Do AI Persona Differences Depend on the Task? Evidence Across Five Business Task Types

At a glance

Research snapshot

Research question

Why this question matters

Persona prompting is often used as if the role itself is the main source of variation.

But tasks differ too.

A model asked to diagnose a problem may naturally produce a different kind of answer from one asked to compare options or define a risk-control plan.

That creates an important design question.

If persona differentiation appears only in certain task types, researchers may be overgeneralizing from narrow prompt formats.

For example:

• Personas might matter for strategic comparison but not measurement
• They might matter for risk control but disappear in diagnosis
• Task structure might overwhelm persona framing entirely

Our broader V2 study had already found that persona-conditioned outputs remained strongly distinguishable after direct persona language was removed.

This analysis asks:

Does that differentiation hold across different kinds of work?

Task design

The five task families

The experiment used five response types.

1. Action plan

The model was asked to recommend concrete steps, experiments or next actions.

2. Comparison

The model compared approaches, options, trade-offs or criteria.

3. Diagnostic

The model identified likely causes, evidence and diagnostic sequences.

4. Measurement

The model defined baselines, metrics, validation plans or monitoring approaches.

5. Risk control

The model identified risks, guardrails, escalation rules or control mechanisms.

The original design contained 30 scenarios in each task family.

After one malformed response was excluded by the quality gate, the primary analysis retained:

• 30 action-plan scenarios
• 30 comparison scenarios
• 30 diagnostic scenarios
• 29 measurement scenarios
• 30 risk-control scenarios

Each complete scenario contained all five persona conditions.

Finding 1

Answer differentiation stayed high across every task family

Using the residual answer representation from the persona cue-masking analysis, mean persona-pair answer diversity was:

Every task mean remained above 0.84.

That matters because persona differentiation did not collapse in the more structured task types.

Measurement still showed residual answer diversity of 0.8478. Comparison, the lowest task, still measured 0.8463.

The practical result is not:

“Personas matter for one special kind of question.”

It is:

The observed persona-associated differentiation remained high across all five task families.

Finding 2

Risk control was highest, comparison lowest—but the spread was small

Risk control had the highest observed answer diversity: 0.8614.

Comparison had the lowest: 0.8463.

Difference: 0.0150.

That difference is about 1.76% of the overall level of task differentiation.

The study prospectively defined a larger task-specificity signal as requiring an answer-diversity range of at least 0.03, together with corrected contrast evidence, cross-cluster direction stability and supporting interaction evidence.

The observed range did not reach that threshold.

The primary task conclusion was therefore:

Broadly persistent across tasks

rather than:

Task-conditioned differentiation

Risk-control prompts did produce somewhat more differentiated answers, but not enough to support the claim that persona conditioning matters dramatically more when the task is about risk.

Finding 3

Grounded search-query differentiation was also high across all five tasks

The task analysis did not examine final answers alone.

It reused the lexical search-query divergence from the grounded-query study.

Mean lexical query divergence was:

Risk control again ranked highest. Measurement ranked lowest.

The contrast between risk control and measurement was 0.0343, with a 95% CI of 0.0160–0.0514 and corrected p-value 0.040. The direction agreed across all five topic clusters.

So query formulation showed somewhat more task variation than final answer diversity.

But the absolute level of query differentiation remained high in every task.

Open full-resolution figure

Finding 4

SEO Manager was the most distinctive answer condition in every task

For each task, we measured a persona's average answer distance from the other four conditions.

SEO Manager ranked highest in all five task families.

The consistency is notable.

SEO Manager was not merely distinctive for technical measurement tasks. It was also most distinctive for comparison, diagnosis, action planning and risk control.

But distinctive is not better.

The analysis does not establish that SEO Manager answers were more accurate, complete, useful or better researched.

Open full-resolution figure

Finding 5

CMO and Founder/CEO were the nearest pair across every task

Across all five task families, the closest observed pair was:

CMO / Founder/CEO

That does not mean the two conditions were interchangeable.

A distance around 0.82 is still substantial, and lexical proximity does not establish that they produce the same facts, recommendations, priorities or risks.

Interaction evidence

Did persona-task interactions matter?

