# Agent Operations ### **How AI Skills Operate in Audience Work** Agent Operations defines how AI Skills behave when running Audience workflows. It covers how a Skill receives input, identifies source type, extracts and evaluates specific attributes, rates credibility, produces the Audience Definition Card, routes to the next step, and handles ambiguity. Skills operate against specific attributes — not just attribute type labels. When a Skill evaluates behavioral coverage, it checks for frequency of use, experience level, feature adoption, interaction mode, and task orientation. Each type has a defined set of sub-attributes the Skill is looking for. Missing sub-attributes are flagged by name, not by type. Confidence thresholds govern when a Skill proceeds independently and when it escalates to a human. Everything below the threshold requires clarification before the Skill continues. ## **1\. Intake and Source Identification** The first thing a Skill does when receiving audience data is identify the source type. Source identification is non-negotiable — it determines credibility, which determines every subsequent step. ### **Source Identification Rules**

Input Signal	Source Type	Credibility Starting Point
Named Helio audience groups with poll response breakdowns	Helio Audience Group	High — confirm attributes are behavioral before finalizing
Demographic targeting criteria from Facebook, Google, or programmatic platforms	Advertising Segment	Low — not qualified for design decisions without behavioral validation
CRM exports with lifecycle stage, tenure, activation milestones, or feature adoption data	CRM / Lifecycle Segment	Medium — confirm behavioral attributes are present before proceeding
Survey data with defined respondent criteria and behavioral screening	Survey — treat as Medium CRM equivalent	Medium — validate assumed attributes before drawing conclusions
Forum posts, app store reviews, community threads	Qualitative — contextual and motivational signals only	Medium for motivation and context, Low for behavioral and lifecycle
Statements from team members, stakeholders, leadership without external data reference	Internal Assumption	Low — document as hypothesis, do not proceed to design without validation
Existing personas or segments without a named research source	Unknown — escalate	Cannot rate until source is confirmed. Ask: where did this come from?

If the source cannot be identified from the input, the Skill escalates before proceeding. Source type is never inferred from content alone. ## **2\. Attribute Extraction** After source identification, the Skill extracts attributes from the input. Extraction operates against the specific sub-attributes within each of the five types. The Skill identifies which sub-attributes are present, which are missing, and which are assumed rather than evidenced. ### **Behavioral Attributes — Extraction Targets**

Sub-Attribute	Present Signal	Missing Signal	Assumed Signal
Frequency of use	Data states daily, weekly, or monthly usage rate explicitly	No usage frequency mentioned	Frequency described in general terms without a data source (e.g. “regular users”)
Experience level	Data distinguishes first-time, returning, or advanced users	No experience level mentioned	Team assigns level based on role label, not behavioral evidence
Feature adoption	Data names specific features used or not used	No feature use mentioned	Assumed from product tier or plan without usage data
Interaction mode	Data specifies device or input method explicitly	No device or input mentioned	Device assumed from demographic or market segment
Task orientation	Data describes whether users are goal-driven, exploratory, or transactional	No task approach mentioned	Orientation described without behavioral evidence

### **Motivational Attributes — Extraction Targets**

Sub-Attribute	Present Signal	Missing Signal	Assumed Signal
Primary motivation	Data states what outcome users are seeking (speed, trust, mastery, discovery, recognition)	No motivation mentioned	Motivation inferred from product category without asking users
Value orientation	Data describes what trade-offs users accept (efficiency vs. creativity, reliability vs. savings)	No value orientation mentioned	Value assumed from target market positioning
Risk tolerance	Data distinguishes early adopters from cautious users	No risk posture mentioned	Risk assumed from lifecycle stage without attitudinal data
Decision style	Data describes whether users prefer guided or exploratory paths	No decision style mentioned	Style assumed from role or seniority
Confidence level	Data captures perceived competence or self-reported difficulty	No confidence data mentioned	Confidence assumed from experience level

### **Contextual Attributes — Extraction Targets**

Sub-Attribute	Present Signal	Missing Signal	Assumed Signal
Device or platform	Data names specific device, OS, or platform	No device context mentioned	Device assumed from demographic or market segment
Environment	Data describes physical setting (workplace, home, field, transit)	No environment mentioned	Environment assumed from role without contextual evidence
Connectivity	Data addresses online, offline, or low-bandwidth conditions explicitly	No connectivity mentioned	Connectivity assumed from environment
Workload and time pressure	Data describes competing priorities or session length constraints	No workload data mentioned	Time pressure assumed from role or frequency
Collaboration mode	Data distinguishes solo use from team or shared system use	No collaboration mode mentioned	Mode assumed from org size or product tier

