Making Generative AI Workable at Work: Accountability, Localization, and Representation in Kenyan and Nigerian Small Businesses

1 Introduction

• The study treats effective use as producing locally credible, accountable outputs with acceptable time, verification effort, and risk—not access alone.Workable participation is an analytic lens for this study, not a general metric of democratization across workplaces.Source: §1 · pp. 1–2
• The study examines reported practices in Kenyan and Nigerian SMBs, where pragmatic experimentation and shared responsibility for client-facing quality make integration work visible.The sites are used to surface adoption, breakdown, and repair dynamics, not as representative cases of the Global South.Source: §1 · p. 2
• Five intertwined practices made GenAI workable: draft labour, interactional steering, tool switching and comparison, release work, and repair of context misfit.Availability alone did not make participation workable because outputs still had to be defensible for circulation.Source: §1 · p. 3
• The paper connects its practice account and design directions to a reframing of democratization as workable participation.It translates the findings into four design directions for workplace GenAI.Source: §1 · p. 3
• Release work is boundary-facing accountability labour that certifies an AI-assisted artifact as defensible before it reaches an external audience.It includes verification gates, traceability reconstruction when references fail, and non-delegation boundaries; it is distinct from articulation work.Source: §1 · p. 3
• Context misfit is a recurring breakdown-and-repair class involving local institutions and place, language practices, and representational fidelity.The paper documents repairs including voice rewriting, constraint layering, local sources and situated knowledge, tool switching, and disengagement when repair costs outweigh value.Source: §1 · pp. 2–3

2.1 Working with Generative AI

• Workplace GenAI can reduce first-draft effort while shifting work toward quality control, verification, traceability, and coordination across tools and sources.Prior work also highlights heterogeneous productivity effects and the role of organizational criteria in deciding whether outputs are reusable.Source: §2.1 · p. 4

2.2 Representation and Bias in Generative AI Models

• Representational and localization failures can be fluent but locally thin, and may involve place knowledge, culturally shaped depictions, language varieties, code-mixing, or accent handling.The paper uses context misfit narrowly for recurring localization and representation failures visible in interaction, rather than as an essentialized regional claim.Source: §2.2 · pp. 5–6

2.3 AI divides as workplace constraints

• AI divides are treated as practice conditions that shape tool selection and switching, selective disclosure, and verification burdens under time pressure and reputational risk.Relevant constraints include cost, privacy, infrastructure, and capability; layered tool ecologies can provide the practical substrate for combining and managing tools.Source: §2.3 · pp. 6–7

3.1 Research Sites and Scope

• Kenya and Nigeria were selected as situated, multilingual cases that make routine use and context-sensitive breakdowns visible, not as representatives of Africa or the Global South.The sites were also selected for practical access to local contexts and professional networks.Source: §3.1 · p. 7

3.2 Defining SMBs and Selecting Organizations

• An SMB was operationally defined as an organization with fewer than 250 employees, using a pragmatic threshold for comparability across sectors.Status combined self-reported size with participants’ descriptions of organizational roles and structure.Source: §3.2 · pp. 7–8

3.3 Participants and Organizations

• The study interviewed 27 participants from 21 SMBs across Kenya and Nigeria, spanning sectors and organizational positions; eligibility means the sample is not an adoption estimate.Most participants held client-facing roles, and recruitment required recent workplace GenAI use.Source: §3.3 · p. 8

3.4 Recruitment

• Recruitment through professional networks and snowball referrals required regular workplace GenAI use, making participants early-stage users during the 2023 fieldwork period.The study did not record exact length of use.Source: §3.4 · p. 8

3.5 Data Collection

• Interviews lasted 60–90 minutes and were conducted in English; demonstrations were supportive context rather than standardized observation, and modality examples are illustrative rather than prevalence estimates.English interviewing may under-observe practices occurring primarily in local languages or code-mixed interaction.Source: §3.5 · pp. 8–10

3.6 Transcription and Qualitative Analysis

• The team used codebook thematic analysis, combining inductive theme development with structured coding, iterative refinement, and consensus-based disagreement resolution.Analysts revisited transcripts and retrieved further examples to check that themes held beyond vivid excerpts; they did not use an inter-rater reliability coefficient.Source: §3.6 · p. 10
• Negative-case analysis and cross-role comparisons helped refine themes, but the findings are practice-oriented accounts rather than an exhaustive map of SMB use or evidence of cross-site differences.The study treated thematic density as sufficient when a second refinement round produced no substantively new themes or subthemes within the dataset.Source: §3.6 · p. 11

