PaperlayerView exact revision on arXiv

Source-linked AI summary

AI Tools Adoption across the Double Diamond Workflow: Phase, Mode, and Barriers in Designer Practice

Sepideh Tajarmakan, Khashayar Hojjati Emami

arXiv:2609.34655v2cs.HC

TL;DR

Existing research offers limited large-scale evidence about phase- and tool-specific AI use and the concerns distinguishing adopters from non-adopters. The authors surveyed designers across the Double Diamond and found that tool choices vary by phase, while perceived usefulness distinguishes the groups on reported barriers.

  • Problem

    Existing research offers limited large-scale evidence on reported AI use by design phase and tool, workflow-wide intensity, and the specific barriers that distinguish adopters from non-adopters.

    Source: INTRODUCTION · p. 2, Non-Adoption Is Described as a General… · p. 3

  • Method

    The study surveyed 443 designers and classified 73 AI tools by output and function to examine reported use across the Double Diamond phases.

    Source: AI Tool Category across the Design… · p. 11, CONCLUSION · p. 18

  • Results

    Conversational use led in Discover and Define, AI-native image generation led in Develop, and perceived usefulness was the only reported barrier distinguishing adopters from non-adopters.

    Source: DISCUSSION · p. 17

  • Takeaways & Limitations

    AI adoption in design is better treated as several phase- and mode-specific behaviors than as a single behavior measured once.

    Source: CONCLUSION · p. 19

    The non-probability sample was concentrated in Iran, limiting generalization, and the cross-sectional design does not support causal inference.

    Source: Limitations · p. 18

Abstract

from arXiv · show

Designers are adopting AI faster than the tools built for them can keep up. General-purpose conversational AI led AI use among designers in three of the four phases of design work, ahead of tools built for design. We surveyed 443 designers from at least 43 countries about AI use in each phase of the Double Diamond: Discover, Define, Develop and Deliver. Tools were classified by output and function. Of these designers, 79.7% confirmed AI use in at least one phase. Adopters used AI in a mean of 2.89 phases, and only 35.1% used it in all four. They added phases in workflow order, from Discover onward. Conversational tools led in Discover, Define and Deliver. AI-native image generation led in Develop, where AI-native tools peaked at 57.6% of tool picks. Embedded AI features in existing design software peaked later, in Deliver (37.5%). Fewer designers used AI in later phases, from 318 in Discover to 182 in Deliver. Those who did named a similar number of tools per phase. Low perceived usefulness was the only barrier that separated non-adopters (40.5%) from adopters (16.4%) after correction for multiple tests. Concern about accuracy was similar in both groups.

INTRODUCTION

The study uses the Double Diamond to examine AI use across design phases, addressing gaps in large-scale evidence about phase-specific behavior, tools, and non-adoption. It reports AI use by phase, tool type, and barrier in a large designer sample.

Analytical framework

  • The Double Diamond organizes the study into Discover, Define, Develop, and Deliver, from exploration and problem definition through solution development and finalization.The model is presented as suitable for UX/UI, product, industrial, and graphic design.Source: INTRODUCTION · p. 2

Research gaps

  • Prior research lacks large-scale evidence of reported AI behavior by design phase and tool, workflow-wide usage depth, and specific differences between adopters and non-adopters.Existing studies often use small samples or literature-derived evidence; larger studies examine acceptance intention or national contexts rather than phase- and tool-specific behavior.Source: INTRODUCTION · p. 2

Study scope

  • The study surveyed 443 designers from at least 43 countries and regions, including UX/UI, product, industrial, and graphic designers, students, and related roles.Closed and open-ended questions followed each Double Diamond phase and addressed four research questions.Source: INTRODUCTION · p. 2
  • The paper reports AI use by phase, tool type, and barrier, distinguishing AI-native applications from AI features embedded in existing design software.Examples include ChatGPT and Midjourney as AI-native applications and Photoshop’s Generative Fill as an embedded feature.Source: INTRODUCTION · p. 2

LITERATURE REVIEW

The review frames AI tools by their observable outputs and functions, then identifies gaps in evidence about phase-specific use, tool choice, usage depth, and non-adoption.

AI Tool Definition

  • The paper treats an AI tool as the application through which a model is put to work on a task, and classifies tools by output and function within that output.The review presents these dimensions as easier for designers to report than internal properties of AI systems.Source: AI Tool Definition · p. 3

AI Concentrates Unevenly Across the Design Process

  • Prior work describes AI concentration in later design stages or differentiated roles by stage, while built-tool research remains distinct from evidence of what designers report doing.Lee et al. (2025) report more built systems for Develop and Deliver; Mirzaei et al. (2026) find stage-differentiated roles without concentration at either end.Source: AI Concentrates Unevenly Across the… · p. 3

Tool Choice and Usage Depth Remain Underspecified

  • Research often treats AI use generally, with uneven tool-level workflow coverage and little evidence on intensity across the full design process.Veggi (2025) documents strong availability for text and image tasks but weaker coverage for 3D modeling and evaluation; Tang et al. (2024) examine intensity for one tool category.Source: Tool Choice and Usage Depth Remain… · p. 3

Non-Adoption Is Described as a General Attitude

  • Prior non-adoption research profiles general orientations, but does not compare adopters and non-adopters item by item on the same barriers measure.Fang et al. (2026) identify five adopter personas with distinct barriers and motivations.Source: Non-Adoption Is Described as a General… · p. 3

METHOD

The study uses a cross-sectional survey and a tool taxonomy to describe AI use across the Double Diamond, with analysis tailored to phase use, tool counts, and reported barriers.

