On this page
TL;DR:
- Roll out in four phases over about a year: prepare, pilot, expand in waves, then build AI into normal work. This timing is our suggestion, and a simpler company can move faster.
- Start from what staff already do. Many are probably using AI on personal accounts, so find that out before you buy anything.
- Split your uses into two queues. Work with no personal data can start early. If a use needs a data protection impact assessment, allow at least 60 days after you submit it.
- Train line managers before their teams, and keep skills groups to about 20 people.
- Pay for licences only where a role has shown the need. Measure real tasks, including the time spent checking AI's work.
Who this guide is for
This guide is for the person told to make AI work across a company of 100 or more people. You may be in HR, operations, IT, strategy or the CEO's office.
It starts after the decision to go ahead. If your CEO has not yet approved anything, read what a CEO should ask before an AI rollout first. If you only need to plan one training session, our guide on how Nairobi companies can train staff on AI is the better fit.
What changes once you pass 100 staff?
Size changes the problem from teaching people to coordinating people. A 20-person firm can agree its AI rules over lunch. A company of 200 has departments with different data, branches on different shifts, and managers who will each interpret the rules differently.
| Issue | Team of 20 | Company of 100 or more |
|---|---|---|
| Training | One group | Several groups, planned around the work calendar |
| Rules | A conversation and a short note | A written policy, a named owner and a way to ask questions |
| Data risk | One or two kinds of data | Staff, customer and financial data, each with different owners |
| Tools | Whatever the team agrees | Licensing, security settings and access rights |
| Staff worries | Handled face to face | Rumours travel faster than official messages |
| Proof that it works | The boss can see it | You need numbers from several departments |
What does the evidence say about scaling AI?
The research points the same way: using AI somewhere is common, spreading it through a company is hard, and results depend on the task. Each study below has limits, so treat them as signals.
| Source | What it found | Limits |
|---|---|---|
| McKinsey State of AI 2025, as reported by PPC Land | 88% of respondents said their organisations use AI in at least one function. Only about a third had begun scaling it. 39% reported any effect on profit (EBIT), and most of those said under 5% | A global survey of what people say about their own firms. 1,993 participants |
| Same survey, on top performers | About 6% of respondents qualified. They were nearly three times as likely to have fundamentally redesigned workflows | Shows a link, and does not prove the redesign caused the results |
| Central Bank of Kenya survey, data as of 31 December 2024 | 50% of institutions surveyed had adopted AI. 70% had no AI strategy. Among adopters, 41% had AI policies. Skills shortage was a top challenge at 46% | Banks, microfinance banks, digital credit providers and credit reference bureaus only |
| 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 | One firm, a controlled experiment, an older model |
| Brynjolfsson, Li and Raymond, 5,179 customer support agents | Issues resolved per hour rose about 14% on average and about 34% for novice and low-skilled workers. Experienced workers saw minimal change | One company and one type of work |
Two lessons matter for your plan. Test each task before you scale it, because AI helps with some and harms others. And plan to change how the work is done, because handing out logins alone is unlikely to show up in your numbers.
Why should you start with what staff already do?
Because many of your people are probably using AI already, and your plan should deal with that first. Microsoft and LinkedIn's 2024 survey of about 31,000 knowledge workers in 31 markets found that 78% of those using AI brought their own tools to work. Microsoft sells AI products, so read that with care.
Kenya's numbers are striking too. The Star, citing the Digital 2026 Mid-Year Global Update, reported that 97.5% of Kenyan internet users aged 16 and over use AI platforms each month. That covers internet users only, and it says nothing about use at work or how deep that use goes. It does tell you not to assume your staff are beginners who have never touched these tools.
Run a short anonymous survey in your first fortnight. Make it clear that nobody will be punished for honest answers. Ask which tools people use, for what, and with what kind of information. You will learn where the risk is, and you will find your future champions.