3.7 Ethics, Consent, and Compensation

• The study received ethics review approval, obtained informed consent, and encouraged participants to avoid sharing confidential client data.Participants could decline questions or stop interviews at any time.Source: §3.7 · p. 11

• Participants described five practices: drafting with AI, iteratively steering outputs, switching tools, performing release work, and repairing context misfit.Context misfit could require repair before circulation or lead participants to abandon the output.Source: §4 · p. 11
• Release work made drafts accountable before circulation, through verification, traceability reconstruction, and boundaries on what could be delegated.Participants’ analytic unit was circulation: assembling a safe-to-ship output package for client-facing release.Source: §4 · p. 11
• When outputs failed local fit, participants described repairing or substituting them, and sometimes abandoning them.Source: §4 · p. 11

4.1.1 Drafting content under delivery pressure.

• Under delivery pressure, participants used AI to unblock writing and generate starting points, then rewrote or localized drafts for professional standards and audiences.One consultant reported work that normally took two hours taking thirty minutes to “maximum forty minutes,” while still requiring revision and localization.Source: §4.1.1 · pp. 11–12

4.1.2 Producing structured artifacts, then tailoring for fit.

• Participants used AI to generate structured materials and candidate ideas, but selected and refined them to fit the project and audience.Examples included generating examination questions for later review and reaching candidate ideas in one session rather than several.Source: §4.1.2 · p. 12

4.1.3 Polishing communication to make it “safe to send”.

• For client communication, participants used AI to calibrate tone and polish language, then revised messages to preserve voice and avoid relationship risks.They used candidate phrasings to be firm without sounding rude, while checking outward-facing messages because mistakes could damage client trust.Source: §4.1.3 · pp. 12–13

4.1.4 Boundary case: using generative AI where expertise is thin.

• When expertise was thin, participants described AI as a scaffold for contributing, but recognized that limited expertise could make outputs harder to evaluate.Participants emphasized verification; one warned that outside one’s expertise, a person may be unable to review or check outputs for false information.Source: §4.1.4 · p. 13

4.2 Interaction routines for iteratively steering output quality

• Participants steered outputs through meta-prompting, external scaffolds, iterative revision, and editing, while reverting to manual methods or avoiding tasks when uncertainty remained.Source: §4.2 · pp. 13–14

4.2.1 Learning how to ask: meta-prompting as a practical skill.

• Some participants used AI to learn how to phrase prompts that would produce usable output, making prompt formulation part of the work.Source: §4.2.1 · p. 14

4.2.2 Scaffolding prompts with external tools and templates.

• Participants used plugins and other external scaffolds to shape outputs toward workplace-acceptable style and audience plausibility.Source: §4.2.2 · p. 14

4.2.3 Iterative refinement loops: prompting as editorial work.

• Participants repeatedly adjusted prompts and scrutinized, edited, and supplemented responses as part of an editorial workflow.Source: §4.2.3 · p. 14

4.2.4 Editing for originality, ownership, and voice as part of quality control.

• Participants rewrote outputs to maintain voice and ownership, including adding substantive input amid concerns about copyright and plagiarism.Source: §4.2.4 · p. 14

4.2.5 Boundary conditions: when steering is not enough.

• Steering could not ensure correctness or traceability, so participants avoided relying on AI for some facts and checked whether cited sources existed.Some returned to manual research when they did not trust an answer; participants also reported unusable or possibly fake references.Source: §4.2.5 · pp. 14–15

4.3 Orchestrating an ecology of tools through switching and comparison

• Participants treated tool choice as an ongoing practice, switching and comparing systems based on perceived capabilities, failures, and uncertainty.Source: §4.3 · p. 15

4.3.1 Switching to exploit differentiated affordances.

• Participants switched tools to match task needs, including richer visual material or more current information.Accounts described using Bing for multimedia or current information, and Google for recency when ChatGPT 3.5 was considered outdated.Source: §4.3.1 · p. 15

4.3.2 Parallel prompting and triangulation.

• Participants ran prompts across multiple tools, comparing outputs and using overlap or divergence to guide what to trust, revise, or verify.Source: §4.3.2 · p. 15

