SWANK AI Guidance Note 16
AI Triage · Operational Governance · Human Review

Core Standard

PRIORITISE ATTENTION, NOT TRUTH

AI triage may help determine what should be reviewed first.

It should not silently transform prioritisation into factual or professional judgment.

Purpose

Artificial intelligence can help organisations process large volumes of:

correspondence;
complaints;
applications;
records;
service requests;
incident reports;
administrative submissions;
and other operational information.

One common use is triage.

A triage system may help determine:

what should be reviewed first;
what should be routed elsewhere;
what may require escalation;
what information appears incomplete;
or what may be treated as lower priority.

This can improve efficiency.

It can also create hidden operational consequences if prioritisation criteria are unclear or if triage becomes indistinguishable from substantive judgment.

The core distinction is:

Triage determines attention.

It should not automatically determine:

truth, credibility, entitlement or final seriousness.

Triage Is Not Determination

A triage system may appropriately assist with functions such as:

identifying urgent language;
detecting missing information;
recognising deadlines;
routing correspondence;
grouping similar requests;
recognising possible safeguarding or operational concerns for human review;
identifying incomplete applications;
or flagging material requiring additional attention.

Those functions can support administrative efficiency.

They are different from asking the system to decide:

whether a complaint is valid;
whether a person is credible;
whether an allegation is true;
whether someone is unreasonable;
whether intervention is required;
or whether a matter is ultimately serious.

Triage should help determine:

what receives attention

rather than automatically determine:

what is true.

Priority Is an Operational Classification

A priority label is not necessarily a factual conclusion.

For example:

High priority

may mean:

Review this quickly.

It does not necessarily mean:

The underlying concern has been established.

Likewise:

Low priority

may mean:

No immediate urgency was detected.

It should not automatically mean:

The matter is unimportant or without merit.

Organisations should preserve this distinction in both policy and practice.

Priority Criteria Should Be Visible

Organisations should understand which variables influence prioritisation.

Relevant questions include:

What makes one case rank above another?
Are particular words disproportionately influential?
Does communication length affect the ranking?
Does previous contact history affect priority?
Are certain communication styles treated differently?
Does the presence of emotional language increase escalation?
Can a human change the assigned priority?
Is prioritisation periodically reviewed?
Does later information alter the ranking?

Opaque prioritisation can create operational inequality even where the underlying AI model appears neutral.

Language Is Not the Same as Urgency

AI systems may rely heavily on language patterns.

That creates a risk that the system confuses:

how something is expressed

with

how urgent it actually is.

A highly emotional message may trigger strong escalation.

A calm message may be treated as routine.

Neither conclusion is necessarily correct.

Urgency should be assessed in relation to the substance and context of the information.

High Urgency and False Positives

Systems designed to detect risk may intentionally favour sensitivity.

That may be appropriate where missing a significant concern could create serious consequences.

But greater sensitivity can also increase false positives.

A system may flag:

emotionally intense language;
certain keywords;
unusual phrasing;
repeated communication;
or indirect references

as indicators of high urgency.

That may justify review.

It does not necessarily establish risk.

Organisations should therefore distinguish clearly between:

FLAG FOR HUMAN ATTENTION

and

ESTABLISHED RISK

Those are different evidential states.

False Positives Can Have Consequences

Over-escalation may result in:

unnecessary intervention;
inappropriate prioritisation;
reputational effects;
increased administrative burden;
misclassification;
disproportionate scrutiny;
or diversion of resources away from genuinely urgent matters.

High sensitivity should therefore be accompanied by meaningful human review.

The purpose of a flag is to prompt attention.

It should not become the final conclusion.

Low-Priority Risk

Errors can also occur in the opposite direction.

A communication may be incorrectly classified as routine because:

the language is calm;
the person communicates indirectly;
a safeguarding concern is embedded in a long document;
unusual terminology is used;
the person has difficulty articulating urgency;
key information appears late in the correspondence;
or the system does not recognise the relevant context.

Systems should therefore be tested for both:

over-escalation

and

under-recognition.

A triage system that performs well only on obvious urgency may still create significant operational risk.

Accessibility and Communication Style

Triage systems should consider whether communication style affects prioritisation unfairly.

People may communicate differently because of:

disability;
neurodevelopment;
education;
language;
culture;
stress;
fatigue;
assistive technology;
AI-assisted drafting;
or personal communication preference.

An unusual communication style should not automatically result in:

lower priority;
higher risk;
suspicion;
or adverse classification.

The system should distinguish substantive indicators from stylistic difference wherever possible.

Long Communications

Important information may be buried inside long documents or correspondence.

Testing should examine whether:

early sections receive more weight;
later material is missed;
repeated information dominates;
context is lost;
or the system reduces long communication to a simplistic label.

Length alone should not determine priority.

A long communication may contain one highly significant issue.

A short communication may contain none.

Repetition and Priority

Repeated communication may influence triage.

That may sometimes be appropriate.

It can also create misleading assumptions.

Repeated correspondence may arise because:

the same issue has already been resolved;
a person is repeating an unchanged request;
the institution has not substantively answered the issue;
different questions are being expressed in similar language;
or earlier communication was misunderstood.

A triage system should not assume that:

repetition = low value

repetition = escalation

without considering context.

Previous Contact History

Historical information may legitimately inform prioritisation.

It can also create anchoring.

A person previously classified as:

high risk;
repetitive;
urgent;
non-urgent;
or difficult to route

may continue receiving similar classifications even after circumstances change.

Organisations should therefore ask:

How much weight does previous history receive?
Can new evidence override historical classification?
Are old labels periodically reviewed?
Can factual corrections affect future triage?
Is recency being balanced appropriately against relevant history?