The four-phase plan at a glance
No study proves that this timing is right. It is the shape we suggest for a company of 100 to 500 staff, and each phase ends with a decision.
| Phase | When | Goal | You move on when |
|---|---|---|---|
| 1. Prepare | Weeks 1 to 6 | Owners, interim rules, a full picture of current use | Policy approved, data queues agreed, baseline recorded |
| 2. Pilot | Weeks 7 to 18 | Test two or three real tasks in two or three teams | The tasks show a measured benefit and errors are being caught |
| 3. Expand | Months 5 to 8 | Roll out in waves, managers first | Most target staff are using AI on real tasks each fortnight |
| 4. Embed | Months 9 to 12 | Build AI into procedures, onboarding and reviews | AI has a permanent owner and a budget line |
If a pilot fails its review, that is useful. Change the task or the tool and run it again before you spread it.
Phase 1: Prepare (weeks 1 to 6)
The purpose of this phase is to put owners, rules and facts in place before anyone is trained.
Name the people
Keep the working group small. It needs people who can make decisions, and it should meet every two weeks.
| Role | Usually held by | Responsible for |
|---|---|---|
| Executive sponsor | A member of the executive team | Removing blocks, protecting time, speaking to staff |
| Programme lead | HR, operations or strategy | The plan, the calendar and reporting |
| IT and security | Head of IT or a senior engineer | Approved tools, settings, access rights |
| Data protection | Your data protection officer or the person doing that job | Data queues, assessments, vendor terms |
| HR | HR manager | Staff messages, policy, job and role questions |
| Finance | A finance manager | Licence costs and tracking value |
| Department leads | One per major function | Choosing tasks and releasing staff |
| Staff voice | A union or staff representative, if you have one | Raising worries early |
Deciding whether to run this yourself or bring in outside help is a separate choice. Our page on in-house or external AI training sets out the trade-offs.
Publish an interim rule
Do this in the first two weeks. It can be a single page: which tools are approved, what must never be pasted in, that a person checks every output, and who to ask. Our guide to an AI usage policy for Kenyan organisations gives you a starting point.
Sort uses into two queues
This is the step that saves the most time. Ask each department for its most repetitive tasks, then sort them.
| Queue | Examples | When it can start |
|---|---|---|
| A: no personal data | Drafting from public information, tidying procedures, formula help, summarising published reports | As soon as the interim rule is out |
| B: personal or sensitive data, or decisions about people | Customer files, payroll, CV screening, staff monitoring, credit decisions | After review, and after a data protection impact assessment where required |
Queue B has a real timeline. Section 31 of the Data Protection Act requires a data protection impact assessment (DPIA) before processing that is likely to result in high risk. Section 31(5) says the report is submitted sixty days before processing. Under regulation 52(3) of the 2021 General Regulations, if you hear nothing within sixty days of submitting it, you may start.
Regulation 49 lists triggers, including automated decisions with significant effects, sensitive data, large-scale processing and "innovative use" of new technology. The regulator's AI guidance note, dated July 2026, is a draft. Digital Policy Alert records that consultation closed on 17 August 2026, and we found no final version as of 10 November 2026. We could not open the regulator's own copy, so we rely on PPC Land, which reports that Annex 2 of the draft lists recruitment screening and employee productivity monitoring as uses that need a DPIA. For how the Act applies to staff using AI, see our guide to the Data Protection Act and AI.
If you start the DPIA in week 2, it can run alongside your Queue A pilot, so nobody sits waiting.
Choose tools from what you already own
Check your current licences before you buy anything. As of November 2026:
- Microsoft 365. Microsoft says Copilot Chat is included at no additional cost with eligible plans. It also says that with enterprise data protection, prompts and responses are not used to train its foundation models. Use a work account, because those terms apply to organisational use.
- Google Workspace. Google's help page draws a line between two kinds of access. Where Gemini is an additional service, chats may be reviewed by humans and used to improve products. Where it is a core service, they are not. Ask your administrator which one you have.
We do not promote one tool. Choose the one that fits the systems your staff already use.
Record your baseline
Pick the tasks you plan to pilot and time them now, with the people who do them. Without a "before" figure, you will have only opinions at the end. Our guide on evaluating ROI on AI training explains how.
