On this page
TL;DR:
- Consent comes first: tell people before you record. In Kenya a voice recording can be personal data, and in a November 2025 decision the regulator ordered KES 700,000 in compensation over a recorded HR call where an automated notice was not enough.
- Get a transcript with timestamps, then give the AI a written rubric. Ask it for a quote and a time for every score, and allow it to answer "not enough evidence".
- Listen to the moments it cites. AI is useful for finding patterns in words and weak at judging tone, heavy accents, Kiswahili and Sheng.
- Turn each review into one change for next week, then check the next recording for that one change.
- As of November 2026, Teams and Zoom do not list Kiswahili for transcripts or captions, and Google Meet lists it only for live captions, in beta. Whisper-based tools accept it, and you must check the result by ear.
Who this guide is for
This guide is for anyone who improves by talking to people: trainers, teachers, sales teams, customer support teams and their supervisors.
You need no technical background, only a recording you are allowed to make and about 30 minutes a week.
The method comes from reviewing live training recordings against a written rubric. It is written here so that any team can use it. Legal points are general information, not legal advice. For a high-risk setup, ask a data protection adviser.
The method on one page
An AI call review works when a person sets the standard, the AI gathers evidence, and a person makes the judgement.
| Step | What you do | What you end up with |
|---|---|---|
| 1. Consent | Tell people what is recorded, why, who sees it and for how long | A notice and a consent log |
| 2. Capture | Record with a tool you are allowed to use | Audio or video, plus a transcript with times |
| 3. Clean | Remove names and check it is the right session | A safe transcript |
| 4. Score | Run the rubric and the prompt | Scores, each with a quote and a time |
| 5. Check | Listen to the cited moments | Scores you agree with |
| 6. Fix | Pick one change for next week | One written behaviour |
| 7. Track | Log the scores and whether the fix happened | A trend over several sessions |
Step 1: Get consent before you press record
Tell every person, before recording starts, what you are doing and why, and keep proof that you did.
What Kenyan law says
Kenya's Constitution protects privacy, including the right not to have "the privacy of their communications infringed" (Article 31). The Data Protection Act, 2019 turns that into working duties for organisations.
| Section | What it requires | What it means for a recording |
|---|---|---|
| 29 | Before collecting data, tell people, as far as practicable, about their rights, the purpose, who else receives it and the security in place | A line saying "this call is recorded" covers only part of this |
| 30 | You need a lawful basis, such as consent, performing a contract, or legitimate interests that do not unfairly harm the person | Choose your basis before you start and write it down |
| 32 | If you rely on consent, you carry the burden of proving it. People can withdraw at any time | Keep a log, and have a plan for someone who changes their mind |
| 33 | A child's data needs a parent's or guardian's consent, and must serve the child's best interests | Teachers: get this in writing before recording any lesson |
| 31 | High-risk processing needs a data protection impact assessment first | Routine monitoring of staff or customers may qualify. Take advice |
| 35 | People have a right not to face significant decisions made solely by automated processing, with limited exceptions | An AI score alone must never decide someone's job, pay or marks |
| 39 | Keep data only as long as reasonably necessary, then delete or anonymise it | Set a deletion date for each recording |
| 25 and 48 | Conditions apply to sending personal data outside Kenya | Find out where your tool stores and processes data |
Section 51 exempts processing by an individual in the course of a purely personal or household activity. A company reviewing its staff or customers is outside that.
The Data Protection (General) Regulations, 2021 add useful detail. Regulation 4 lists what to tell people when you rely on consent, including who you are, the purpose, any automated decision-making, any sharing, and the right to withdraw. It allows the information to be given in writing, orally, or by audio or video message. It also says consent is not freely given if it is presumed because someone did not object. So if consent is your basis, "nobody left the call" is weak proof.
Teachers should also read the ODPC's education sector guidance, which warns that pervasive student surveillance technology can cause serious harm.
The Act's definition of biometric data includes voice recognition. If a product builds voice profiles to identify speakers, ask the vendor about it and take advice before you use it.
Two ODPC decisions worth knowing
A recorded HR call: KES 700,000 awarded. On 3 November 2025 the Office of the Data Protection Commissioner (ODPC) ordered a Kenyan telecoms employer to pay a former employee KES 700,000 and issued an enforcement notice. Press reports say the case concerned a recorded consultation call with HR during a retrenchment. The employee said he had not consented and was told the recording would be deleted. The recording was later shared with a sister company for use in separate legal proceedings. The ODPC found that an automated recording notice at the start of the call did not show that all the section 29 duties were met, and it also found breaches of the rights to be informed and to erasure.
