On this page
TL;DR:
- For a manager, AI literacy means knowing enough to judge work that AI helped to produce. Coding and technical theory can wait.
- Five things matter most: how AI goes wrong, how to review AI-assisted work, what data must stay out, where a person must decide, and what rules your team works by.
- Match your checking to the risk. An internal note needs a quick read. A client figure, a contract or anything about a person needs every fact traced to its source.
- Make honest use safe. In a study covering 47 countries, over half of employees using AI at work said they avoid revealing it.
- Your name stays on your team's work, whoever or whatever drafted it.
Who this guide is for
This guide is for team leads, supervisors and heads of department in Kenyan organisations. You may never open an AI tool yourself, but your team probably does, and their work lands on your desk.
It covers what you personally need to understand. If your problem is managers who attend training and then never use AI, read how to get middle managers to use AI. If you want training for a group of managers, see our page on AI training for managers in Kenya.
What does AI literacy mean for a manager?
It means having the knowledge to make sound decisions about AI-assisted work: when to trust it, when to check it, and when to keep AI away from a task.
The term has reached the law in Europe. Article 4 of the EU AI Act, as changed by the Digital Omnibus on AI in force since 27 July 2026, requires providers and organisations deploying AI to take measures to support AI literacy among their staff, taking account of their knowledge, the context of use and the people affected. That law is aimed at the EU market, so check with a lawyer if you have EU links. The idea is useful anywhere, because it ties literacy to the job and to the people who could be harmed.
For a manager, that comes down to six things.
| What you need to understand | Why it matters to you | A quick test |
|---|---|---|
| How generative AI produces text | You stop treating fluent writing as proof | Can you explain why a tool might invent a report? |
| Where it tends to fail | You know which work needs closer checking | Can you name two tasks in your team where it is weak? |
| How to review AI-assisted work | Errors stop at your desk | Do you ask staff what they checked? |
| What data must stay out | You protect customers and staff | Could your team recite the "never paste" list? |
| Where a person must decide | You stay within the law and stay fair | Who decides on hiring, appraisal and discipline? |
| What your team's rules are | People know what is allowed | Is it written down anywhere? |
You can safely skip how neural networks work, the names of the latest models and every new feature. Those change often, and you can look them up when you need them.
How does AI actually get things wrong?
Generative AI writes what is likely to come next, so a false statement can arrive in the same confident voice as a true one. The US National Institute of Standards and Technology calls this confabulation: "confidently stated but erroneous or false content".
Research gives managers several reasons to take this seriously. Every study here has limits, so treat them as signals.
| Study | What it found | Why it matters to you | Limits |
|---|---|---|---|
| Dell'Acqua and others (2023), 758 consultants at Boston Consulting Group | On tasks AI handled well, consultants finished 12.2% more tasks, 25.1% faster, with over 40% higher quality. On a task beyond AI's ability, they were 19 percentage points less likely to be correct | The same tool helped on one kind of task and hurt on another. Different consultants did each task. The researchers call the edge of AI's ability a "jagged frontier" | One firm, an experiment, an older model |
| University of Melbourne and KPMG (2025), over 48,000 people in 47 countries | Two in three employees using AI at work said they have relied on its output without evaluating it. Over half said they have made mistakes because of AI. Over half said they avoid revealing when they use it | Unchecked use is common, and much of it is hidden from you | Self-reported, collected from late 2024 to early 2025. Kenya was not one of the countries |
| Lee and others (2025), 319 knowledge workers | People who trusted AI more reported less critical thinking. People more confident in their own skill reported more. Effort moved towards checking and combining AI's output | Staff who feel confident in a subject tend to think harder about the output. Staff who trust the tool more may think less | A self-reported survey. Most authors work at Microsoft, which sells AI |
| BetterUp Labs and Stanford (September 2025), over 1,000 US workers | 40% said they received "workslop" in the last month: polished AI content with little substance. Each case took nearly two hours to deal with | Weak AI work moves the cost from the sender to whoever receives it. Often that is you | US only, self-reported, co-run by a coaching company. Published sample sizes differ (1,004 and 1,150) |
Two real cases show what this looks like when it reaches the outside world.
In 2025, Deloitte agreed to repay part of an AU$440,000 contract with an Australian government department. The Register reported that the report contained fabricated citations and a made-up quote from a court judgment. A university researcher found them. The corrected version disclosed that a generative AI tool had been used.
In Kenya, Nairobi Law Monthly and Radio Generation reported on 28 May 2026 that judges had struck out pleadings after finding AI-generated content and citations to cases that do not exist. Neither report named the cases. The same reports covered the Judiciary's draft AI policy and practice directions, circulated for comment. The draft directions say "wholesale adoption of AI-generated text without independent review" would be treated as a failure of professional judgement. As of November 2026 we found no report that they have been finalised. They are written for the courts and the lawyers who appear before them. A manager signing off a report faces the same question.