Tell staff about jobs
Say what you know and what you do not know. Make no promise you cannot keep. Silence will be filled with rumours. If you ever plan redundancies, section 40 of the Employment Act has its own rules on notice, selection and severance, and a rollout plan does not replace them. Take legal advice before that conversation.
Phase 2: Pilot (weeks 7 to 18)
A pilot should answer one question: does AI make this specific task better once checking time is counted?
Choose two or three teams and two or three tasks. Good pilot tasks are frequent, low risk, easy to check and owned by a willing manager. Use this scoring table with your department leads.
| Question | Score 0 | Score 1 | Score 2 |
|---|---|---|---|
| How often is it done? | Rarely | Monthly | Weekly or daily |
| How easy is the output to check? | Very hard | Possible with effort | Quick and clear |
| What happens if it is wrong? | Serious harm | Embarrassing, fixable | Minor |
| Does it use personal data? | Yes, sensitive | Yes, ordinary | No |
| Is the manager willing? | Opposed | Neutral | Keen |
Start with the highest scores. Treat a 0 on "what happens if it is wrong" as a veto, however high the total.
Run the pilot this way:
- Train the pilot teams on the tool and on your rules, using masked real work.
- Pick one or two champions in each team, and give them protected time.
- Hold a 30-minute check-in each week. Ask what worked, what failed and what was caught.
- Keep an error log. It costs little, and it becomes your best training material.
Suggested gate for moving on. These are starting points of ours, so adjust them. At least two tasks show a measured improvement after checking time is counted. No serious data incident has occurred. Managers say quality held or improved. Pilot staff would recommend the approach to colleagues.
Phase 3: Expand in waves (months 5 to 8)
Expand by wave so that each group learns from the last one. Managers go first, because they decide whether their teams get time to practise. Our piece on getting middle managers to use AI explains why, and we also run AI training for managers.
Keep skills groups to about 20 people, and group them by the type of work they do. Space sessions out, and block month-end for finance.
A made-up example: 180 staff
Suppose a company has 180 people across a head office, branches and a warehouse. This example is invented to show the arithmetic.
| Group | Staff | Plan | Groups |
|---|---|---|---|
| Executive team | 8 | One 90-minute session on direction and rules | 1 |
| Line managers | 22 | Two groups, before their teams start | 2 |
| Head office, by function | 80 | Four groups of about 20 | 4 |
| Branch staff | 30 | Live online, timed around shifts | 2 |
| Warehouse and drivers | 40 | A short session on safe use and practical help on a phone | 2 |
| Total | 180 | 11 groups |
Do not give everyone a paid seat on day one
Paid licences add up quickly. Microsoft's US price page, checked on 10 November 2026, lists Microsoft 365 Copilot Business at US$21 per user per month paid yearly. It also shows a first-year US$18 promotion for eligible existing Business customers until 31 December 2026. Some advanced features are billed by usage. Kenyan prices, tax and eligibility may differ, and larger firms may be on enterprise terms, so check with your Microsoft partner.
Using the US$21 figure for our 180-person example:
| Choice | Sum | Cost per year |
|---|---|---|
| A paid seat for everyone | 180 × 21 × 12 | US$45,360 |
| Paid seats for 40 people who have shown the need | 40 × 21 × 12 | US$10,080 |
A fair rule is that a person gets a paid seat when they can name the task it would help and why the free or included version falls short. Review usage after 90 days and move unused seats. For the wider costs, see how to budget for AI training.
Frontline and shift staff
Do not leave them out. Staff who never sit at a computer still hear about AI, and some will use it on their phones. They need three things: what they may and may not paste in, how to check an answer, and one or two uses that help them. Run these sessions on their shifts, and take care with Kiswahili outputs, which should be read by a fluent speaker before anything goes to customers.
Phase 4: Embed (months 9 to 12)
AI becomes part of the company when it appears in your normal documents and routines. If it still depends on the programme team, it will fade when they move on.
- Update standard operating procedures for each task that has changed.
- Add AI rules and basic skills to onboarding for new staff.
- Move ownership from the project team to a permanent role and a budget line.
