The detail behind
the story.

GradLaunch is a doctoral research project at Chester Business School examining whether structured, bounded use of generative AI can improve graduate employability outcomes. It runs in two phases: an exploratory qualitative phase, reported here, and a randomised controlled trial testing an AI-supported employability intervention against standard careers resources.

Phase 1 asks how final-year and final-stage undergraduates actually use generative AI in job-seeking, what they understand about how recruitment works, and what support they are missing. Its findings inform the design of the Phase 2 intervention.

Twenty-one semi-structured interviews were conducted in August 2026 with final-year and final-stage undergraduates across thirteen UK institutions.

Each interview followed a twenty-question topic guide covering current job-search activity, AI use across CV work, interview preparation and LinkedIn, beliefs about recruitment processes, how AI output is judged, concerns and perceived risks, and experience of university careers support. The topic-guide outline is published below.

The corpus comprises 23,687 words of dialogue, of which 15,605 words are participant speech.

Participants ranged in age from 21 to 38 and included full-time campus students, online and evening learners, and mature students studying alongside full-time work. Disciplines spanned business and management, engineering, nursing, computing, law, criminology, psychology, history, cybersecurity, health and care management, fashion marketing, and film and television production. Intended destinations included graduate schemes, NHS Band 5 nursing posts, care management, freelance and self-employed routes, and short-contract creative work.

Participants were recruited through public posts on LinkedIn announcing the research and inviting final-year and final-stage undergraduates to take part. In return for an interview, the researcher offered to answer any questions participants had about AI and employability. Participation was entirely self-selecting; no participant was approached directly and no institution circulated the invitation on the researcher's behalf.

Ethical approval was granted by the Chester Business School Research Ethics Committee on 29 May 2026.

All participants received a participant information document and gave written informed consent before taking part. The consent obtained covers publication of verbatim quotations in anonymised form, which is how every quotation on this site is presented.

Participants are identified by code throughout. Names used during the interviews have been replaced, and identifying references to employers, colleagues and specific workplaces have been removed or generalised.

In return for taking part, participants were offered answers to their own questions about AI and employability. This reciprocal arrangement is disclosed because it forms part of how the sample was constituted: those who responded were, by definition, people who wanted that conversation.

Phase 1 addresses four questions:

1. How do final-year and final-stage undergraduates use generative AI across employability tasks, and what governs where they use it heavily and where they hold back?
2. What understandings of graduate recruitment do they hold, where do those understandings come from, and how do those understandings shape their AI use?
3. On what basis do they judge whether AI-generated material and AI advice are any good?
4. What role does AI occupy relative to the human support available to them, and what does that reveal about the support they lack?

The analysis uses reflexive thematic analysis (Braun & Clarke, 2006; 2019). Transcripts were coded in NVivo against a codebook of twenty-five codes organised under four candidate themes, with coding revisited as the themes developed. Analytic memos were written throughout and form part of the audit trail.

Theme significance is not inferred from frequency. Reflexive thematic analysis does not treat the number of participants mentioning something as a measure of its importance, so no counts or percentages are reported. Where prevalence is described it is described in words, and qualifiers such as "most" or "all" are avoided unless the full corpus supports them.

Contrasting and disconfirming accounts are presented alongside supporting evidence in each theme rather than relegated to a limitations note. Where the data complicates an interpretation, that is shown.

The researcher conducting this study is also developing GradLaunch, the intervention Phase 2 will test. That interest can shape what appears significant in qualitative material, and it is stated here rather than left implicit.

Two safeguards are applied. First, accounts that complicate or contradict the emerging interpretation are actively sought and reported — the "where the story gets less tidy" passage in each theme exists for this purpose. Second, the design responses in this report are framed as hypotheses requiring testing, not as requirements the findings have established. Phase 1 can say what an intervention should respond to. It cannot establish that a particular intervention works.

Generative AI was used in this analysis and the use is disclosed in full.

AI assistance was used to produce a first coding pass over the corpus, to locate candidate extracts, to draft the codebook structure, and in preparing and editing this report. Theme development, the interpretation of the material, decisions about what the evidence supports, and all final judgements are the researcher's own. A first-pass coding frame produced with AI assistance is a starting point for analysis, not a substitute for it, and the coding was revisited independently.

