AI workflow-mapping methodology
How we map where
AI fits in the work.
Deployed Intelligence is a task-based method for describing organisational work, assessing what that work demands, and applying a specified technology and deployment policy to identify appropriate forms of human and AI participation.
AI workflow mapping means asking which specific steps of work a specified system could carry, which require human review, which should remain human-led, and which lie outside that system’s scope—and answering those questions through the same traceable process across the organisation.
On this page
The conceptual structure, the analytical pipeline, the ten work dimensions, the deployment-profile layer and the current evidence status.
§ 01 · Core principle
Measure the work,
not “automatability.”
Deployed Intelligence does not begin by assigning an automation score. It begins by describing what the work demands of whoever or whatever performs it.
A direct automation score combines too many assumptions into one number: the capability of today’s models, the organisation’s risk posture, the quality of its data, the availability of system integration and the nature of the work itself.
Instead, the methodology creates a relatively technology-neutral description of the work. It asks about properties such as the consequence of error, the intensity of value judgment, the stability of the procedure, the contextual information required, and the importance of human relationships.
A separate deployment profile then interprets those properties for a specified technology and policy. If the technology changes, the same work map can be evaluated again under a new profile—provided the underlying work has not materially changed.
§ 02 · Unit of analysis
Jobs contain tasks.
Tasks contain steps.
The step is the primary analytical unit because different parts of the same task may require very different working arrangements.
Job-level analysis is often too coarse. A resource-consents planner may retrieve plans, compare a proposal with rules, tabulate measured effects, weigh competing considerations and defend a professional recommendation—all within one task.
Some of those steps may be suitable for a system to carry subject to review. Others may be best performed by a person with AI support. A small number may require human decision authority because the legitimacy of the judgment is part of the work.
Decomposing work also creates possible handover points. A useful step is described so that it has a recognisable beginning and end, performs one coherent piece of cognitive or procedural work, and can pass its output to the next part of the workflow.
§ 03 · Work dimensions
Ten dimensions describe
what the work demands.
Every step is assessed against the same versioned metric registry. Each score is accompanied by a written rationale.
The dimensions are designed to remain distinct. High consequence does not automatically mean high value judgment. A rare area of expertise does not automatically imply complex reasoning. A stable process may still contain tacit judgment.
Consequence Magnitude
The severity of harm if the step is performed poorly, incorrectly or not at all.
Contextual Awareness Requirement
How much of the current surrounding situation must be perceived and held in mind, including how quickly it changes.
Human Relationship Sensitivity
The interpersonal and emotional skill required to read, influence, reassure or respond appropriately to people.
Procedural Stability
How consistently the method remains the same across different instances of the step.
Persistent Context Dependency
The depth and time horizon of historical context that must be carried forward for continuity.
Creative Generation Requirement
The degree of genuine novelty the performer must produce rather than selecting, configuring or reproducing existing material.
Explicitness of Reasoning
How fully the reasoning can be articulated, specified and transferred to another person or system.
Cognitive Complexity
The intrinsic structural complexity of the reasoning required, independent of who performs it.
Value Judgment Intensity
How much the outcome depends on values, priorities, principles or norms rather than facts or unambiguous rules.
Expertise Scarcity
How rare the specialised knowledge required by the step is across the working population.
Metrics are not the only inputs
The methodology also records direct step attributes, including physical demand, interaction mode, exposure and reversibility. These attributes can perform different analytical work from the metric scores—for example, determining whether a chat-based system is outside the scope of a physical step.
Not every metric must influence every technology profile. Expertise Scarcity, for example, is currently reported as a prioritisation dimension but is deliberately excluded from the default chatbot verdict logic because its relationship with AI suitability remains contested.
§ 04 · Analytical pipeline
A provisional map,
then evidence and review.
Language models assist with several stages, but their outputs are treated as structured analytical claims—not as self-validating descriptions of reality.
Define the role and reference context
Collect role titles, organisational context, job descriptions and practitioner input. Record the richness and limits of the source material.
Generate tasks and decompose them into steps
Produce a provisional role map and describe the ordered or conditional work through which each task is completed.
Check the map with people who know the work
Test whether tasks belong, what is missing, where wording is wrong, and whether steps should be reordered, merged, split or treated as conditional.
Characterise, attribute and score each step
Apply the versioned metric registry and record direct attributes. Preserve every score rationale and the configuration that produced it.
Apply technology capability and deployment policy
Evaluate each step under a named deployment profile and retain the full trace of gates, capability checks and policy rules.
Aggregate as distributions, not one score
Show the mix of working arrangements across tasks, roles and the organisation without pretending they collapse into one universal automation percentage.
§ 05 · Deployment profiles
Capability and policy
remain separate.
A deployment profile is not a generic description of “AI.” It is a versioned interpretation of a specified technology, deployment setup and organisational policy.
Is this technology the right instrument?
Direct attributes determine whether the profile applies at all. A chat-based system, for example, is outside the scope of constitutively physical work.