A strongly task-conditioned persona pattern would imply that the relative distinctiveness of personas changes meaningfully by task.

For example, SEO might be highly distinctive in measurement but ordinary in comparison, or CMO might separate strongly in risk control but converge elsewhere.

That pattern was not strong enough to satisfy the study's predefined task-conditioning rule.

Persona profiles were relatively stable across tasks.

The analysis therefore did not identify a clear persona × task interaction large enough to overturn the broad-persistence conclusion.

This matters for study design.

If the role structure changed dramatically by task, a research team might need entirely different persona panels for different question types.

The historical data did not support that conclusion.

Robustness

The pattern also held across topic clusters

Each task family appeared across five topic clusters:

• AI optimisation
• Answer Alignment
• Content authority
• Conversion/pipeline
• Search visibility

Across task × cluster cells, answer differentiation ranged approximately 0.825–0.884.

Lexical query differentiation ranged approximately 0.836–0.927.

These cells contained only five or six scenarios each, so they should not be treated as standalone findings.

But they provide a useful robustness check: the broad task-persistence result was not obviously driven by one topic cluster.

Study design

What this means for AI research design

The practical implication is not that task framing can be ignored.

Task design still changes what the model is asked to do, what evidence may be relevant, the structure of the answer and what success looks like.

But if a research programme is using personas to sample different decision contexts, this experiment suggests you should not assume persona differentiation disappears just because the task becomes more structured.

In this study, persona-associated differences remained high when the model was asked to:

• Diagnose
• Measure
• Compare
• Plan
• Control risk

That supports a more deliberate matrix design:

persona × task

rather than choosing either personas or tasks as the only source of variation.

Interpretation

Implications for AI Answer Alignment

For Answer Alignment research, different task forms can represent different moments in a buyer or stakeholder decision.

A company may be evaluated through questions such as:

• What is going wrong?
• How should we measure it?
• Which option is better?
• What should we do next?
• What could go wrong?

If persona-associated differences remain high across all of those task forms, then testing only one task style may under-sample the ways an AI system can represent the same company or category.

That does not mean every research programme needs all five personas × all five tasks.

It means the dimensions are not obviously interchangeable.

Persona variation remained visible even after the task changed.

Open question

The unresolved question: does task-persistent diversity remain substantive?

The main V2 programme now shows that persona-associated differentiation:

• Survives direct cue masking
• Appears in grounded search-query formulation
• Persists across five task types
• Remains materially additive in a five-persona panel

But those results are still primarily representational.

The next question is whether the task-persistent differences correspond to different factual propositions, recommendations, decisions and risks.

That substantive layer is currently awaiting independent human calibration.

Until that validation is complete, the correct conclusion is:

Persona-associated representational diversity persists across tasks. Whether substantive decision diversity persists to the same degree is not yet established.

Conclusion

Conclusion

Persona-conditioned AI differences did not disappear when the task changed.

Across action planning, comparison, diagnosis, measurement and risk control, residual answer differentiation remained consistently high.

Risk control produced the highest observed answer diversity. Comparison produced the lowest.

But the overall spread was small: 0.015, or about 1.76% of the overall differentiation level.

Search-query differentiation showed somewhat more task variation, but remained high across all five task types.

The strongest conclusion is therefore:

In this matched grounded-AI experiment, persona-associated answer differentiation was broadly persistent across business task types rather than confined to one narrow prompt format.

That makes task design an important companion variable—not a substitute for persona design.

For the broader programme, see Persona Prompts Change More Than Tone: Evidence from 750 Grounded AI Responses.

Research context

Related Kojable research

• Persona Prompts Change More Than Tone: Evidence from 750 Grounded AI Responses

  The V2 capstone.

• How Many AI Personas Do You Need? What a Five-Persona Panel Actually Preserves

  Persona-panel coverage.

• Do Persona Prompts Change How AI Searches? Evidence from 1,937 Grounded Search Queries

  Query-layer differentiation.

• Beyond Persona Labels: Do AI Personas Still Differ After Explicit Cues Are Removed?

  Cue-ablation research.

• Do AI Role Prompts Change the Decision Lens?

  Separate marketing-role framing analysis.

• AI Answer and Source Research

  Browse the Kojable Research index.