Kenya’s County AI Pilots Need Decision-Recovery Drill

Imagine a parent seeking a county bursary for a student. An AI helpdesk tells them that an unstamped fee statement is sufficient. They follow the advice, only to discover that a required document has been mishandled. This example is hypothetical; the underlying paperwork is real. Makueni County’s publicly accessible 2024/25 bursary application form asks for family income, ward and village details, supporting evidence of family circumstances, and school-stamped fee information. A confident but inaccurate answer about any of those requirements could send an applicant in the wrong direction.

The practical question for county executives is whether staff could detect that error, find the controlling record, correct the advice, and explain the correction to the family. Too many AI pilots measure how quickly a system produces an answer. Public services need a second test: how reliably can an employee recover when that answer is wrong?

The same service already takes different forms across Kenya. Lamu County published downloadable bursary application forms for the 2026 academic year, while Makueni’s published application page shows an online form with fields for documents and supporting evidence. Neither example proves that an AI tool is being used in those offices. They show why a national standard for safe AI use must work with varying records, submission routes, and staff routines.

That Which Cannot Be Measured

That variation is measurable. A February 2026 Kenya Vision 2030 assessment of county automation found sharp differences in the accessibility of digital services. It described some counties as relying heavily on manual and cash-based processes and identified digital literacy, device access, implementation costs, and staff training as persistent obstacles. The assessment was a desk review, so it cannot tell us how every applicant experiences a particular service. It does establish that a single Nairobi-style digital workflow would be a poor starting assumption for all 47 counties.

Kenya has good reason to experiment. Its National AI Strategy for 2025–2030 includes digital infrastructure, data and AI governance, and innovation. But the county-level operating test should be simpler than a new procurement exercise. Before an AI assistant helps employees explain bursaries, licenses, health appointments, or land records to residents, the department should run a decision-recovery drill using invented cases and the actual published service rules.

Start with one service that already generates repeat questions. The departmental manager selects a current application form, any applicable instructions, and a dated paper or electronic copy of the authoritative rules. Staff should be able to identify which version governs a case, who can correct the record, and where the case goes when the evidence conflicts. This matters because Kenya’s ICT ministry has itself acknowledged fragmented public-sector data systems and limited data-management capacity. If a fee statement is in a physical file while an electronic record contains an outdated figure, a chatbot cannot decide which one controls through fluent prose.

Next, stage a short exercise with fictional applicants. Give one officer an AI-generated answer that invents a bursary eligibility rule; another receives a summary that drops a supporting document; a third sees a correct answer paired with the wrong academic-year form. Ask each officer to stop the answer before it reaches the resident, retrieve the authoritative source, make a correction, and record the reason. An officer who cannot locate the original should be able to pause the response and use the existing human service route. The ICT ministry explicitly includes critical evaluation of technology in its digital-skills framework. The drill makes that skill observable. Faster output counts for little if staff cannot identify an unsupported condition.

Mind Your Language 

Language needs its own test. Kenya’s Constitution recognises Kiswahili and English as official languages and protects linguistic diversity. In practice, an applicant may ask a question in Kiswahili or a local language while the source form uses English. A mistaken translation of “must attach” into “may attach,” or a muddled explanation of who qualifies, can change what the applicant does. Test an explanation in the languages that staff actually use with residents. Have a capable bilingual officer compare the meaning against the original instruction, and keep an approved glossary for recurring eligibility terms. A polished translation alone is no evidence of accuracy.

Connectivity creates another kind of failure. Kenya’s Last Mile County Connectivity Project documents a continuing effort to connect government facilities beyond county headquarters. A ward office should therefore rehearse what happens when the network drops midway through checking a case. Staff need access to dated, controlled offline instructions and a way to log an unresolved question for later reconciliation. They should never have to invent a rule because the portal is unavailable. The offline route must also make clear when information is too incomplete to give an answer.

Use What Already Exists

A workable drill needs modest resources. Kenya already has a Digital and AI Skilling Programme for Public Servants that includes county officials. Decision-recovery exercises would put those skills to work at the service counter. Use a handful of common cases, 30-45-minutes of staff time, and an existing supervisor. Record whether the employee caught the planted mistake before sending advice, how long it took to locate the source, and whether the correction was clear enough for the applicant to act on. Review missed cases and repeat the exercise after changing either the source material or the AI system. The county should publish basic service-quality measures, such as the share of tested errors caught and the proportion of unresolved inquiries that received a human follow-up, without exposing applicant identities.

Protect those identities during testing, too. Bursary applications can include family income, identity documents, and sensitive family circumstances. Kenya’s Office of the Data Protection Commissioner has warned counties about third-party handling of residents’ personal data. Use synthetic applications for the drill and approved systems for real records. Error recovery should leave a clear internal trail showing the source consulted, the advice corrected, and the official responsible for the final response.

An applicant also needs a workable correction channel. The Commission on Administrative Justice accepts complaints about delayed or undelivered public services. Counties should try to correct ordinary AI mistakes at the service counter before forcing residents into an escalation process. If a resident spots a mistake after receiving advice, staff should give them a case reference, a way to supply the missing document, and a clear explanation of who will respond. County managers should examine repeated correction requests for a broken instruction or stale form. Kenya’s 2026 national AI policy consultation underscores the need for responsible deployment and public trust. A correction route makes those ambitions visible at the service counter.

The strongest workplace AI adoption programmes give employees the means to challenge an automated recommendation and recover the right answer. Counties can adopt that principle now, without waiting for a national platform or an expensive new system. Pick a real service, plant a plausible mistake, and test whether the staff member can prevent it from becoming a resident’s problem.

This article was written by Gleb Tsipursky, PhD, a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results.