise puts agents into the organisation, the job of learning changes shape. The people who used to do the work now supervise it, and the agents that do the work were never hired, never inducted and never trained. This page sets out how we develop both populations against one capability model — including what a digital employee actually needs before it is allowed near a customer, and what it means to redesign a job around a machine.
One model, two carriers
The same capability dimensions, different instruments for humans and agents
Job redesign first
Redesign the job, then write the curriculum — never the other way round
Shadow before autonomy
Agents earn scope by working beside a human and being measured
Evidence in one record
A person's record and an agent's file live in the same system
Training a human for the old job and deploying an agent for the new one is how organisations get slower.
Most agent programmes do one of two things. They either train people in a new tool and leave the process unchanged, or they automate the process and leave the people where they were. Both produce an organisation that has adopted a technology without changing how work actually moves.
We start from the job, not the tool. For each role in scope, we write down what the work actually consists of, split it into the parts a machine can carry and the parts that require accountable human judgement, and then design two curricula: one for the person who supervises, one for the agent that executes. The result is a role that is genuinely different, and a learning system that can prove both halves of it are competent.
- Capability dimensions are shared — a "claims assessment" dimension applies to the handler and to the assessment agent, with different evidence instruments for each
- Decisions are placed, not removed — every step keeps a named human owner, even if the agent drafts, recommends or pre-fills
- Judgement is taught as judgement — scenario-based, with escalation criteria the trainee must be able to state out loud
- Confidence is calibrated — the human learns to read an agent's uncertainty correctly, which is the single most transferable skill in the new operating model
What changes for the human role
Execution becomes a smaller share of the job. Judgement, escalation and communication become a larger share. Most role descriptions have not caught up with that.
Manual execution
Data entry, drafting, formatting, first-pass lookups, routine replies.
Judgement and intervention
Deciding when the agent is wrong, handling the exception, owning the customer outcome.
Design of the work itself
Where the agent should stop, what it should escalate, and how its scope should grow.
A role redesigned without a curriculum produces anxious staff and an over-supervised agent. Both are avoidable with a plan.
Digital workforce in numbers
Onboarding, a performance record, a retirement plan
The curriculum a digital employee receives before it acts
Not uptime — accuracy, escalation, override and cost
A person, for every agent in production
Digital employees need three things software never needed
This is the gap that organisations discover during their first incident: they deployed something with authority and never onboarded it. The three requirements are not difficult; they are simply not in anybody's default workflow.
- Installed, configured, switched on
- Monitored for uptime and error rate
- Version number as its identity
- No training record
- Retired when it becomes inconvenient
- Nobody can say what it decided last week
- Given a role, a job description and explicit non-goals
- Completed a curriculum before touching production work
- Supervised by a named person, at a stated frequency
- Performance measured against an evaluation threshold
- Cost per task tracked, and its authority tier reviewed
- Retired deliberately, with its decisions reviewed and its knowledge preserved
Where the work goes, step by step.
The table below is the artefact we produce first in every engagement, before a single course is commissioned. It is deliberately concrete: a role, a task, a split, a decision right, an evidence requirement.
| Task in the redesigned role | Carried by | Human retains | Evidence required |
|---|---|---|---|
| Intake and case classification | Agent | Override on ambiguous intent; taxonomy changes | Classification accuracy against a human-labelled evaluation set |
| Policy retrieval and eligibility check | Agent | Authority to interpret policy gaps | Citation correctness on every retrieved rule |
| Drafting the customer response | Agent | Voice, tone and any commercial commitment | Review-and-edit rate; unexplained-edit rate |
| Exception and edge-case handling | Human | The whole case | Simulation score and observed practice on exception scenarios |
| Prioritisation under SLA pressure | Agent proposes, human disposes | Final prioritisation and the customer relationship | Escalation rate; SLA breach rate |
| Quality assurance of agent output | Human | Standard-setting and case acceptance | Inter-reviewer agreement; calibration session records |
| Process improvement and scope expansion | Joint | Decision to widen the agent's remit | Named owner sign-off plus regression test results |
Note the shape of the table. The human column is not residual work — it is the work that carries accountability. When a customer's case fails, someone has to be able to explain the decision. The agent can hold the effort; it cannot hold the responsibility, and our curriculum is built around that line.
The six-module onboarding curriculum every digital employee receives
Delivered through the LMS, recorded like any other onboarding, and versioned with the agent. Module 6 is for the supervisor, not the agent — and is the one most often skipped.
What the agent is for, and — more importantly — what it must never do. Non-goals are the module that prevents most incidents.
Which decisions it may make at which authority tier, and the exact conditions under which it must stop and ask.
What it may read, what it may not, and how to recognise when a source is stale or ungoverned.
What it may change, what is irreversible, and what a human must see before it happens.
What it does when it cannot complete, and to whom it hands the work. Practised, not just described.
For the named human: how to review its output, what to override, and how to tell a real problem from noise.
Six artefacts an agent needs before it touches live work.
This is the checklist we use for every digital employee we commission, whether it runs in our own delivery or in a client's environment. If any of the six is missing, the agent is not onboarded — it is deployed.