- Refresh the policy with what the error log taught you.
- Report to the board or executive team each quarter.
- Keep a list of tools you stopped using, and why.
If your board wants a recognised structure, the US National Institute of Standards and Technology's AI Risk Management Framework is voluntary and is built around four functions: Govern, Map, Measure and Manage.
How should you communicate during the rollout?
Communicate early, and let leaders speak first. Staff need to hear the purpose from someone senior before they hear it from a colleague.
| When | Who speaks | What people need to hear |
|---|---|---|
| Before the survey | Executive sponsor | Why we are doing this, and that honest answers are safe |
| After the interim rule | HR and IT | What is allowed, what is not, and who to ask |
| Pilot starts | Department heads | Which teams, which tasks, and that it is a test |
| After the pilot | Sponsor and pilot staff | What worked, what failed, what happens next |
| Before each wave | Line manager | When, what to bring, and how your work will be covered |
| Each quarter | Sponsor | Results, including what did not work |
What should you measure?
Measure whether real work has changed. Login counts tell you very little. Use a small set of numbers and report them the same way each month.
| Area | Measure | Warning sign |
|---|---|---|
| Use | Share of target staff who used AI on a real task in the last two weeks | High after training, then falling |
| Time | Task time, including checking and corrections, against baseline | Faster drafts, but more rework |
| Quality | Manager review of a sample of outputs | Errors reaching customers |
| Risk | Near misses, data incidents, policy questions | Zero reports, which often means nobody is reporting |
| Confidence | A short staff survey each quarter | Rising use with rising worry |
What usually goes wrong?
| What goes wrong | Why | What to do |
|---|---|---|
| Licences bought, then unused | Purchase came before the tasks | Pilot first and review seats after 90 days |
| Pilots that never end | No gate or decision date | Set the review date on day one |
| A DPIA starts late and stalls the plan | Nobody sorted the queues | Screen uses in the first fortnight |
| Managers quietly block it | They were trained last, or not at all | Train them first and set a shared target |
| Champions burn out | The role was added to a full job | Protect some hours and thank them publicly |
| Every department picks its own tool | No approved list | Publish one, with a route to ask for additions |
| One serious mistake kills the programme | No error log and no agreed response | Decide in advance how incidents are handled |
| Staff fear for their jobs | Leaders said nothing | Speak early, and be honest about what is unknown |
Three prompts for the programme lead
Remove names and personal details before you paste anything.
1. Score your task list
I am running an AI rollout at a Kenyan company with [number] staff. Below is a list of tasks from department heads. Context: we want two or three low-risk pilot tasks. Constraints: for each task, score 0 to 2 on frequency, ease of checking, harm if wrong, personal data and manager willingness, using the scale I give here: [paste scale]. Show the results in a table with the total. Mark any task with 0 on "harm if wrong" as excluded. Do not add tasks that are not on my list. Customisation: the reader is our executive sponsor, so end with your top three and one sentence on why. Tasks: [paste]
2. Turn the numbers into a monthly report
Below are this month's rollout figures: [paste]. Write a one-page update for our executive team. Context: we are in [phase]. Constraints: use plain English, lead with the decision we need from them, show what improved and what did not, and do not claim any cause the numbers cannot support. Mark anything unclear with [CHECK]. Customisation: they will read this on a phone before a meeting.
3. Draft honest answers for staff
Write a short question-and-answer note for staff about our AI rollout. Context: people are asking whether jobs will be cut, whether they will be monitored and whether they must use AI. Here are the facts management has agreed: [paste]. Constraints: use only these facts, say "we do not know yet" where they are silent, make no promises about jobs, and stay under 250 words. Customisation: it will be read on WhatsApp by staff with different levels of English, so use short sentences.
Have HR and your lawyer read that third note before it goes out.
How much of this can you skip?