A recorded call in a bank's investigation: no breach. Here the ODPC found no breach over a recorded phone call used in an internal investigation. The employment contract had disclosed the monitoring, the bank relied on legitimate interest, and a court order supported the processing.
We read press reports of the first decision and a case-law library entry for the second, not the official texts. Whether either has been appealed is unknown.
Our reading of the two together: tell people clearly and in advance, say who will receive the recording, and use it only for what you said.
Are there separate call-recording rules?
We could not find a Kenyan statute that sets out a general "one-party" or "all-party" consent rule for ordinary people on a call. Here is what we could confirm as of November 2026:
- Kenya Information and Communications Act, section 31 makes it an offence for a licensed telecommunication operator, outside the course of its business, to intercept a message or disclose its contents. It is aimed at operators.
- Computer Misuse and Cybercrimes Act, section 17 makes it an offence to intentionally intercept data without authorisation, with a fine of up to KES 10 million, up to five years in prison, or both. The Act's definition of interception includes recording non-public transmissions. We found no Kenyan ruling on how this applies to a person who is a participant in the call.
- The Data Protection Act applies to organisations whether or not an interception offence applies.
Whatever the law allows, secret recording has no place in coaching.
A notice you can copy
Change the brackets. Delete any promise you cannot keep.
For a training session or class:
Before we begin, I would like to record this session so that [organisation] can review how it ran and improve the next one. The recording and an automatic transcript will be seen only by [roles]. We will use [AI tool, from provider] to help analyse the transcript, and this may involve processing outside Kenya. We will delete the recording by [date]. Nobody's marks, pay or job depend on this review. You can say no now, or change your mind later, and you will not lose out. If you prefer not to appear, tell me privately and I will [alternative, such as a written summary of the session]. Questions can go to [name and contact]. Is everyone happy for me to start?
For a sales or support call:
This call is recorded so that we can check our service and train our team. An AI tool may help us review the transcript, and we keep it for [number] days. If you would rather not be recorded, tell me and I will stop. We can carry on either way.
Log the date, session, how notice was given and any objections. Our guide to an AI usage policy for Kenyan organisations covers what that policy should say, and public bodies should also read data protection and AI in Kenya's public sector.
An employee may not feel free to refuse a manager. If that is your situation, consent may be the wrong basis, and a clear policy resting on legitimate interests may fit better. Get advice on this.
Do not record counselling, medical conversations, disciplinary hearings or anything involving a safeguarding concern for coaching purposes.
Step 2: Capture the session and get a transcript
Use the tool you already pay for, and make sure the transcript has times on it. Features below were checked on official pages on 8 November 2026 and change often.
| Tool | Free | Paid or work plans | Transcript and language notes |
|---|---|---|---|
| Google Meet | Recording is not on Google's list for free personal accounts | Recording: Business Standard and Plus, Essentials, Enterprise editions, Education Plus (staff), Teaching and Learning Upgrade, Workspace Individual, and Google One with 2 TB or more. Transcripts: a narrower list, including Business Standard and Plus and Workspace Individual. "Take notes for me" needs an eligible Workspace edition or Google AI plan | Transcripts and notes support English, French, German, Italian, Japanese, Korean, Portuguese and Spanish. Notes handle one language per meeting, and mixed-language meetings are not supported. Live captions list Swahili in beta; the transcript list does not |
| Microsoft Teams | Personal accounts cannot start a recording without a Microsoft 365 Personal, Family or Premium subscription. Those recordings expire after 30 days | Microsoft lists Office 365 Enterprise E1, E3, E5 and F3 and Microsoft 365 Business plans, if IT has enabled recording. Intelligent meeting recap sits under Teams Premium, and Microsoft moved some Premium features into Teams Enterprise on 1 April 2026. For 1:1 calls, recap needs Teams Premium or a Microsoft Copilot licence | Transcripts show speaker names and time stamps, and organisers can download .docx or .vtt. The listed languages do not include Kiswahili |
| Zoom | Basic: 40-minute meetings, recording to your computer only. Computer recordings get no audio transcript. AI allowance includes 3 hosted meeting summaries a month | Pro: USD 14.16 per user per month billed annually, or USD 16.99 monthly, with 10 GB of cloud recording and audio transcripts through cloud recording | The automated caption list has Afrikaans, Somali and Zulu, but no Kiswahili |
| Whisper (open source) | Free. MIT licence for code and model weights. Runs on your own machine. Models need about 1 GB (tiny) to about 10 GB (large) of graphics memory | Hosted versions cost money and send audio to the provider | Swahili is in its language list |
| Buzz (free app built on Whisper) | Free, MIT licence, works offline. Exports TXT, SRT and VTT. Current Mac versions need Apple silicon, and Windows shows a warning because the app is unsigned | None needed | Same strengths and limits as Whisper |
For phone calls, Google's Phone app can record on supported devices. Both parties hear a disclosure, and the file stays on your phone. Google says the feature is not available in all countries and does not list them, so check your own handset. Apple's feature page lists the languages for iPhone call recording, and Kiswahili is not among them.