### **Lifecycle Attributes — Extraction Targets**

Sub-Attribute	Present Signal	Missing Signal	Assumed Signal
Stage of relationship	Data explicitly names lifecycle stage: trial, active, habitual, power, dormant, churned	No lifecycle stage mentioned	Stage inferred from activity level without explicit segmentation
Tenure	Data states duration of product use in days, months, or years	No tenure mentioned	Tenure inferred from plan type or cohort date
Activation milestone	Data identifies specific actions completed: onboarded, upgraded, first purchase	No activation milestone mentioned	Milestone assumed from account status
Frequency of engagement	Data states engagement cadence: daily active, weekly, monthly	No engagement frequency mentioned	Frequency assumed from subscription tier
Renewal or purchase cycle	Data describes subscription, one-time, or recurring purchase pattern	No purchase cycle mentioned	Cycle assumed from product category

### **Demographic Attributes — Extraction Targets**

Sub-Attribute	Present Signal	Missing Signal	Assumed Signal
Role or profession	Data states job function explicitly	No role mentioned	Role inferred from company size or industry
Industry or discipline	Data names specific industry or domain	No industry mentioned	Industry assumed from company type
Region or language	Data specifies region or language of participants	No region or language mentioned	Region assumed from CRM or billing address
Age or education	Data includes age range or education level when relevant to comprehension or accessibility	Not mentioned — acceptable if not relevant to the design situation	Assumed from demographic targeting criteria
Accessibility needs	Data identifies assistive technology use or accommodation requirements	Not mentioned — flag if design involves accessibility-sensitive decisions	Never assumed — escalate if relevant and missing

### **Extraction Output** After running extraction, the Skill produces an attribute summary with three lists: • **Present attributes:** Sub-attributes confirmed by evidence in the input data, named specifically. • **Missing attributes:** Sub-attributes not mentioned or measurable from the input, named specifically. • **Assumed attributes:** Sub-attributes described without an evidenced data source, named specifically with a flag. Assumed attributes reduce credibility. Three or more assumed attributes in a single type lower the overall credibility rating by one level. ## **3\. Confidence Thresholds and Credibility Rating** The Skill combines source type credibility with attribute extraction results to produce a final credibility rating and a confidence score. Confidence determines whether the Skill proceeds independently or escalates. ### **Confidence Scoring Model**

Factor	Full Credit	Partial Credit	No Credit
Source type	Helio Audience Group	CRM, Survey, or Qualitative	Advertising Segment or Internal Assumption
Behavioral attributes	3+ sub-attributes present with evidence	1–2 sub-attributes present; remainder missing or assumed	No behavioral sub-attributes present
Motivational attributes	2+ sub-attributes present with evidence	1 sub-attribute present; remainder missing or assumed	No motivational sub-attributes present
Contextual attributes	2+ sub-attributes present with evidence	1 sub-attribute present; remainder missing or assumed	No contextual sub-attributes present
Lifecycle attributes	Stage and at least one additional sub-attribute present	Stage only, no supporting sub-attributes	No lifecycle sub-attributes present
Assumed attribute count	0–1 assumed sub-attributes total	2–3 assumed sub-attributes total	4+ assumed sub-attributes total
Group definition	Defined by behavior or motivation, not by label	Partially behavioral with some label dependence	Defined entirely by demographic label or role title

### **Confidence Thresholds**

Confidence Level	Score Range	Skill Behavior	Output
High Confidence	Full credit on source + 4–7 factors at full or partial credit	Skill proceeds independently. Completes the Audience Definition Card, assigns credibility rating, and routes to the next step without escalation.	Complete Audience Definition Card with High or Medium credibility rating
Medium Confidence	Partial credit on source + 3–4 factors at full or partial credit	Skill completes a partial card, flags all assumed attributes by name, rates credibility, and notes what validation is needed before Collecting can begin. Does not route to Collecting without flagging the gap.	Partial Audience Definition Card with credibility note and named gaps
Low Confidence	No credit on source, or fewer than 3 factors at any credit level	Skill stops. Does not attempt to complete the card. Escalates to a human with a specific statement of what is missing and what source type is needed to fill it.	Escalation message with named gaps and recommended next step