You can shorten the plan if your risk is low. Use this table to decide.
| Your situation | What to change |
|---|---|
| Under 150 staff, one site, little personal data | Merge the Prepare and Pilot phases, and aim for about six months |
| Bank, insurer, hospital or other regulated body | Allow more time. The Central Bank of Kenya closed consultation on draft AI guidance for banks on 7 November 2026, so check where that stands |
| Staff already using AI heavily | Spend less on basics and more on rules, checking and sharing |
| AI will make or shape decisions about people | Treat those uses as a separate project with legal review |
| Leadership cannot release any staff time | Wait. A rollout without protected time rarely lasts |
Some problems will not be fixed by a rollout. If a process is broken, your data is a mess or your managers do not trust their teams, AI tends to expose those problems. Sort them out first, or at least do not expect AI to solve them. Our AI readiness assessment can help you see where you stand.
Questions people ask
How long does an AI rollout take for a company with 100 or more staff?
We plan for about a year, from preparation to having AI in normal routines. A simple, low-risk company can do it in about six months. If a use needs a DPIA, allow at least 60 days after submitting it.
Should we train everyone at the same time?
No. Train in waves, with managers first and groups of about 20. Each wave learns from the last, and your business keeps running.
Do we need a chief AI officer?
Most companies of this size do not. You need a senior sponsor, a programme lead and a clear owner once the project ends. The title matters less than having someone accountable.
Is there a law in Kenya that controls AI rollouts?
There is no AI-specific law in force as of November 2026. The Data Protection Act and its 2021 Regulations already apply to any personal data. 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. See our explainer on the AI Bill.
Should we ban public AI tools while we prepare?
A total ban is hard to enforce and tends to push use out of sight. A short interim rule works better: use approved tools, never paste personal or confidential data, and check everything.
What should the first pilot be?
Pick something frequent, easy to check and free of personal data, in a team with a willing manager. Routine drafting and report summaries are common starting points.
Who should pay for the rollout?
That depends on how your budgets work, but keep one owner for the total. If costs are spread across departments with nobody watching, unused licences go unnoticed.
Sources
- The Data Protection Act, No. 24 of 2019, Gazette copy on the ODPC website, section 31, last checked 10 November 2026
- The Data Protection (General) Regulations, 2021, ODPC, regulations 49 and 52, last checked 10 November 2026
- PPC Land, report on the ODPC's draft AI guidance note, last checked 10 November 2026
- Digital Policy Alert, status of the ODPC draft Guidance Note on Artificial Intelligence, last checked 10 November 2026
- Digital Policy Alert, CBK consultation on draft AI guidance for banks, last checked 10 November 2026
- The Star, "Senate committee shoots down Nyamu's Bill on AI" (15 October 2026), last checked 10 November 2026
- Employment Act, No. 11 of 2007 (Revised Edition 2022), hosted copy, section 40, last checked 10 November 2026
- Central Bank of Kenya, Survey on Artificial Intelligence in the Banking Sector, last checked 10 November 2026
- PPC Land, report on McKinsey's State of AI 2025 survey, last checked 10 November 2026
- Dell'Acqua and others, Navigating the Jagged Technological Frontier (working paper, September 2023), last checked 10 November 2026
- Brynjolfsson, Li and Raymond, Generative AI at Work (NBER Working Paper 31161), last checked 10 November 2026
- Microsoft and LinkedIn, 2024 Work Trend Index, last checked 10 November 2026
- The Star, "Countries with the highest use of AI tools" (28 April 2026), last checked 10 November 2026
- Microsoft, Microsoft 365 Copilot pricing, last checked 10 November 2026
- Microsoft Learn, Enterprise data protection in Microsoft Copilot and Copilot Chat, last checked 10 November 2026
- Google, Use Gemini Apps with a work or school account, last checked 10 November 2026
- NIST, AI Risk Management Framework, last checked 10 November 2026
Where FuKazee fits in
FuKazee helps Kenyan organisations with the people side of a rollout. We can train leaders and line managers, run staff groups in waves, and help you test whether a pilot worked. Since 2022 we have trained more than 3,000 people.
We assume nobody knows AI, we avoid jargon, and we say plainly that nobody becomes an expert in a few days. 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. Your data protection and legal decisions stay with your own advisers.
See how this works on our page for AI training in Kenya.
To plan a rollout for your company, 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.