Which route to pick:
- No budget, and your sessions are under 40 minutes. Record to your computer on Zoom Basic, then transcribe the audio with Buzz. Nothing needs to leave your laptop for the transcription step, which helps with privacy.
- You already have a work Microsoft 365 plan. Use Teams, because the transcript comes with speakers and times.
- You already have a qualifying Google Workspace plan. Use Meet. Google's transcript help page does not say how often timestamps appear, so look at one file before you build a routine around it.
- Sessions in Kiswahili or mixed language. Use Whisper or Buzz, and plan to correct the transcript by hand.
Two practical points for Kenyan teams. Microsoft's planning figure for a one-hour Teams recording is 400 MB, so consider staff data bundles before asking people to stream it. And noise hurts transcripts, so test a simple external microphone in the room you actually use.
Both Google and Microsoft now offer admins a consent prompt, off by default. Google's began rolling out on 5 May 2026 for eligible editions. Microsoft's keeps people muted until they answer and records the answer in the attendance report. Ask your IT admin to switch them on.
Step 3: Clean the transcript before any AI sees it
Replace names with labels such as Trainer, Learner A and Customer 1, and remove anything that identifies a person or client. Phone numbers, ID numbers, M-PESA codes and health details should go too.
Then check the file. Confirm the date and the opening lines match the session you mean to review. Notes from two similar sessions are easy to swap, and an AI will write a confident review of the wrong one.
Check the AI tool's data setting as well. As of November 2026, Google says Gemini chats may be used to improve its services and read by human reviewers when Keep Activity is on, and it warns against entering confidential information. Claude's consumer plans have a Model Improvement setting that controls training use. Whichever tool you choose, ChatGPT included, read the provider's current privacy page.
Work plans often carry different data terms, so read yours. Microsoft 365 Copilot Chat lets people with a qualifying work subscription upload files. If your material is sensitive, a model running on your own machine is an option. See our guide to running open-source AI models locally.
Step 4: Score against a rubric you wrote
A rubric makes the AI judge what you care about, and it makes scores comparable from week to week. Without one, you get general praise and generic tips.
Score only what can be seen in the words. Here is a free rubric. Copy it, then change it.
| Area | 1: Weak | 3: Mixed | 5: Strong | Evidence to look for |
|---|---|---|---|---|
| Purpose | No stated goal | Goal stated, but vague | Clear goal in the first few minutes, returned to at the end | The opening and the close |
| Accuracy | Wrong or risky claims stated as fact | Some claims unsupported | Claims are correct, or limits are stated | Numbers, rules, product claims, absolutes such as "always" and "never" |
| Balance of talk | One person talks for long stretches | Some questions, mostly closed | Others speak often, at length, and in their own words | Length of turns and longest monologue |
| Practice or action | Nobody tries anything | One short attempt | People do the thing, get feedback and try again | Moments when the learner or customer acts |
| Checking understanding | "Any questions?" only | Some specific checks | Specific questions, and the answers change what happens next | The question and what followed |
| Plain language | Jargon left unexplained | Some terms explained | Terms explained using examples from the listener's world | First use of each technical term |
| Time and next steps | Runs over, no next step | Near time, next step vague | On time, with who does what by when | Timing and the last few minutes |
Change these two rows to suit your work:
| Role | "Balance of talk" looks like | "Practice or action" looks like |
|---|---|---|
| Trainer | Learners explain and ask | Learners attempt the task on their own material |
| Teacher | Students reason aloud, and the teacher builds on their answers | Students solve, and the teacher responds to the working |
| Sales | The buyer talks about their problem before any pitch | The buyer agrees a specific next step |
| Support | The customer finishes describing the problem without interruption | The customer confirms the fix worked |
The prompt
Paste this into your AI tool, with the rubric and your cleaned transcript below it.