Study Design

  • A cross-sectional online questionnaire collected self-reported AI-use data from designers between August 2025 and February 2026, favoring a broad snapshot over tracking individual change.The final sample of 443 valid responses exceeded the a priori estimate of approximately 385.Source: Study Design · p. 4, Sample Size · p. 4

Sampling Procedure

  • Recruitment combined purposive, convenience, and snowball sampling through professional networks and design-focused social channels, targeting professionals and students in several design roles.Experience and country were recorded but were not eligibility criteria.Source: Sampling Procedure · p. 4

Consent and Ethical Considerations

  • Participation was voluntary and confidential; respondents consented before proceeding, and 34.1% provided an email address for possible follow-up interviews.Although the consent statement described the study as anonymous, the volunteered emails mean some responses could in principle be linked to a person.Source: Consent and Ethical Considerations · p. 4

Questionnaire Structure

  • The 70-item questionnaire routed respondents according to general and phase-specific AI-use answers, with phase tool selections and barriers asked through branching sections.Phase items were answered only by those reporting use in that phase; samples declined from 318 in Discover to 182 in Deliver.Source: Questionnaire Structure · pp. 4–5

Data Cleaning and Filtering

  • Researchers standardized tool names and open-text entries, excluded unmatched tool mentions from counts, and merged phase-specific datasets without altering cleaned values.Open-text entries naming multiple tools were split into separate binary tool mentions; no thematic or interpretive analysis was conducted.Source: Data Cleaning and Filtering · pp. 5–6

AI Tool Classification

  • The study identified 73 distinct tools, classified by primary output and function; tools were assigned their most established use, while agentic tools were excluded from phase-level interpretation because they were reported only twice.The taxonomy covered five output categories and ten function-nests; tools with multiple uses were classified by their primary use at data collection.Source: AI Tool Classification · p. 6

Data Analysis Plan

  • Non-parametric tests were used for group and phase comparisons, with Holm correction for phase and barrier post-hoc testing and effect sizes reported where specified.Barrier endorsement was compared between adopters and non-adopters using separate chi-square tests for ten categories, followed by Holm correction.Source: Data Analysis Plan · pp. 6–7

Validity and Reliability

  • The authors report validity and reliability checks, but note that non-probability sampling concentrated in Iran limits external validity and that phase-use items did not always agree.Two raters showed almost perfect agreement for tool Output and Function classification (κ = 0.937 and κ = 1.00, respectively).Source: Validity and Reliability · pp. 7–8

RESULT

In this sample, most designers reported confirmed AI use, usually across only part of the Double Diamond, with partial adoption accumulating from Discover onward. Tool choices varied by phase, while fewer designers reported use in later phases; low perceived usefulness was the only barrier that differed significantly between adopters and non-adopters after correction.

RESULT

  • The 443-person sample covered at least 43 countries and regions, but Iran accounted for 47.0%, so results should not be read as broadly cross-national.Product and Industrial Designers were each 26.9% of participants; 60.3% were aged 20 to 29.Source: RESULT · p. 8, Table 3 · p. 8

AI Adoption across the Double Diamond

  • 79.7% of designers confirmed AI use in at least one phase; adopters used AI in a mean of 2.89 phases, and 35.1% used it in all four.Partial adopters most often added phases in workflow order, from Discover onward; daily users averaged 3.21 phases versus 2.33 among those using AI a few times a month.Source: AI Adoption across the Double Diamond · p. 9, Table 4 · p. 9

AI Tool Category across the Design Process

  • The study identified 73 distinct AI tools; counts ranged from 32 in Deliver to 47 in Develop, and adopters picked a mean of 1.80 to 2.09 tools per phase.Source: AI Tool Category across the Design… · p. 10
  • Text led in Discover and Define, Image led in Develop (72.1%), and Deliver combined Text with Image.Video was absent in Define, then grew through Develop and Deliver, exceeding 3D Geometry in both.Source: AI Tool Category across the Design… · p. 10, Figure 2 · p. 11
  • By function, ChatGPT accounted for 45.0% of Discover picks; Conversational tools rose in Define, Image Generation led in Develop, and Conversational tools led again in Deliver.Discover’s embedded design features outpaced AI-native image generation; Define shifted from Figma AI to FigJam AI, while Develop’s AI-native image generation overtook Conversational tools 51.5% to 13.0%.Source: AI Tool Category across the Design… · pp. 11–12
  • AI-native tools peaked at 57.6% of picks in Develop, while embedded AI peaked at 37.5% in Deliver; just 9 of 73 tools were Conversational & General-Purpose, yet they drew most picks in three phases.The two modes peaked in different phases, with embedded AI leading over AI-native tools in Deliver.Source: AI Tool Category across the Design… · pp. 12–13, Table 7 · p. 13

AI Use Intensity across Workflow

Barriers within AI Adopters and Non-adopters

DISCUSSION

The discussion finds that AI use is concentrated in early workflow phases despite the later-stage focus of much purpose-built tooling, with different tool modes peaking at different phases. Usefulness alone separated adopters from non-adopters, while the authors emphasize that the descriptive, geographically concentrated survey supports limited generalization and no causal claims.