4.3.3 Arbitration when outputs diverge.

• When outputs diverged, participants had to reconcile, select, and adapt them for the task and audience; comparison did not remove human judgment.Plausible alternatives could leave users deciding which account to stand behind.Source: §4.3.3 · pp. 15–16

4.3.4 Boundaries, fallbacks, and paid tiers.

• Switching did not satisfy every accountability demand: participants routed factual research to search and weighed paid tiers against time, features, and access constraints.Source: §4.3.4 · p. 16

4.4.1 Verification triggers.

• Verification was triggered by domain standards and external accountability, including stale or incomplete information and the need to defend claims.Participants looked for auditable commitments to current sources and research references that could withstand challenge.Source: §4.4.1 · p. 16

4.4.2 Cross-checking routines.

• Cross-checking routed factual work to search and authoritative primary sources, sometimes through coordinated review by institutional roles.A legal services partner described a librarian checking legislation against the Gazette and government printers.Source: §4.4.2 · p. 17

4.4.3 Traceability breakdowns.

• When outputs lacked retrievable sources, participants reconstructed evidence trails so claims could be justified before external circulation.Verification therefore involved rebuilding a checkable justification, not only checking correctness.Source: §4.4.3 · p. 17

4.4.4 Non-delegation boundaries.

• Participants withheld high-stakes decisions from AI and required human review where responsibility or authoritative grounding mattered.AI could support analysis or drafting, while workers retained financial judgment and factual accountability.Source: §4.4.4 · p. 17

4.5 Context misfit and repair: localization work when outputs do not travel

• Participants repaired outputs that failed local requirements through rewriting, local constraints, source substitution, tool switching, or abandonment.Breakdowns included local words and accents, Westernized assumptions, and unconvincing representations of African people and places.Source: §4.5 · pp. 17–18

4.5.1 Sounding like me.

• Workers revised fluent drafts to preserve their voice and meet audience expectations, making professional representation socially safer.They edited overly formal or robotic language to sound like themselves and remain presentable.Source: §4.5.1 · p. 18

4.5.2 Language breakdowns.

• Local-language and accent failures could alter meaning or make a capability unusable, leading some participants to avoid code-mixing or switch to English.One participant abandoned a tool when it struggled to pronounce local names naturally.Source: §4.5.2 · p. 18

4.5.3 Locality gaps.

• Locality gaps included unstable cultural answers and legal outputs that missed locally decisive materials, so workers used generic text as scaffolding or rejected it.Participants described substituting situated knowledge when local specificity mattered.Source: §4.5.3 · pp. 18–19

4.5.4 Representation and repair.

• Generated images and voices could misrepresent African people, places, and accents; participants layered constraints and regenerated, but sometimes disengaged when repair costs persisted.One participant concluded a tool was not ready for their market after a month of paid use.Source: §4.5.4 · p. 19

5.1 Drafting and release work

• Perceived speed gains depended on whether workers could make AI-assisted drafts safe to ship through feasible release work.Release work gates circulation through verification, traceability reconstruction, and non-delegation boundaries; workers sometimes certified drafts alone, not only through coordinated work.Source: §5.1 · pp. 19–20
• The key measure is not drafting speed alone, but whether workers can assemble evidence and accountability boundaries before circulation.When sources are not retrievable, traceability reconstruction must fit deadlines and institutional standards.Source: §5.1 · p. 20

5.2 Steering as editorial practice

• Participants used iterative prompting, templates, refinement, and post-editing as an editorial workflow for producing drafts they were willing to release.Rewriting also preserved voice and ownership and managed plagiarism concerns, making prompt skill a situated quality practice.Source: §5.2 · p. 20
• Steering did not resolve factual uncertainty or unusable citations, so manual research remained a fallback and responsibility was redistributed rather than removed.These routines shaped drafts participants could credibly stand behind before circulation.Source: §5.2 · p. 20

5.3 Tool ecologies and arbitration

• Participants moved among tools to exploit different affordances, pursue freshness, obtain multimodal assets, and triangulate outputs.Agreement, divergence, and overlap informed what they trusted, revised, or verified further.Source: §5.3 · p. 21
• Cross-tool divergence required participants to arbitrate among alternatives, adding reconciliation overhead rather than eliminating judgment work.Evaluation of one tool alone can miss routing across tools, search, and paid tiers.Source: §5.3 · p. 21