You are helping me review a recorded [training session / lesson / sales call / support call]. Context: [topic, audience, length, what success looks like]. The transcript below was produced by automatic speech recognition and may contain errors, especially with names, accents, Kiswahili and Sheng.
Score the session using only the rubric below. Follow these rules.
- For each area, give a score from 1 to 5, one exact quote, and its timestamp. Do not paraphrase the quote.
- If the transcript does not contain enough evidence, write "not enough evidence" and give no score.
- Judge only the words. Do not judge tone, confidence, mood or emotion, and do not guess anyone's identity, gender, age or background.
- If a passage looks garbled or mistranscribed, mark it [CHECK] and do not score from it.
- List every factual claim that needs verifying, with its timestamp. Do not decide whether a claim is true unless you can show why.
- Count these: the longest uninterrupted turn by the lead speaker, the number of times others tried something, and the number of questions from others.
Then give me: (a) a table of area, score, quote and timestamp; (b) three things to keep, each with evidence; (c) three things to fix, each with evidence and a specific behaviour to try next time; (d) the five moments I should listen to in order to check your review, in order of importance. Do not invent any quote or timestamp. If you cannot find one, say so.
RUBRIC: [paste rubric]
TRANSCRIPT: [paste cleaned transcript]
Here is an invented example of one good output row, to show the format:
| Area | Score | Quote | Time |
|---|---|---|---|
| Checking understanding | 2 | "Is that clear? Okay, let us move on." | 00:41:12 |
The score may be right or wrong. The value is that you can go to 00:41:12 and find out.
Step 5: Check the AI's judgement against the recording
Listen to the cited moments before you act on any score. In the published research we found, AI scoring of teaching was much weaker than its advice sounded.
Wang and Demszky (2023) tested ChatGPT on real classroom transcripts. Its scores correlated weakly with those of human raters: the best result was 0.35 on one teaching measure and 0.17 on another. Its feedback was often relevant, but 82% of its suggestions in one test pointed at moments where the teacher was already doing the suggested thing. That study used an older model (gpt-3.5-turbo), so treat it as a warning and test your own setup.
Narrow, specific feedback has better evidence behind it. In a randomised trial with 1,136 instructors in an online course, an automated tool that gave feedback on one behaviour (taking up students' ideas) raised that behaviour by 13%, with only suggestive evidence of benefits for students. Our reading: aim the AI at one observable behaviour and be wary of a single overall score.
A simple checking routine should take about 15 minutes:
- Go to the five moments the AI listed and listen, playing 30 seconds either side.
- Mark each one: quote exact, quote wrong, or quote right but judgement wrong.
- Compare every quote against the recording. Any quote you cannot find means the review is unreliable. Run it again.
- For the first three sessions, score the rubric yourself before reading the AI's scores, then compare.
If agreement stays low in one area, stop asking the AI about that area.
Before sharing a recording with anyone else, watch for what appeared on screen. Google says a Meet recording captures the active speaker and anything presented, so anything on a shared screen, such as an open folder or a client's chat, will be in the file.
Where AI judgement is unreliable
Be most careful with tone, accents, and any speech that is not standard English.
| Weak spot | What we know | What to do |
|---|---|---|
| African-accented English | The AfriSpeech-MultiBench study found commercial systems accurate on clean speech but varying significantly by country and domain, and open-source systems weaker on noisy, non-native dialogue. Whisper's own model card warns of uneven performance across accents and dialects | Correct key passages by hand. Do not score someone on words the machine misheard |
| Kiswahili | Not on the Meet transcript, Teams or Zoom lists above. Meet lists it for live captions only, in beta. Whisper accepts it. Its model card reports lower accuracy on languages with less data | Use a fluent reviewer. Treat AI output as a rough guide |
| Sheng and code-switching | We found no peer-reviewed error rate for Sheng. The AfriSwitch benchmark (a 2026 preprint) tested five systems on mixed-language African speech: the best averaged a 35.93% word error rate, and none scored below 24% on any language. Its abstract does not say whether Sheng was included | Assume a mixed-language transcript is unreliable until you have checked it. A speaker switching languages should never lower a score |
| Tone, warmth and confidence | A transcript does not record them. A major 2019 review found facial movements are not reliable signs of emotion across people and cultures. The EU's AI Act treats detecting emotion at work and in education as prohibited, apart from medical or safety uses. That law does not bind Kenya, and it shows how doubtful the practice is | Have a person judge tone. Avoid products that promise "sentiment" or "engagement" scores for individual staff |
| Silences | One study found about 1% of Whisper transcriptions contained whole invented phrases, more often for speakers with long pauses. Cornell reported that the rate has fallen since | Check any shocking or out-of-place line against the audio |
| Names and technical terms | The same MultiBench study found multimodal systems struggle with domain-specific named entities | Give the AI a list of correct names and terms, using labels where needed |
| Politeness and culture | Our view: indirect phrasing and deference can read as weakness or agreement in a transcript | Have someone who knows the audience read the fixes |
Step 6: Turn findings into next week's fixes
Choose one fix and write it as a behaviour someone could see on the next recording.