Why can someone check AI's work and still miss the errors?
Polished writing lowers our guard. Reading for sense and reading for accuracy are different jobs, and a tidy paragraph can pass the first without ever facing the second.
A stronger tool may also make people relax. In a 2022 working paper by Fabrizio Dell'Acqua, as described in a review by Jason Collins, 181 professional recruiters assessed 44 job applications. Those given a better AI spent less time and effort and followed it more closely. Collins, who has a PhD in economics, points out that overall accuracy was similar across the groups, so the finding is about effort and needs care. The warning is still useful: as your team's tools improve, their checking may weaken.
Three habits help:
- Read twice for important work. Read once for sense, then again only for facts and numbers.
- Decide what you are checking before you start. "Does this look fine?" catches very little.
- Ask where a claim came from before you judge whether it sounds right.
How should a manager review AI-assisted work?
Set the amount of checking by what happens if the work is wrong. Then ask the person who produced it what they checked.
The levels below are our suggestion, so adjust them to your work.
| Level | Examples | What review looks like |
|---|---|---|
| Light | Meeting agenda, brainstorm, your own notes | The author reads it. You skim it if it reaches you |
| Standard | Internal report, staff notice, routine customer reply | The author checks facts and numbers. You sample some of them |
| Full | Anything going to clients, regulators, the board or the public. Figures, contracts and quotes | Every fact, number, name and reference is traced to its source. A named person signs off |
| Human decision | Hiring, appraisal, discipline, pay, credit, complaints about a person | AI may help organise information. A person decides and can explain why |
Six questions for any AI-assisted work
- Where did this come from? Ask for the document or link behind each claim.
- Can I trace the numbers? Recalculate at least the headline figure yourself.
- What has been left out? A summary can drop the objection that mattered most.
- Is this about us? If a paragraph could sit in any company's report, it may be saying nothing.
- Is it current? Prices, laws, names and job titles go out of date.
- Who owns it? The author should be able to defend it without looking at the screen.
Three things to say to the author
- "What did you check, and how?"
- "Which part are you least sure of?"
- "If a client questions paragraph three, what will you tell them?"
These questions work whether or not AI was used, so nobody has to feel accused.
Can you spot the problems? A worked example
Everything in this example is invented for practice. Imagine you manage a small sales team and have this data in front of you:
- Second quarter sales: KES 10.8 million. Third quarter sales: KES 12.4 million.
- Kisumu sales rose. Eldoret sales fell.
- Nobody gave the tool any data on complaints.
A staff member sends you this paragraph, drafted with AI:
Sales for the third quarter rose 18% to KES 12.4 million, driven by strong performance in Kisumu and Eldoret. According to the Kenya Retail Trends Report 2026, the sector grew by 9%, so we grew at double the market rate. Customer complaints also fell. We recommend opening a second Kisumu branch by January.
It reads well. Try to find the problems before you look at the table.
| What to look at | The problem | How you catch it |
|---|---|---|
| "18%" | From 10.8 to 12.4 million is about 14.8% | Recalculate the headline number |
| "Kisumu and Eldoret" | Eldoret fell | Compare each claim with the data |
| "Kenya Retail Trends Report 2026" | The name sounds real, and nobody has a copy | Ask for the link. If it cannot be found, cut it |
| "double the market rate" | It rests on the wrong figure and the unverified report | When one number falls, check what was built on it |
| "complaints also fell" | The tool had no complaints data | Ask what information the tool was given |
| The recommendation | It does not follow from the numbers. A branch needs costs and demand | Keep decisions like this with people |
Try this with your team. Take a real AI draft with sensitive details removed, or plant a few errors in one, and see who finds what. Then talk about how each person caught them.
What errors should you look for most?
Look hardest at anything specific: names, numbers, dates, references and quotes. These are easy to invent and easy to verify, so a short check pays off.
| Error type | What it looks like | What to do |
|---|---|---|
| Invented sources | A plausible report, case or article nobody can find | Open every reference yourself |
| Wrong arithmetic | Totals or percentages that do not match the sheet | Recalculate in Excel or on a calculator |
| Out-of-date facts | An old price, an old law or a person who has left | Check the current official source |
| Lost nuance | A summary that keeps the conclusion and drops the condition | Read the original where it matters |
| Filler | Smooth sentences with no fact, name or decision in them | Ask, "What does this actually tell the reader?" |
| Weak translation | Kiswahili that is correct but sounds wrong for your reader | Have a fluent speaker read it before it goes out |
Three prompts to help you check a draft
AI can be a useful second reader, though it can be wrong about its own mistakes too. Use these to find what to check, and then do the checking yourself. Remove names and personal details before you paste anything.
