Making Generative AI Workable at Work: Accountability, Localization, and Representation in Kenyan and Nigerian Small Businesses
B · Useful, limited Microsoft Research Africa qualitative interview study with a clear method and a useful concept (release work), but a small purposive sample interviewed in 2023 on tools that have since changed.
Summary based on the full text·Original: about 75 minutes to read
In short
A CSCW paper (arXiv, 22 October 2026) by Najeeb Gambo Abdulhamid, Stephanie Nyairo, Liz Ankrah and Jacki O'Neill of Microsoft Research Africa. From 21 interviews with 27 workers in Kenyan and Nigerian small businesses, the authors find that drafting with generative AI was rarely the hard part. The work that decided whether gains materialised was making a draft defensible before it reached a client, which they call "release work", plus repairing outputs that did not fit local language, law or representation ("context misfit").
Key numbers
- 27workers interviewed, across 21 small and medium-sized businesses
- 18 / 9participants in Kenya and in Nigeria
- 60-90 minuteslength of each semi-structured interview
- July to October 2023fieldwork window, within the first year of public ChatGPT access
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Five practices made the tools usable: AI as draft labour, iterative steering, switching and comparing across tools, release work before circulation, and repair of context misfit.
- Release work has three parts: verification gates tied to domain standards, reconstruction of sources when citations fail or point to dead links, and non-delegation rules (for example, financial decisions and current law stay with the person).
- A law firm kept a librarian checking the Gazette and government printers, and told staff to treat AI output only as a starting point before going back to statutes and parliamentary reports.
- Context misfit showed up as "Americanised" legal answers, failure to pronounce or transcribe local names and accents, unstable answers on local cultural knowledge, and poor image and voice representation of African people and places. One participant abandoned a paid voice tool after a month.
- Several participants routed factual and statistical questions back to search and kept sensitive data (monthly sales) out of chatbots.
- Four design directions: provenance marked before circulation rather than rebuilt after, user-owned voice profiles, claim-level arbitration across tools, and grounding in local sources with honest early failure when a locale cannot be served.
How they got it
Semi-structured interviews in English, 60-90 minutes each, with workers recruited through professional networks who had used at least one generative AI tool at work. Thematic analysis with deviant-case checks. Participants held roles from interns to CEOs across legal, IT, marketing, finance, real estate, media and other sectors.
Limits
- Purposive sample of digitally connected early users; not a prevalence estimate, and the authors make no claims about differences between the two countries.
- Interviews date from July to October 2023, mostly with ChatGPT 3.5-era tools; the authors say the study is not an assessment of current models.
- English-only interviews may under-observe practice in local languages.
- Accounts are self-reported; no observation of actual outputs or measured time.
Why it matters for companies
The paper names a cost most AI business cases leave out: the checking and source reconstruction needed before anything goes to a client. That cost sits with the person who signs the output, so gains depend on how cheap verification is in a given domain. Firms working across markets should expect extra repair work wherever their language, law or customers sit outside the tools' defaults, and should test local fit before buying seats. Explicit non-delegation rules and a provenance step at the point of release are cheap controls to copy.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Work and AI