DISCUSSION

  • Discover had the highest phase adoption, led by Conversational, General-Purpose tools, reversing the built-tool emphasis on Develop and Deliver.The discussion says early-stage work was supported by a tool not designed for design.Source: DISCUSSION · pp. 16–17
  • Define was also among the two most AI-saturated phases, with adoption showing no gap relative to the other phases.The authors say this agrees with prior findings on LLM support across phases.Source: DISCUSSION · p. 17
  • At every breadth level, leading adoption combinations added workflow phases in order rather than skipping ahead.Frequent users also spread AI use across more phases, so habitual use and breadth went together in this dataset.Source: DISCUSSION · p. 17
  • Just 9 of 73 identified tools were general-purpose, yet they accounted for most picks in three phases; 14 tools appeared in all four phases.Another 33 tools appeared in only one phase.Source: DISCUSSION · p. 17
  • Develop was the phase where AI-Native Image Generation led, while AI-Native and Embedded modes peaked in different phases.The authors suggest this may reflect a shift from directing conversation to curating machine output; Deliver had the lowest adoption, with fewer users in later phases but no reduced intensity among engaged users.Source: DISCUSSION · p. 17
  • Perceived usefulness was the only barrier distinguishing adopters from non-adopters; unfamiliarity, ethics, and organizational restrictions did not separate the groups.The authors caution that these descriptive patterns do not establish mechanisms.Source: DISCUSSION · p. 17

Limitations

  • The sample was concentrated in Iran (47.0%) and recruited through unpaid, non-probability methods, limiting generalization; the cross-sectional design rules out causal inference.Phase-specific samples also declined from Discover (n = 318) to Deliver (n = 182), weakening comparisons for small subgroups; measures were self-reported.Source: Limitations · p. 18
  • The English instrument, preset tool lists, manual classification of open-text responses, and unvalidated custom items further constrain measurement.The general multi-select item showed a recognition gap against phase-specific screening questions and would have undercounted adoption if used alone.Source: Limitations · p. 18

CONCLUSION

This study maps reported AI adoption across the Double Diamond, showing that tool use varies by phase and collaboration mode. Its conclusion is that design AI adoption is better treated as several behaviors than as a single behavior measured once.

CONCLUSION

  • The study examined reports from 443 designers using a taxonomy of 73 tools classified by output and function across four Double Diamond phases.It considered adoption, workflow breadth, tool choice, intensity, and recurring tools.Source: CONCLUSION · p. 18
  • Conversational and general-purpose AI filled early-stage work, while AI-Native and Embedded AI use peaked in different phases.The conclusion frames these patterns as pointing to two distinct modes of collaboration.Source: CONCLUSION · pp. 18–19
  • Usefulness was the only concern that separated adopters from non-adopters; unfamiliarity and ethics did not.Source: CONCLUSION · p. 19
  • The authors present phase and mode patterns as a practical map of where AI currently earns a place and where human judgment is still preferred.Source: CONCLUSION · p. 19

CRediT Author Statement

The end matter records contributor roles, AI-assisted editing and figure-design uses, data access conditions, and funding and competing-interest statements.

CRediT Author Statement

  • Sepideh Tajarmakan is credited with conceptualization, methodology, investigation, data curation, analysis, software, visualization, and original drafting; Khashayar Hojjati Emami with supervision, validation, and review and editing.Source: CRediT Author Statement · p. 19

Declaration of Generative AI Use

  • Grammarly assisted with language editing and Claude with figure design; the authors conducted and take responsibility for all analysis, testing, and interpretation.Source: Declaration of Generative AI Use · p. 19

Data Availability Statement

  • Questionnaire, taxonomy, cleaned data, and validity files are restricted pending completion of the four-paper series, with case-by-case access requests.The records exclude participant-identifying information.Source: Data Availability Statement · p. 19

Funding

Declaration of Interest Statement

Appendix Table 1: Example of Binary Tool Coding (n = 3 Participants)

The appendix illustrates binary coding of named or selected tools in participant responses.

Appendix Table 1: Example of Binary Tool Coding (n = 3 Participants)

Appendix Table 3: Nests Function within Each Output by Phase, with AI Tool Names

Appendix Table 3 organizes tools by output, function, and phase. Its percentages are shares of tool selections within each phase, not shares of participants.

Appendix Table 3: Nests Function within Each Output by Phase, with AI Tool Names

Appendix Table 4: Tool Identification Coverage for AI Adopters, by Phase

The table header identifies tool, count, and percentage fields.

Appendix Table 4: Tool Identification Coverage for AI Adopters, by Phase

Loading 2609.34655v2…