Previous classification should inform context.

It should not become a permanent identity.

Human Review

Where triage affects access to time-sensitive or consequential services, meaningful human override should remain available.

A human reviewer should be able to:

inspect the underlying communication;
identify context the system may have missed;
correct a priority score;
escalate a wrongly deprioritised case;
reduce an unjustified escalation;
review contradictory information;
and document a different conclusion.

Human review should not simply confirm the AI’s ranking.

Override Must Work in Both Directions

A well-governed triage system should allow human reviewers to:

raise priority

and

lower priority.

If staff can escalate a case easily but cannot practically de-escalate it, the system may become directionally biased.

Likewise, if low-priority classifications are difficult to change, genuine concerns may remain buried.

Operational flexibility should exist in both directions.

Triage and Escalation

AI triage may feed directly into escalation pathways.

For example:

classification → priority score → escalation → intervention

Where that occurs, organisations should examine whether the chain contains sufficient human judgment.

A triage output should not silently become:

priority → risk → decision

without clear review.

The farther a triage score travels into consequential action, the greater the need for evidential discipline.

Triage and Source Traceability

A reviewer should ideally be able to understand why a case received a particular priority.

Relevant information may include:

which source material was analysed;
what signals influenced the score;
whether important records were missing;
whether historical information was considered;
whether the classification changed;
and whether a human overrode the result.

A priority decision that cannot later be explained may weaken accountability.

Testing Triage Systems

Triage systems should be tested using realistic communication conditions.

Testing may include:

calm descriptions of genuinely urgent issues;
emotionally intense but low-consequence messages;
very long correspondence;
short messages;
unusual terminology;
accessibility-assisted communication;
second-language communication;
repeated contact;
conflicting records;
incomplete information;
and material concerns embedded within otherwise routine information.

The purpose is to determine whether the system identifies substance, rather than merely familiar signals.

Measuring More Than Accuracy

A triage system should not be assessed only by overall accuracy.

Organisations may also examine:

false-positive rates;
false-negative rates;
differences across communication styles;
override frequency;
time to human review;
proportion of priority changes after review;
missed deadlines;
downstream consequences;
and recurring reasons for misclassification.

A system can achieve strong average performance while producing poor outcomes for particular types of communication.

Priority Should Be Reviewable Over Time

Priority may change as new information emerges.

Systems should permit reassessment where:

new evidence is supplied;
circumstances deteriorate;
urgency reduces;
a factual error is corrected;
a deadline changes;
or a human reviewer identifies missing context.

A priority score should not become permanently attached to a case merely because it was assigned first.

High-Stakes Environments

Stronger safeguards may be necessary where triage affects:

healthcare;
safeguarding;
education;
social care;
public services;
emergency response;
benefits;
housing;
disciplinary processes;
or access to essential support.

In these contexts, a triage error may alter:

response time;
access to review;
intervention;
or exposure to harm.

The greater the consequence, the stronger the case for:

human oversight;
transparent criteria;
testing;
reviewability;
correction;
and ongoing monitoring.

Triage Should Not Infer Human Characteristics

Particular caution is required if triage systems infer matters such as:

credibility;
motivation;
emotional stability;
cooperation;
intent;
capacity;
or future behaviour

from language alone.

Those are complex human judgments.

A communication system may appropriately identify:

language associated with urgency

without concluding:

this person is objectively high risk.

The distinction matters.

Questions for Organisations

Where AI is used for triage or prioritisation, organisations may ask:

What exactly is the system prioritising?
Which variables influence the ranking?
Does the system distinguish urgency from truth?
Are false positives understood?
Are false negatives understood?
Could unusual communication styles affect ranking?
Does communication length distort priority?
Does historical classification influence later decisions disproportionately?
Can a human raise or lower priority?
Can new information change the ranking?
Can the institution explain why a case received its priority?
Is prioritisation being used only to direct attention, or is it silently becoming substantive judgment?

SIAAF Relevance

This Guidance Note principally relates to:

Domain 02 — Decision Integrity & Human Oversight

Does a human remain capable of reviewing and overriding AI-generated priority classifications?

Domain 04 — Communication & Feedback Integrity

Does important information move through the system accurately enough to reach appropriate attention?

Domain 05 — Escalation, Challenge & Contestability

Can a priority decision be challenged, corrected or reconsidered?

Domain 06 — Risk, Harm & Operational Resilience

Does the organisation understand the consequences of both over-escalation and under-recognition?

It may also engage:

Domain 03 — Evidence & Traceability

Where the basis for a priority classification needs to remain visible and reviewable.

Domain 07 — AI Literacy & Organisational Readiness

Where staff need to understand that a triage output is an operational signal rather than a factual determination.

SWANK AI Standard

PRIORITISE ATTENTION, NOT TRUTH

AI triage may help determine:

what should be reviewed first

what requires human attention

what should be routed elsewhere

and

what may need escalation.

It should not silently determine:

credibility

entitlement

truth

final seriousness

professional judgment.

A priority classification is a routing decision.

It should not become a factual conclusion merely because an algorithm produced it.

Related SWANK AI Guidance

Guidance Note 01 — Handling AI-Assisted Complaints and Correspondence

Guidance Note 07 — Public-Sector AI and Accessibility

Guidance Note 08 — Reassessment Alongside Escalation

Guidance Note 12 — AI and Administrative Fairness

Guidance Note 15 — Testing AI Under Contradiction and Uncertainty

Guidance Note 20 — Governing AI in High-Stakes Environments

SWANK AI

Independent AI & Institutional Assurance

We do not just review AI. We review the institutional systems responsible for governing it.