Knowledge set
Versioned, curated and bounded. It states what the agent is entitled to retrieve, from which sources, and — just as importantly — what it is not allowed to know. Unbounded retrieval is a governance failure waiting for a customer.
Skill set
An enumerated list of capabilities, each with a testable definition of done. "Can handle a claims query" is not a skill; "can classify a query into eleven policy categories with stated confidence and escalate the two ambiguous ones" is.
Evaluation set
Real cases from your own history, including the ones that historically went wrong. Held as a regression suite: every change to knowledge, tools or prompts is re-scored against it before release.
Permission scope
Least privilege by default, granted per tool and per data class, with a documented approval path for expansion and a single-action revocation. Scope grows on evidence of performance, never on optimism.
Compliance evidence
Which framework the agent was tested against, which controls it satisfies, who approved it, and what changed in each version. Presented to an auditor in the same shape as a human operator's training file.
Performance file
A longitudinal record — task success, escalation rate, cost per completed task, and the ninetieth-day check after go-live. Comparable across versions, so you can see whether a release actually improved the work.
The digital workforce business lines
A new business, and one most organisations have not yet realised they need. Each line has its own evidence model rather than a completion rate.
| Service line | What it is | How it is evidenced |
|---|---|---|
| Agent onboarding as a service | Design and deliver the six-module curriculum for a specific agent estate | Curriculum records in the LMS, per agent, versioned |
| Supervisor enablement | Training for the named humans who review and override agent work | Assessment record per supervisor, with a validity period |
| Capability managed outsourcing | Running the supervision, evaluation and evidence operation on your behalf | Continuous operational record with a defined handover point |
| Agent estate governance | The ownership model, authority tier reviews and retirement discipline | Portfolio record: owner, tier, threshold, cost, review date |
| Agent reskilling programmes | Redeploying people from execution to supervision, authorship and ownership | Role transition records in the LMS, tied to the new role's authority |
A five-stage path, from baseline to certified competence.
The same five stages apply to a person and to an agent, with different instruments at each stage. What differs is the evidence required to pass, not the architecture of the journey.
Baseline
Before anything is taught, we measure. For a person, a practical assessment against the real task. For an agent, a first run of the evaluation suite to establish the starting point. Without this, every later number is uninterpretable — you cannot tell improvement from the effect of a good month.
Micro-learning on the specifics
Short modules built around your process, not generic practice. Deliberately small, because the material that matters is the delta between how your organisation works and how the general case works. Deployed in the flow of work, not scheduled in a room nobody attends.
Scenario simulation
MetaSim business and sales simulation puts both populations through the same decision points: what did you do, what did it cost, what would you do differently. For agents, the simulation doubles as a pre-production rehearsal, run in a sandbox with no customer impact.
Shadow mode
The agent produces work alongside a competent human, and neither output reaches the customer. The disagreement between them is the raw material of the coaching: it is where you learn which cases the agent handles poorly and what coaching the human needs. The error rate during this period is the single most useful number in the programme.
Certification and in-the-work coaching
Both parties are certified against the same standard, then supported in live work by a coach for a defined period. For agents this includes watching for drift — an agent that performs well in month one and degrades by month three has not been certified, it has been launched.
Four things to measure, and what the worrying number looks like
Each measure has a failure signature that looks like success. Knowing which signature to watch for is what separates oversight from monitoring.
| Measure | What it tells you | The number that should worry you |
|---|---|---|
| Accuracy against threshold | Whether the agent is doing the task to the standard you set | A threshold that was lowered to make the number look better |
| Escalation rate | Whether it knows when it should stop | Falling towards zero — that usually means it stopped escalating rather than that it got better |
| Override rate by approver | Whether the humans supervising it are reviewing properly | A rate near zero across every approver, which means rubber-stamping |
| Cost per task | Whether the economics still hold at this volume | Rising month over month with no change in traffic mix |
| Failure clusters | Whether it is actually improving | Clusters moving rather than reducing — you are fixing symptoms |
Three roles that did not exist five years ago, and now need a job description and a development plan.
Introducing agents creates positions. Most organisations discover this after the first failed deployment, when it turns out nobody owns the agent, nobody can maintain its knowledge set, and nobody is responsible for its performance. We build these roles into the capability plan from day one.
Agent Owner
A business leader accountable for an agent's remit, scope and outcomes. Not an engineer — a person who owns the business result and defends the line between what the agent does and what stays human.
Agent Engineer
Builds and maintains the agent: knowledge sets, tools, evaluation suites and regression runs. A hybrid role that needs both technical depth and enough process knowledge to write a meaningful evaluation case.
AI Steward
Sits in the business and keeps the human side honest: are we still reviewing agent output, or rubber-stamping it? Runs the calibration sessions and watches for skill decay in the human population.
Each role needs a capability profile, a development path and a definition of what competent looks like. Without that, these roles are filled by whoever is closest to the project, and the organisation inherits a new layer of unmanaged risk.
Workforce planning, connected
The capability data this generates should not sit in a learning platform. It is the raw material of workforce planning, and it belongs in the same system as the rest of the talent record.