## **4\. Output Contract** The Audience Definition Card is the Skill’s primary output. Every completed card must meet the minimum output contract before it is handed off to Collecting or returned to the team. ### **Minimum Output Contract**

Field	Minimum Requirement	What Fails the Contract
Audience Name	Descriptive name built from defining attributes, not a demographic label or role title	Name is a marketing label, job title, or demographic category without behavioral grounding
Audience Type	One of the four types: Participants, Customers, Stakeholders, Project Team	Type is missing or listed as “Unknown” without escalation
Source	Named data source with source type identified	Source listed as “various” or left blank
Defining Attributes	Minimum two behavioral or motivational sub-attributes confirmed by evidence, named specifically	Only demographic or lifecycle attributes present with no behavioral or motivational evidence
Design Situation	One to two sentences describing the conditions that make this group experience the situation differently	Design situation is absent or restates the audience name without adding interpretive content
Connected UX Metrics	Minimum one UX metric connected to a defining attribute	No UX metric present
Credibility Rating	High, Medium, or Low with a one-line rationale	Credibility rating missing or stated without rationale
Gaps	Named missing sub-attributes, or “None identified” if coverage is complete	Gaps field left blank when assumed attributes are present

A card that fails the minimum contract is not handed off. The Skill flags which fields are incomplete, states what input is needed to complete them, and holds the card until the gaps are resolved or the escalation is acknowledged. ## **5\. Routing Logic** After producing the Audience Definition Card, the Skill routes to the next step based on credibility rating and card completeness.

Condition	Route	What the Skill Communicates
Card complete, credibility High	Collecting	State the audience type, the defining attributes, the connected UX metrics, and the recommended Collecting method based on the audience type. Flag any gaps for follow-up.
Card complete, credibility Medium	Collecting with validation note	State what validation step Collecting needs to include before drawing conclusions from assumed attributes. Name the assumed sub-attributes specifically.
Card partial, credibility Low	Back to Audience — Helio study needed	State which sub-attributes are missing, which source type would fill them, and what kind of Helio study or screener would establish a behaviorally defined group.
Design situation does not match stated user need	User Needs	State the mismatch clearly: the audience definition points to a situation the current user need statement doesn’t capture. Return to User Needs with the attribute profile as new evidence.
Multiple groups detected in input data	Back to Audience — split required	Name each group by its distinguishing attributes. Flag that averaging them will produce unreliable findings. Request confirmation on which group to prioritize before continuing.
Advertising and Helio data diverge	Stakeholders	State the divergence specifically: what the advertising segment describes vs. what the Helio group describes. Flag this as a strategy question, not a usability question.
Source cannot be identified	Escalate to human	State that source type could not be determined from the input. Ask: where did this data come from? Do not rate credibility or proceed until source is confirmed.