1. List what needs verifying
I am a manager reviewing the text below. Context: it was drafted with AI help and will go to [reader]. List every factual claim, number, name, date and reference in a table. For each, say what kind of source would confirm it. Constraints: do not tell me whether a claim is true, and do not add new facts. Customisation: put the claims that would cause the most harm if wrong at the top. Text: [paste]
2. Compare a summary with its source
Below are a source document and a summary of it. Context: I need to know whether I can rely on the summary. Constraints: list anything in the summary that the source does not support, quoting the source where you can. Then list the most important points or conditions the summary leaves out. If you are unsure, write [CHECK]. Customisation: use short bullet points. Source: [paste] Summary: [paste]
3. Argue against the draft
Read this recommendation as a sceptical finance director would. Context: our team proposes [decision]. Constraints: give the three strongest objections, and name the information that is missing. Do not rewrite the text. Customisation: plain English, under 200 words. Text: [paste]
What rules should your team work by?
A short team agreement is enough, and it should sit under your organisation's AI policy. If you have no policy yet, our guide to an AI usage policy for Kenyan organisations gives you a starting point.
Here is a template you can adapt in one team meeting:
How we use AI in this team 1. We use these tools: [approved list, using work accounts]. 2. We never paste in: customer or staff personal details, payroll, contracts, or anything marked confidential. 3. The person who sends the work owns the work, however it was drafted. 4. We check according to the review levels: light, standard, full, human decision. 5. We say when AI did a substantial part of a piece of work. Nobody is criticised for saying so. 6. Decisions about people are made by people. 7. If something goes wrong, we say so quickly. Fast honesty is never punished. 8. We review these rules every three months.
Why rule 5 needs your protection
Staff have reasons to hide AI use. Across four experiments with 4,439 participants, Reif, Larrick and Soll found that people who use AI at work expect to be judged as less competent and less motivated, and that they are. So your reaction to the first person who tells you matters a great deal. If you sneer, the rest of the team are likely to stop telling you.
A ban is unlikely to solve this. In the Melbourne and KPMG study, 67% of employees whose organisation had banned generative AI reported behaviour such as uploading sensitive company information or copyrighted material to public AI tools. The figure was 56% where a policy guided use, and 33% where there was no policy. That is a link and does not prove the ban caused it, but the ban did not stop those people. Written rules alone are not enough either: the report calls for clear guidance and education.
Junior staff need extra thought
AI can lift beginners quickly. Among 5,179 customer support agents, Brynjolfsson and colleagues found in a working paper revised in 2023 that issues resolved per hour rose 14% on average and 34% for novice and low-skilled agents, with minimal effect on experienced staff. In our view, that gain comes with a risk: a new person can get an acceptable answer without learning why it is acceptable. Ask junior staff to explain the reasoning behind their AI-assisted work, and have them do some tasks without it.
For the people side of this, our piece on soft skills and AI training goes further.
What are you still accountable for?
You remain accountable for your team's output and for how personal data is handled. As of November 2026, Kenya has no AI-specific law in force. The Artificial Intelligence Bill, 2026 is still a Bill: The Star reported in October 2026 that a Senate committee recommended not proceeding with it, and we found no report of a full Senate vote. Our explainer on the AI Bill has more.
The Data Protection Act already applies. Three points matter most for a manager.
- Automated decisions. Regulation 22 of the Data Protection (General) Regulations, 2021 says that where decisions are made by automated means with no human involvement, people must be told and must be able to get human intervention and give their point of view.
- Risky uses. Regulation 49 lists processing that needs a data protection impact assessment. The list includes automated decisions with significant effects, sensitive data, large-scale processing and "innovative use" of new technology.
- Staff data. As of November 2026, the regulator's AI guidance is still a draft. PPC Land reported on 3 August 2026 that its Annex 2 expects an impact assessment for recruitment screening and employee productivity monitoring.
So before anyone pastes CVs or appraisal notes into a tool, or uses AI to rank staff, stop and speak to HR or your data protection lead. Our guides on the Data Protection Act and AI and AI training for HR teams explain more. This section is general information. For a specific case, take legal advice.
A 30-day self-check for managers
You can build a working understanding in a month, using your own work.
| Week | Do this | You are done when |
|---|---|---|
| 1 | Use an approved AI tool on three of your own low-risk tasks. Check every fact in the output | You know how far you can trust it on those tasks |
| 2 | Run the worked example with your team, then try one of your own | Each person has caught an error |
| 3 | Agree your team rules and write them down | Everyone can find them |
| 4 | Review five pieces of AI-assisted work using the six questions | You know which errors your team makes most |
After that, keep a simple log of errors caught. It takes little time, and it shows you what your next team discussion should cover.
What will AI literacy not fix?
Better understanding will not fix a team that has no time to check its work. If deadlines leave no room for review, more knowledge will only make you aware of the risk you are carrying.