Can the specified system contribute to the core step?
Metric bounds represent current claims about the technology’s capability. A breach may limit the system to assistance or make the core work human-required.
How much human control should remain?
Named rules consider consequence, value judgment, relationship demands and other governance factors. The most restrictive applicable result is retained.
The map can change for two different reasons
A more capable or better-integrated technology may expand the capability envelope. Separately, an organisation may choose a more cautious or permissive deployment policy. Those changes should not be confused: risk appetite does not make the technology more capable, and technical capability does not settle what the organisation should permit.
§ 06 · Outputs
Five outcomes,
never one automation score.
Each step receives a working arrangement under the selected profile. Tasks and roles are then shown as distributions across those arrangements.
AI carries
The system performs the primary work. People govern and monitor the workflow rather than reviewing each instance.
AI carries, human approves
The system performs the primary work; a person checks the result before it has an effect.
Human leads, AI assists
A person performs the core step while AI drafts, retrieves, analyses or checks within it.
Human required
A person must perform or own the core decision because judgment, relationship, presence or accountability is essential.
Beyond this profile
The technology is not the right instrument for the step. This is a scope result rather than a rung on the autonomy ladder.
“Beyond this profile” is kept separate from “Human required.” A physical step may be outside the scope of a chat-based system, but that says nothing about whether another technology—such as robotics—could eventually participate.
The distributions are not staffing forecasts. They do not account automatically for task frequency, implementation effort, system performance, demand growth, process redesign or how saved time would be used.
§ 07 · Evidence status
What has been tested.
What remains open.
The methodology has a substantial internal evidence base, but it is not presented as a completed academic validation exercise.
Internal experiments are used to select and pin production configurations, test score stability, examine prompt and model effects, preserve important construct boundaries and identify where profile thresholds materially change the output.
Engineering and analytical controls
- Versioned metric and profile registries
- Full run and verdict provenance
- Deterministic profile-layer logic
- Model and scoring-strategy experiments
- Repeat-stability and grain-sensitivity tests
- Internal tests of key metric boundaries
Practitioner and external evidence
- Task-list recognition and missing-work review
- Practitioner assessment of decompositions
- Human interpretation of selected metric boundaries
- Capability versus human-control judgments
- Rationale and score-concordance testing
Policy and profile assumptions
- Required review at specified consequence levels
- Human authority for value-defining decisions
- Current capability-envelope thresholds
- Risk-appetite settings
- Acceptable forms of organisational autonomy
Real-world deployment outcomes
- Actual system accuracy in a client workflow
- Implementation cost and integration difficulty
- Employee or customer acceptance
- Compliance with sector-specific obligations
- Staffing or productivity outcomes
§ 08 · Limits and use
A decision map,
not a deployment decision.
The output identifies where to investigate and what questions to ask. It does not replace technical evaluation, assurance, consultation or accountable organisational decision-making.
It is not a headcount model
A step-level disposition does not translate directly into jobs or full-time equivalents. Tasks occur at different frequencies, workflows change, demand may grow, implementation takes effort and people often redirect saved time into other work.
It is not proof that a system performs adequately
A capability profile is an evidence-informed claim about a technology class. A real deployment still requires tests using the actual model, tools, data, permissions, users and quality standards.
It is not a substitute for practitioners
Role descriptions and model-generated task maps cannot fully capture local variation, tacit knowledge, exceptional cases or how work has changed over time. Practitioner review is a substantive part of the process.
It is not a universal policy answer
Organisations may reasonably adopt different positions on review, accountability and acceptable risk. Those positions should be stated and compared rather than hidden inside a single score.
It is not limited to local government
The methodology is designed to apply wherever work can be represented meaningfully as tasks and steps. The public council example demonstrates the pipeline; it does not establish universal validity across every profession or operating environment.
§ 09 · Common questions
The short
answers.
Is this an automation-potential score?
No. Deployed Intelligence first assesses properties of the work itself, including consequence, value judgment, relationship sensitivity, context and procedural stability. A separate technology and policy profile then interprets those properties for a specified system.
Can a language model reliably assess work?
A language model can apply a fixed rubric consistently enough to support structured analysis, but its outputs are not treated as ground truth. The methodology records provenance, tests stability and construct boundaries, flags uncertainty, and is adding practitioner validation and empirical technology evaluation.
Does the methodology predict whether jobs will be replaced?
No. It analyses tasks and steps rather than treating a job as one automatable unit. It maps possible working arrangements and does not directly predict staffing levels, implementation feasibility or employment outcomes.
Does the methodology work outside local government?
It is designed to be domain-general wherever work can be represented as tasks and steps. The public council case study is a composite demonstration rather than proof of universal validity, so each new domain still requires suitable context and practitioner review.
What changes when AI technology improves?
The underlying work map and work-characteristic scores can be retained while the steps are re-evaluated under a new versioned deployment profile, provided the work itself has not materially changed.