No AI-assisted analysis features were used inside the qualitative analysis software. NVivo's AI Assistant is disabled at institutional level and was unavailable throughout; all coding within NVivo was carried out manually by the researcher.

No participant data was processed using tools outside those described in the approved ethics application.

Quotations reproduce the interview transcript wording. Selected excerpts may omit surrounding text; cuts within a sentence are marked. Each quotation carries a participant code and the question number it answers, so any extract can be traced back to its position in the corpus.

Headings, framing sentences and the notes beneath each quotation are editorial interpretation, not participant speech.

Diagrams are conceptual. Their arrows indicate an interpretive relationship or a possible sequence, not a measured effect. The overview illustration maps the sources of advice available to a job-seeking student; it does not represent how often each is used or how strongly each influences decisions.

The sample is not representative of UK final-year undergraduates. It leans toward mature students, students studying online or in the evening, students working alongside their degree, and widening-participation routes. This shaped what the study could see: the prominence of insider access and of practical constraints on getting help follows partly from who took part. A campus-based, full-time, traditional-age sample would likely produce a different emphasis.

Recruitment through LinkedIn shaped the sample twice over. Participants had to be active enough on the platform to encounter the invitation, and motivated enough by the offer of AI and employability guidance to respond. Both filters point in the same direction: toward students already engaged with AI and already thinking about the transition into work. Students who use neither AI nor LinkedIn are absent by construction.

This carries a further consequence that should be stated plainly. LinkedIn is not only the recruitment channel but also one of the information sources participants describe drawing on when forming their picture of how recruitment works. The study reached participants through a channel that is itself part of what the study examines. That does not invalidate the accounts, but it means the corpus is likely to over-represent students immersed in online careers discourse rather than students insulated from it.

These are accounts, not observations. Participants describe what they do and believe. Self-report is subject to recall, to rationalisation after the fact, and to what people are willing to tell an interviewer. Where a participant's stated practice appeared to shift under closer questioning, that is shown rather than smoothed over.

Beliefs about recruitment are not evidence about recruitment. Nothing here establishes how any employer actually screens applications. A participant describing automated ranking is evidence of a belief, not of a process.

No causal claims are made. The relationships described — between exposure and understanding, between available information and belief, between belief and behaviour — are interpretive. This design cannot establish that one produced another.

Preferences are not effects. Participants describing what they would want from a support tool tells us about felt needs. It does not tell us that such a tool would improve their outcomes. That is what Phase 2 is for.

All interviews were conducted by a single researcher. One interviewer means one interviewing style, one set of instincts about which answers to probe further, and one relationship with each participant. Follow-up questions were not standardised, so some accounts were pressed harder than others. Where a participant's stated practice shifted under closer questioning, that shift is visible in the quoted material rather than hidden.

The analysis was also carried out by that same researcher. In reflexive thematic analysis this is not treated as a deficiency to be corrected: Braun and Clarke are explicit that themes are generated through the researcher's engagement with the data rather than discovered independently of it, and that inter-rater reliability measures are conceptually inconsistent with the method. The relevant safeguard is not a second coder but transparency about position and interpretation, which is what the researcher-position statement above is for.

The analysis is in progress. Themes presented here are candidates. Continued engagement with the corpus may revise, merge or discard them.

Transcript verification remains incomplete. The website uses the earlier edited transcript files. Their wording has not been checked against original interview records.

The twenty topics covered by the guide:

1. Degree and career direction
2. Current job-search stage
3. Most recent AI use
4. Tools used
5. CV workflow
6. Interview preparation
7. LinkedIn
8. Recruitment mental model
9. What makes a CV stand out
10. Value of the degree
11. Estimated application volume
12. Judging AI advice
13. Rejected AI output
14. Concerns and risks
15. Employer detection
16. University careers support
17. AI alongside human support
18. Ideal support tool
19. Closing reflection
20. Participant question

Follow-up probes varied by participant and are visible in the quoted material.

Braun, V. & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101.

Braun, V. & Clarke, V. (2019). Reflecting on reflexive thematic analysis. Qualitative Research in Sport, Exercise and Health, 11(4), 589–597.

Institute of Student Employers. Student Recruitment Survey 2025.

Phase 1 Findings · Working research report · September 2026