| Vague finding | Usable fix |
|---|---|
| "Talk less" | No explanation runs longer than 10 minutes without someone else doing something |
| "More practice" | By minute 20, every learner has tried the task once on their own example |
| "Be more accurate" | Before the session, check the five claims I use most against their sources |
| "Listen to customers more" | Ask two questions about the problem before naming any product |
Write the fix into the thing you use to run the session: the deck, the lesson plan or the call script. Keep what is working too, and name it, because people drop good habits they do not know they have.
If a review finds a wrong claim, correct it with the people who heard it, in the next session or in writing.
Step 7: Track improvement over time
Log each review in one sheet, and look for patterns across several sessions.
Use these columns: date, session type, rubric version, the seven scores, longest monologue, number of practice moments, minutes over time, wrong claims found, this week's fix, whether the fix happened (yes or no), and how far you agreed with the AI.
Some practical rules:
- Prefer counts over scores. Practice moments and minutes over time are harder to argue with.
- Keep the rubric fixed for at least a month. If you change it, note the new version and do not compare across versions.
- Plan on six to eight sessions before you read a trend. That number is our judgement, and we know of no study that sets it.
- If the same problem shows up in three reviews, change the training design or the system around the person.
- Delete recordings on schedule and keep only the scores and the anonymised notes.
For managers, one rule protects the whole system. Agree in writing that these reviews are for coaching. In our view, staff who suspect that AI scores feed into appraisals will perform for the rubric and stop trying new things. If a review ever raises a serious concern, a person must listen to the original, speak to the employee and decide. For more on what makes training last, see why corporate AI training doesn't stick.
What goes wrong, and how to avoid it
Most failures come from skipping a step.
| Mistake | Fix |
|---|---|
| Recording first, explaining later | Give the notice before the record button, every time |
| Using the free tool on a client's confidential call | Clean the transcript, check the tool's settings, or stay local |
| Scoring people on bad transcripts | Fix the audio and correct key passages first |
| Comparing colleagues on a leaderboard | Compare each person with their own earlier sessions |
Questions people ask
Is it legal to record a phone call in Kenya?
We could not find a Kenyan statute that sets a general one-party or all-party rule for people on a call. The Constitution protects the privacy of communications, the Computer Misuse and Cybercrimes Act has an interception offence, and organisations must follow the Data Protection Act. Tell the other person before recording, and take legal advice for anything sensitive.
Do I need consent to record a staff training session?
You need a lawful basis, and you must tell people the details listed in section 29 before you record. Consent is one basis, and legitimate interests may be another. Consent can be awkward between a manager and an employee, so a clear written policy and advance notice often matter as much.
Can I paste a transcript into free ChatGPT, Gemini or Claude?
Only after removing names and identifying details, and after checking the tool's data settings. Free consumer plans may use conversations to improve their models, depending on the settings. For client or staff material, prefer an approved work plan or a local model.
Which tool can transcribe Kiswahili?
Whisper and tools built on it, such as Buzz, accept Swahili. As of November 2026, Kiswahili is not on the language lists we checked for Google Meet transcripts, Teams transcription or Zoom automated captions. Expect to correct the output, and have a fluent reviewer check it.
Can we use AI call scores in performance reviews?
We advise against it. Section 35 of the Data Protection Act gives people a right not to face significant decisions based solely on automated processing, subject to limited exceptions. Transcription errors can also fall unfairly on people with particular accents. Use the scores for coaching.
How long should we keep recordings?
The Act says only as long as reasonably necessary for the purpose. For coaching, that usually means a short period. Our suggestion is 30 days, unless a law or a dispute requires longer. Keep the anonymised scores.