- Capability inventory — which roles hold which capabilities, at what depth, evidenced how
- Gap analysis — where the strategy depends on a capability nobody in the organisation has
- Sequencing — which gaps block the roadmap and which can be developed in parallel
- Agent substitution view — which portions of each role are being carried by machine, and what remains
- Decay monitoring — capabilities that were certified two years ago and have not been re-evidenced
The talent record and the learning record should be the same record. Most enterprises maintain both, and they disagree with each other by the time anyone checks.
Retiring an agent is a process, not a switch
Every organisation has a plan for deploying agents and none for retiring them. The retirement is where accumulated decisions get reviewed — which is sometimes uncomfortable and always necessary.
Identify
The agent is superseded, its workflow has changed, or its cost no longer justifies it.
Review its decisions
Everything it decided during its life is sampled and assessed. Some of it may be wrong, and that has consequences.
Preserve its knowledge
The skills, judgments and edge cases it accumulated are captured as assets for the next agent — or archived deliberately.
Redirect the humans
The people who relied on it get told what happens to the work, and retrained if it changes.
Revoke and archive
Access removed, records archived with the retention period applied, and the retirement recorded as a governance event.
Five measures that survive a sceptical question.
Each of these has a defined computation, a defined population and a defined evidence source. None of them can be improved by encouraging people to click more buttons, which is the point.
| Measure | How it is computed | Reported for |
|---|---|---|
| Capability baseline | Assessment result at entry, against the role's capability profile | Both populations |
| Time to proficiency | Calendar days from first exposure to passing the role certification | Both populations |
| Shadow-period error rate | Share of shadowed tasks where agent and human outcome diverged materially | Digital employees |
| Ninety-day post-launch performance | Task success and escalation rate measured 90 days after go-live, against the launch baseline | Digital employees |
| Human review quality | Share of agent outputs edited without a stated reason; inter-reviewer agreement in calibration | Employees |
| Escalation quality | Of cases the agent escalated, the proportion that genuinely required human judgement | Both populations |
Why learning is the foundation of an agentic business, not a line item
This is the single idea that separates our enterprise learning business from a training vendor. An agentic business needs three populations taught, and only one of them currently is.
- Learning is for employees
- Courses are assigned by an L&D system
- Completion is the metric
- The LMS is a compliance archive
- External partners are somebody else's problem
- Software is deployed, not onboarded
- Learning covers every capability carrier — employees, partners, and digital employees
- Capability is assigned by role and authority, not by department
- Operating evidence is the metric, not completion
- The LMS is a governed semantic store the agents reason over
- Partner enablement sits inside the same system, with its own records
- Digital employees are onboarded with a curriculum and given a performance record
One platform holds both files, because your evidence should not depend on who did the work.
UpLearn LMS, Kora and the evaluation tooling write into the same capability record. When a customer asks who is qualified to do a task, the answer covers the human who did it last quarter and the agent that did it this week — with the same structure, the same audit trail and the same retrieval.
Same record shape
A capability identifier, a competency level, an issue date, an expiry date and an evidence pointer — identical whether the holder is a person or an agent version.
One audit path
Regulator, insurer or internal audit asks a question about a service: which entities performed this work, under what competence, and what evidence supports it. One query, not two systems and a reconciliation.
One renewal cycle
Certification, attestation and re-evaluation all expire on the same schedule, and the escalation path when they lapse is identical — so nothing quietly falls out of currency on either side.
We do not run AI certification crash courses.
A certificate that takes a day to obtain and expires in a year is a marketing artefact, not a capability. We have been asked to produce exactly this and we have declined: it is fast to build, cheap to sell, and it guarantees that the number on the dashboard means nothing.
We also do not train people to operate a specific vendor's interface and call that transformation, and we do not certify a population we have not assessed. If the underlying business process is not ready for agents to work in, the honest answer is a different conversation, and we will say so rather than sell a curriculum around it.
What we will do is build the assessment, the curriculum and the evidence for the specific capability your organisation depends on — and keep the record for as long as the risk lasts.
Competence in supervising agent work is assessed against real cases. A short course and a badge is a different product, sold by other people, more cheaply.
Where you use a specific tool changes annually. What changes slowly is knowing how to brief, judge and correct the work — and that is what we teach.
If nobody has established what the population can already do, a pass rate measures the course, not the capability. We assess first, and we report both numbers.
Shadow mode is how the error rate becomes a number instead of an anecdote. Skipping it to save a week is the most expensive week you can save.
What the digital workforce business is not
We do not sell digital labour by the hour, and we will not build an outsourcing business whose economics depend on you not understanding what the agents are doing. The supervision is yours by design, ours by contract, or somebody's — and the most dangerous arrangement is the one where nobody can say which.
What this page answers
Every topic below is covered in the sections above, and each links to the page that goes deeper on it.
Choose one role and we will redraw it around an agent.
Give us a role where you already suspect work could be reassigned. In three weeks we will produce the task-by-task human/agent split, the curriculum for both sides, and the six-artefact onboarding checklist for the agent — before anyone commits to a platform.