**Routing Across Role Tiers** Most Audience work operates at the Contributor tier — practitioners defining groups, extracting attributes, and completing Audience Definition Cards for use in Collecting. Two situations signal that the work has crossed into Expert or Leader territory and requires a different kind of response. Expert crossover: When audience signals from different sources conflict and the resolution requires interpreting what the divergence means for the design direction. This is no longer card completion work — it is analytical synthesis that requires a practitioner with cross-functional context. The Skill flags the conflict, names the diverging sources, and routes to a human for interpretation before proceeding. Leader crossover: When the audience definition surfaces a strategy question — who should the product serve, does the current targeting reflect the actual user base, is audience drift significant enough to change product direction. These questions cannot be resolved within the Audience workflow. The Skill names the question explicitly, states that it requires a leadership decision, and holds the workflow until that decision is made and communicated back. ## **6\. Ambiguity Handling** Ambiguity in Audience work takes three forms: source ambiguity, attribute ambiguity, and group boundary ambiguity. Each has a defined handling rule. ### **Source Ambiguity** Source ambiguity occurs when the input contains audience data but the origin is unclear. Common examples: a persona deck without a research citation, a segment labeled by the team without naming its source, a spreadsheet with no column headers explaining the data type. • **Rule:** Do not infer source type from content. Ask directly: where did this data come from? Do not rate credibility until the source is confirmed. • **Escalation trigger:** Any input that contains audience definitions without a named source. ### **Attribute Ambiguity** Attribute ambiguity occurs when a description could map to more than one sub-attribute, or when the same language is used for both evidence and assumption. Common examples: “active users” (frequency? lifecycle stage? both?), “mobile users” (device? environment? interaction mode?). • **Rule:** When a description maps to more than one sub-attribute, extract all plausible matches and flag each as “possibly present — confirm.” Do not collapse ambiguous descriptions into a single attribute without confirmation. • **Escalation trigger:** More than two ambiguous sub-attributes in the same type, or ambiguity in the behavioral type specifically, which is the most consequential for credibility rating. ### **Group Boundary Ambiguity** Group boundary ambiguity occurs when the input contains data that appears to describe more than one audience but has not been formally split. Common examples: Helio results that show bimodal distributions, CRM segments that contain both habitual and dormant users, survey data with high variance across respondents. • **Rule:** Flag the potential split. Name the distinguishing attribute that suggests two groups are present. Do not produce a single Audience Definition Card that spans both. Ask: should these be treated as separate groups? • **Escalation trigger:** Any dataset where the variance in a single behavioral or motivational sub-attribute is high enough that a single design situation cannot be written to cover both ends. ## **7\. Escalation Rules** Escalation is not a failure state. It is the Skill correctly identifying that a human decision is required before the work can proceed. Escalation messages are specific, not generic.

Escalation Trigger	What the Skill Says	What It Does Not Say
Source cannot be identified	“I can see audience data here but I can’t determine where it came from. Before I can rate credibility or extract attributes, I need to know: is this from a Helio study, a CRM export, an advertising platform, or an internal assumption?”	“I don’t have enough information to proceed.” — too generic to act on
Confidence score is Low	“This audience definition doesn’t have enough behavioral evidence to support a design decision. Specifically, [name the missing sub-attributes]. To move forward, the team needs a Helio [poll / screener / usability study] that establishes [name the group] by behavior rather than [label / assumption].”	“More research is needed.” — too vague to act on
Multiple groups detected	“The data appears to contain two distinct groups: [name Group A by its distinguishing attributes] and [name Group B by its distinguishing attributes]. Combining them will produce unreliable findings. Which group should I build the Audience Definition Card for first?”	“Your audience may need further segmentation.” — doesn’t name the groups or the decision
Design situation doesn’t match stated user need	“The audience attributes point to [describe the situation]. The current user need statement describes [describe the stated need]. These don’t align. Before Collecting begins, the team should return to User Needs with this attribute profile and confirm whether the need statement needs to be updated.”	“There may be a mismatch between your audience and your user need.” — doesn’t identify the specific gap
Accessibility needs relevant but absent	“This design situation involves [name the accessibility-sensitive element]. The audience definition doesn’t include accessibility needs as an attribute. This can’t be assumed — should this be added as a defining attribute, and if so, what accommodation requirements apply to this group?”	“Accessibility attributes may be missing.” — doesn’t name the specific design element or the decision required

## **8\. Skill Boundaries** The Audience Skill does not replace human judgment in these situations: • **Strategic decisions about who the product should serve.** The Skill can surface that the audience is drifting or that advertising and Helio data diverge. It cannot decide which audience the product should prioritize. That is a leadership decision. • **Choosing a Collecting method.** The Skill routes to Collecting with a brief based on the Audience Definition Card. Collecting makes the method decision based on the audience type, attributes, and connected UX metrics. • **Resolving stakeholder disagreement about the customer.** The Skill can show where signals conflict. It cannot align stakeholders. That requires a human conversation with the evidence the Skill has surfaced. • **Determining whether a new audience represents a better opportunity.** The Skill can detect drift and document the new group. Whether to pursue it is a product strategy decision. • **Interpreting qualitative data without behavioral evidence.** Forum posts, community threads, and app store reviews surface motivational and contextual signals. The Skill flags these as Medium credibility for those types only and does not elevate overall credibility without behavioral evidence from a separate source. When the Skill reaches a boundary, it names what it has surfaced, states that a human decision is required, and holds the workflow at that point. It does not proceed past a boundary by filling in assumptions.