Sources
Law and regulator
- Constitution of Kenya, 2010, Article 31, Parliament of Kenya (last checked 8 November 2026)
- Data Protection Act, No. 24 of 2019, gazetted text hosted by ICNL (last checked 8 November 2026; now Cap. 411C, so check Kenya Law for later amendments)
- Data Protection (General) Regulations, 2021, ODPC (last checked 8 November 2026)
- Kenya Information and Communications Act, Communications Authority of Kenya (last checked 8 November 2026)
- Computer Misuse and Cybercrimes Act, Cap. 79C, Kenya Law (last checked 8 November 2026)
- ODPC, Guidance Note for the Education Sector (last checked 8 November 2026)
- allAfrica, news report on the ODPC's KES 700,000 decision (10 November 2025) (last checked 8 November 2026)
- Money254, report on the same decision (last checked 8 November 2026)
- CIPIT Case Law Library, ODPC determination on a bank's recorded call (last checked 8 November 2026)
- EU Artificial Intelligence Act, Regulation (EU) 2024/1689, Recital 44 (last checked 8 November 2026)
Tools
- Google, Record a video meeting (last checked 8 November 2026)
- Google, Use Transcripts with Google Meet (last checked 8 November 2026)
- Google, Take notes for me in Google Meet (last checked 8 November 2026)
- Google, Use live captions in Google Meet (last checked 8 November 2026)
- Google Workspace Updates, explicit consent for notes, recordings and transcripts (last checked 8 November 2026)
- Google, Use the Phone app to record calls (last checked 8 November 2026)
- Google, Gemini Apps Privacy Hub (last checked 8 November 2026)
- Apple, iOS feature availability (last checked 8 November 2026)
- Microsoft, Record a meeting in Teams (last checked 8 November 2026)
- Microsoft, Manage meeting and call recordings in Teams for personal use (last checked 8 November 2026)
- Microsoft, Start, stop and download live transcripts in Teams (last checked 8 November 2026)
- Microsoft Learn, Manage transcription and captions for Teams meetings (last checked 8 November 2026)
- Microsoft Learn, Teams recording storage and permissions (last checked 8 November 2026)
- Microsoft Learn, Manage participant agreement for recording and transcription (last checked 8 November 2026)
- Microsoft Learn, Teams Premium overview for admins (last checked 8 November 2026)
- Microsoft Support, How Copilot Chat works with and without a Microsoft 365 Copilot licence (last checked 8 November 2026)
- Zoom, Pricing (last checked 8 November 2026)
- Zoom, Enabling or disabling computer recording (last checked 8 November 2026)
- Zoom, Starting a computer recording (last checked 8 November 2026)
- Zoom, Enabling or disabling automated captions (last checked 8 November 2026)
- OpenAI, Whisper repository (last checked 8 November 2026)
- OpenAI, Whisper model card (last checked 8 November 2026)
- OpenAI, Whisper language list (last checked 8 November 2026)
- Buzz, GitHub repository (last checked 8 November 2026)
- Anthropic, Is my data used for model training? (last checked 8 November 2026)
Research
- Wang and Demszky (2023), "Is ChatGPT a Good Teacher Coach?" (last checked 8 November 2026)
- Demszky, Liu, Hill, Jurafsky and Piech, "Can Automated Feedback Improve Teachers' Uptake of Student Ideas?" (last checked 8 November 2026)
- Koenecke and others (2024), "Careless Whisper: Speech-to-Text Hallucination Harms" (last checked 8 November 2026)
- Cornell Chronicle, "AI speech-to-text can hallucinate violent language" (last checked 8 November 2026)
- Ashungafac and others (2025), "AfriSpeech-MultiBench" (last checked 8 November 2026)
- Ashungafac, Awobade and Olatunji (2026), "AfriSwitch" (last checked 8 November 2026)
- Barrett and others (2019), "Emotional Expressions Reconsidered" (last checked 8 November 2026)
Where FuKazee fits in
FuKazee trains teams and individuals in practical AI for the work in front of them, and we have trained more than 3,000 people. We assume nobody in the room knows AI yet, we avoid jargon, and we show the free version of each feature so that people can keep practising. We never promote one tool. Microsoft 365 and Copilot are part of what we teach, along with the four Cs (Clarity, Context, Constraints, Customisation), safe handling of data and how to check every output.
Nobody becomes an expert in a few days, and a call review habit needs weeks of steady practice. If your team of trainers, teachers, sales or support staff wants to set up a routine like this one, we can build the rubric and notice around your own calls and sessions. Read about our AI training in Kenya.
To talk it through, email us at lead@fukazee.com
Want guided support instead?
Not sure which programme fits? Tell us who you are and we will point you to the right path: corporate AI training, the ScholarShift Accelerator for postgraduates, or KCSE study plans and tuition.
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.
