Automation and integration

PUBLISHED

Connecting AI systems to others, and running them unattended.

Statements

—

Version

v1.29

Last reviewed

30.08.2026

In scope

Programmatic access and its differences from conversational use

Tool and function calling, and the design of what a system may invoke

Construction of workflows and pipelines incorporating AI steps

Triggers, scheduling and event-driven execution

Placement of human review, and the criteria determining where it is required

Error handling, fallback and graceful degradation when an AI step fails

Idempotency, retry behaviour and the consequences of non-determinism under repetition

Monitoring a running automation

Cost and rate control at volume

Out of scope

Excluded	Where it lives
Software engineering in general	`X-SWE`
Agentic autonomy, planning and multi-step tool use	`X-AGT`
Infrastructure and deployment	`Out of framework`

Boundary notes

Where this domain abuts another, and how the line is drawn

Against general software integration

The domain most at risk of scope creep. The test: if the statement would be true of any software integration with no AI involved, it does not belong here.

Against X-AGT

D4 covers automation that executes a defined workflow. X-AGT covers systems that determine their own steps. The line is who decides the sequence.

Against D7 Ethics and Governance

Human-in-the-loop placement is D4. Whether human review is required as an obligation is D7.

Statements by level

Every identifier is a permanent address. Indicators are normative; they state what would be observed in a person who meets the statement.

L1 Aware

4 statements

D4.L1.01Knowledge

Describe what changes when an AI step runs without a person present.

Indicators

States that no one is available to catch an error at the moment it occurs

Identifies the volume effect: one fault repeats across every run

Recognises that output becomes input to something else

D4.L1.02Judgement

Identify the steps in a workflow they perform that could run unattended.

Indicators

Distinguishes a repetitive step from a step requiring judgement

States what would happen if the step produced a wrong result silently

Refers the decision rather than automating it alone

D4.L1.03Knowledge

State why an automated step needs a record of what it did.

Indicators

States that an unlogged run cannot be investigated

Identifies what must be recorded to reconstruct a run

Recognises that a summary is not a record

D4.L1.04Knowledge

Recognise that a system connected to other systems can act, not only answer.

Indicators

Distinguishes reading from writing, sending or paying

Identifies which of their tools can take action

Reports a capability they did not expect a tool to have

L2 Applied

6 statements

D4.L2.01Practice

Operate an automated AI workflow within its stated limits.

Indicators

Runs it on the inputs it was designed for

Stops it rather than working around an unexpected result

Reports drift from expected behaviour

D4.L2.02Skill

Configure the handover between an AI step and the step that follows it.

Indicators

States the format and constraints the next step requires

Validates the output before passing it on

Provides a defined behaviour for an invalid result

D4.L2.03Skill

Record what an automated run did, on what input, with what result.

Indicators

Records input, output, version and timestamp for each run

Retains records long enough to investigate a late-discovered fault

Makes the record accessible to whoever must investigate

D4.L2.04Skill

Identify and report a failure in an unattended step.

Indicators

Distinguishes a failed run from a run that produced a wrong result

Establishes how many runs were affected

Reports before the next scheduled run rather than after

D4.L2.05Practice

Apply the organisation’s rules on which actions an automated step may take.

Indicators

Establishes which actions require human authorisation

Configures the step to hold rather than proceed where authorisation is required

Refers cases the rules do not cover

D4.L2.06Skill

Test an automated workflow before it runs on real work.

Indicators

Tests with inputs that should fail as well as inputs that should pass

Verifies the failure path, not only the success path

Records the test result before release

L3 Proficient

8 statements

D4.L3.01Skill

Design an automated AI workflow, stating where human judgement is required.

Indicators

Places the human step where the consequence of error is highest

Justifies each point at which no human is present

States what the human step is expected to catch

D4.L3.02Skill

Design error handling for an unattended AI step.

Indicators

Distinguishes retry, hold and abort, and states when each applies

Prevents a fault from propagating to downstream systems

Ensures a held run is visible to someone who can act

D4.L3.03Skill

Design the observability of an automated workflow.

Indicators

Selects signals that reveal wrong results, not only failed runs

Sets alert thresholds that produce action rather than noise

Provides for the fault nobody anticipated

D4.L3.04Judgement

Assess the blast radius of an automated step and constrain it.

Indicators

States what the step can reach and change

Limits scope, rate and reversibility in proportion to consequence

Records the assessment for review

D4.L3.05Skill

Design an AI system that acts across multiple steps, with authorisation boundaries.

Indicators

States which actions the system may take without confirmation

Prevents an instruction reaching an action it was not authorised for

Provides for stopping the system mid-sequence

D4.L3.06Skill

Diagnose a fault in a multi-step automated workflow.

Indicators

Reconstructs the run from records rather than by re-running

Identifies the step at which the result became wrong

Distinguishes a fault in the step from a fault in the handover

D4.L3.07Judgement

Evaluate whether automation is warranted for a given process.

Indicators

States the cost of the failure mode alongside the saving

Identifies where automation would obscure a signal a person currently notices

Recommends against automation where that is the answer

D4.L3.08Skill

Design the rollback of an automated workflow and its effects.

Indicators

Identifies effects outside the workflow that must also be reversed

States what cannot be reversed and what follows from that

Tests the rollback rather than documenting it

L4 Advanced

4 statements

D4.L4.01Practice

Establish the conditions under which AI systems may act unattended in the organisation.

Indicators

States the classes of action permitted without human authorisation

Requires evidence proportionate to consequence before autonomy is granted

Provides a route to withdraw autonomy already granted

D4.L4.02Practice

Define accountability for the output of automated AI workflows.

Indicators

Assigns a named accountable owner to each automated workflow

States that accountability is not discharged by the automation

Provides for the owner changing role

D4.L4.03Practice

Establish operational standards for monitoring, alerting and incident response in AI operations.

Indicators

Sets minimum monitoring for each class of workflow

Defines response times and who is accountable for them

Verifies that the standard is met rather than assumed

D4.L4.04Judgement

Govern the accumulation of automation, including workflows nobody owns.

Indicators

Maintains a record of what runs unattended and what it can reach

Identifies and retires unowned or undocumented automation

Prevents accumulation rather than periodically clearing it

Editorial notes

Known gaps, open questions and contested points

Published because the record is more useful than the appearance of completeness.

Needs a statement at L3 on deciding not to automate. Currently absent from every comparable framework.

All eight

Domains index

Filtered view

Statements in D4

Proficiency

Levels

This page displays version 1.29 · last reviewed 30.08.2026