It also will not fix unclear goals, weak source data or a manager who does not read the team's work at all. And nobody becomes an expert in a few days. Expect your judgement to keep improving with practice.
Questions people ask
Do managers need to learn coding or prompt engineering?
No. You need to know how AI fails and how to review its work. Basic prompting is useful because it shows you what the tool can and cannot do, and you can learn it by practising on your own work.
How can I tell if a staff member used AI?
Often you cannot, and we would not build a process on guessing. Review the work on its quality and ask what was checked. Staff who feel safe saying so will usually tell you.
Should staff have to declare AI use?
We suggest yes, when AI did a substantial part of the work. Keep it simple and free of blame. Clients or regulators may have their own rules, so check those for outside work.
Can I use AI to help write performance reviews?
Be very careful. Appraisal notes are personal data, so they should stay out of public tools. You could use an approved tool to improve the wording of general text with names removed, but the judgement and the evidence must be yours.
Should I ban AI in my team?
We advise against a ban unless your work truly requires one. Survey evidence suggests rules get broken anyway, and hidden use is harder to manage. Clear rules and approved tools give you more control.
How much time should reviewing AI-assisted work take?
That depends on the risk level. Include checking time when you measure any saving. If review takes longer than writing from scratch, that task may not suit AI.
Is AI literacy the same as AI training?
They are different. Literacy is the understanding you carry. Training is one way to build it, along with practice on your own work and honest discussion with your team.
Sources
- National Institute of Standards and Technology, Generative AI Profile, NIST AI 600-1 (July 2024), last checked 10 November 2026
- EU AI Act, Article 4: AI literacy, last checked 10 November 2026
- EU AI Act guide, Digital Omnibus on AI, Regulation (EU) 2026/1744, last checked 10 November 2026
- Dell'Acqua and others, Navigating the Jagged Technological Frontier (working paper, September 2023), last checked 10 November 2026
- University of Melbourne and KPMG, Trust, attitudes and use of artificial intelligence: A global study 2025 (full report), last checked 10 November 2026
- KPMG, summary of the 2025 global study on trust in AI, last checked 10 November 2026
- Lee and others, The Impact of Generative AI on Critical Thinking (CHI 2025), last checked 10 November 2026
- BetterUp, "The hidden costs of workslop" (29 September 2025), last checked 10 November 2026
- Allwork.Space, report on the BetterUp Labs and Stanford workslop survey (September 2025), last checked 10 November 2026
- Niederhoffer and others, Harvard Business Review article on workslop (September 2025), last checked 10 November 2026
- The Register, "Deloitte refunds Aussie gov after AI fabrications slip into $440K welfare report" (6 October 2025), last checked 10 November 2026
- Nairobi Law Monthly, "Judiciary moves to regulate AI in courts over AI-generated filings" (28 May 2026), last checked 10 November 2026
- Radio Generation, report on the Judiciary's draft AI policy (28 May 2026), last checked 10 November 2026
- Jason Collins, review of Dell'Acqua's "Falling Asleep at the Wheel" recruiter study, last checked 10 November 2026
- InfoDocket, abstract of Reif, Larrick and Soll, "Evidence of a social evaluation penalty for using AI", PNAS (May 2025), last checked 10 November 2026
- Brynjolfsson, Li and Raymond, Generative AI at Work (NBER Working Paper 31161), last checked 10 November 2026
- The Star, "Senate committee shoots down Nyamu's Bill on AI" (15 October 2026), last checked 10 November 2026
- The Data Protection (General) Regulations, 2021, ODPC, regulations 22 and 49, last checked 10 November 2026
- PPC Land, report on the ODPC's draft AI guidance note (3 August 2026), last checked 10 November 2026
- Digital Policy Alert, status of the ODPC draft Guidance Note on Artificial Intelligence, last checked 10 November 2026
Where FuKazee fits in
FuKazee trains Kenyan managers and their teams to use AI on the work in front of them. Since 2022 we have trained more than 3,000 people. We assume nobody in the room knows AI yet, we avoid jargon, and we say plainly that nobody becomes an expert in a few days.
For managers, that means practice on real reports, appraisal wording with names removed, team rules and reviewing AI-assisted work. We show the free version of every feature and never promote one tool, and Microsoft 365 and Copilot are part of what we teach. We start with the four Cs: Clarity, Context, Constraints and Customisation.
See how it works on our page for AI training for managers in Kenya.
To talk through a session for your managers, email us at lead@fukazee.com
Want guided support instead?
Plan AI training around your team's real work. Answer a few quick questions about your sector, team size and goals, and we will recommend the right programme and format. Sessions are built on your workflows and include a usage policy.
Rodney has run FuKazee's training since 2022 and still teaches inside every programme line: corporate sessions, the postgraduate cohort and the KCSE examiner webinars. First Class Honours in Journalism and Mass Communication, University of Nairobi.