5.4 Context misfit and repair

• Context misfit encompassed failures of voice, local languages and accents, domain realities, and credible representation, producing repair demands and credibility risks.Participants repaired outputs or abandoned capabilities when repair costs outweighed value; these dimensions are recurring, not exhaustive.Source: §5.4 · p. 21
• Locality failures were linked to uneven geographic knowledge production, but the paper does not claim specific datasets were tools’ training corpora.The authors trace how structural unevenness resurfaces in SMB work as breakdowns requiring repair and credibility management.Source: §5.4 · pp. 21–22
• Localization spanned speech and institutional knowledge: accent and pronunciation issues could alter meaning, while legal outputs missed locally decisive materials.Repairs included defaulting to English, substituting local sources, adding situated knowledge, switching tools, or abandoning capabilities.Source: §5.4 · p. 22

5.5 Workable participation

• Participants who were not AI specialists used generative tools to draft, refine, and package workplace materials for professional communication and participation.The paper relates this account to scholarship distinguishing access from equitable benefit.Source: §5.5 · p. 22
• Workable participation means producing locally credible, accountable outputs with acceptable time, verification effort, and risk.It was evidenced when release-facing accountability work could be completed within delivery timelines.Source: §5.5 · pp. 22–23
• Participants repaired misfit through substitution and layering, including primary-source checks, situated knowledge, local constraints, tool switching, or disengagement.Access expanded drafting possibilities, but participation depended on stabilizing verification and localization within acceptable time, effort, and risk.Source: §5.5 · p. 23

5.6 Contributions

• Release work is boundary-facing accountability labour that determines whether an AI-assisted draft can circulate to an external audience.It involves verification gates, traceability reconstruction when references fail, and non-delegation boundaries; perceived speed gains depended on completing it within deadlines and institutional standards.Source: §5.6 · p. 23
• Context misfit is a recurring breakdown-and-repair class whose localization and representational failures become consequential through repair or abandonment.The repair repertoire includes rewriting for voice, layering constraints, substituting local sources and situated knowledge, switching tools, and disengaging when repair costs outweigh value.Source: §5.6 · pp. 23–24
• The paper connects these contributions through workable participation and translates them into four design directions targeting release work, steering, ecological use, and localization.The framing focuses on locally credible, accountable outputs within acceptable time, verification effort, and risk.Source: §5.6 · p. 24

5.7 Design implications

• Design for the boundary crossing to a client, where consequential friction appeared, rather than for first-draft production.The proposed design principle centers the moment an output leaves the worker and enters external circulation.Source: §5.7 · p. 24
• Make provenance legible before circulation by marking claims as grounded or unverified and providing a portable evidence trail.The trail could include source, retrieval date, and fallback query; citation features should not create false confidence in unusable references.Source: §5.7 · p. 24
• Treat voice as a cumulative, user-owned asset with an inspectable, overridable profile and support for prompt formulation.The proposed design responds to learned steering and authorship work while guarding against homogenization and overfitting past text.Source: §5.7 · p. 24
• Support arbitration across tools by surfacing agreement and divergence at the level of specific claims.Agreement should guide where to look, not stand in for verification, because similar tools may share blind spots.Source: §5.7 · p. 24
• Enable local grounding and persistent, inspectable locale, voice, and representation constraints, while signaling early when requirements cannot be met.Workers could ground outputs in their own sources or local materials and disengage before repair costs accumulate; representation controls should use community-grounded defaults.Source: §5.7 · pp. 24–25

5.8 Scope and limitations

• The study uses Kenya and Nigeria for analytic leverage, not representativeness, and does not support prevalence estimates or claims about differences between the sites.English-language interviews may under-observe local-language practices; fieldwork from July to October 2023 is a snapshot, not an assessment of current tool performance.Source: §5.8 · p. 25

6 Conclusion

• Across accounts, five interlocking practices made generative AI workable: draft labour, steering, tool comparison, accountability work, and context-misfit repair.These practices addressed client-facing deliverables and mismatches in language, institutions, or representation.Source: §6 · p. 25
• Incorporation involved verification triggers, accountability boundaries, and cross-tool comparison; context misfit prompted repair or abandonment when costs exceeded value.The repair repertoire included voice rewriting, localization, substituting situated knowledge, and switching tools.Source: §6 · p. 25
• Reported gains were conditional on release work that made outputs defensible for circulation, and the findings describe practices in two sites rather than representative global patterns.The paper identifies four design directions: verification and traceability, steering and post-editing, tool comparison and arbitration, and localization including representational fidelity.Source: §6 · p. 25