An AI-generated draft is not the same thing as work a business can responsibly send to a client. A study submitted to arXiv on 22 October 2026 examines that distinction through early generative AI use in Kenyan and Nigerian small businesses. In the researchers’ account of participants’ experiences, producing text was rarely the hardest part; making it defensible for an external audience was the consequential work. The useful question this raises is not simply whether AI can produce something quickly, but what remains to be done before someone will stand behind it.
A new paper about earlier workplace use
The researchers report 21 semi-structured interviews involving 27 participants from 21 small and medium-sized businesses, conducted between July and October 2023. Those dates matter: the submission is new, but the experiences describe an earlier period of workplace AI use, not business practices or tool performance in October 2026.
The paper lists all four authors as affiliated with Microsoft Research Africa; its peer-review status is unconfirmed. The authors also explicitly treat Kenya and Nigeria as two situated cases rather than a representative sample of small businesses or the Global South.
This makes the research useful for examining how people describe integrating AI into their work, not for calculating a national productivity effect. Its relevance elsewhere is therefore a question to investigate: does an organisation’s assessment of AI include the effort required to turn an output into an acceptable deliverable?
The work required before release
The authors describe workers adjusting prompts, rewriting outputs, preserving their own voice and comparing or switching tools. Their account is not of a single instruction followed by a finished product, but of people steering output towards something they could use.
They distinguish a further stage, which they call “release work”: the checking and allocation of responsibility that determine whether AI-assisted material can circulate externally. This includes verification against professional standards, reconstructing the source trail when references fail, and deciding which responsibilities cannot be delegated.
The distinction is valuable analytically. Editing asks whether an output can be improved. Release work asks whether it is ready to leave the organisation, and who accepts responsibility for that decision. A polished sentence does not, by itself, answer either question.
Small-business settings make this integration work particularly visible, the researchers argue, because the people using the tools often also carry responsibility for client-facing quality, reputation and deadlines. Their rationale suggests that evaluation should follow responsibility through the workflow, rather than stop when the model finishes generating.
Localisation is more than a finishing touch
The study identifies three dimensions of mismatch between AI output and local context: knowledge of institutions and places, including regulations and market norms; language practices, including multilingual and mixed-language communication; and faithful representation of people and places in generated language and media.
Read this way, localisation is not merely a final adjustment to spelling or tone. The authors’ framework invites a more demanding question: is the material credible in the setting where it will be used? Fluency and local suitability should be assessed separately.
Reported responses included rewriting for voice, adding constraints, substituting local sources and knowledge, switching tools and abandoning use when repair costs outweighed value. These are descriptions of participants’ experiences, not a ranking of tools or a measurement of how frequently any repair was needed.
The option to stop using a tool is especially important to the interpretation. Disengagement need not be treated as a failure to adopt AI successfully. As an evaluation principle, it can instead mean recognising that a particular task offers too little value after the necessary repairs. The retained passages do not quantify that threshold.
What language benchmarks can add
Separate research provides limited context for the language dimension. IrokoBench, published in April 2025 conference proceedings, presents a human-translated benchmark covering 17 low-resource African languages. Its researchers report a significant performance gap between high-resource languages, such as English and French, and the low-resource African languages they evaluated.
That finding is relevant background, but it does not verify the business interviews, measure workplace correction time or establish why a particular participant needed to revise an output. A benchmark result and an account of workplace experience answer different questions.
The practical implication is to avoid treating strong performance in one language as sufficient evidence of suitability for another. For a business evaluating AI, language and local context deserve explicit attention rather than being assumed to follow from general fluency.
Why completion time needs its own evidence
A separate developer experiment illustrates the importance of measuring completed tasks. In a randomised trial involving 16 experienced open-source developers and 246 tasks, researchers found that allowing AI increased completion time by 19%. The trial examined February–June 2025 tools and developers working in mature projects they knew well; its estimate cannot be transferred to these businesses or to current systems.
Nor should that older result become a permanent verdict. In a February 2026 update, METR said conversations with participants suggested developers were probably benefiting more than in early 2025. However, selection effects left only very weak evidence about the size of the improvement. That follow-up came from the same research organisation, not an independent replication.
Neither developer result establishes that the mechanisms described in the Kenyan and Nigerian interviews caused a slowdown. The narrower lesson is methodological: task completion requires its own evidence. A finding tied to particular people, tools and dates should not become a universal claim that AI either saves or wastes time.
Count the work someone can stand behind
The main study’s authors argue that participants’ reported gains depended on accountability and repair labour. That is not a finding of net productivity loss—or a quantified gain. The qualitative evidence available here supplies no timed comparison, error-rate estimate or numerical measure of the verification burden.
Their definition of effective use offers a more useful starting point: producing locally credible, accountable outputs with acceptable time, verification effort and risk. An editorial implication is to evaluate AI against the whole deliverable, not merely its first draft.
For a small business planning a trial, that could mean asking how long an acceptable output takes, what corrections it needs, whether its sources can be checked, whether it suits its intended audience, and who signs it off. These are evaluation questions, not a proven recipe for productivity gains.
The unresolved question is concrete: after verification, correction and localisation, does AI help the business deliver work it can stand behind—or simply move effort to the person responsible for pressing send?