Back to blogAI Training & Skills

AI for Accountants and Auditors in Kenya: What to Hand Over, and What Never To

November 11, 2026 · 10 min read · Gavin

Share
A finance professional checks figures on his laptop beside a calculator and printed schedules, testing AI work for accountants

The short answer: AI is good at the writing around the numbers: variance commentary, review notes, audit programme drafts, report narrative, policy summaries, client emails and data cleaning. It must never sign off a figure, form an audit opinion or see a client's identifiable data on a personal account. Test every AI number against the source before it leaves your desk.

Most accountants we train already use ChatGPT for something: a quick email, a formula, a definition. Few use it for the work that actually takes their week, because they are not sure where the line is. That caution is right. A wrong number in a management pack or an audit file is not a typo. It is a professional risk.

This guide draws the line. It is the same approach we use in hands-on AI training in Nairobi, with finance examples. It covers the tasks to hand over, the tasks to keep, and the testing habit that makes AI safe for finance work in Kenya.

Can accountants and auditors in Kenya use AI safely?#

Yes, if AI drafts and you decide. Use it for wording, structure and first passes on clean, anonymised data. Keep judgement, sign-off and client confidentiality with a qualified person. The professional and legal duties do not move to the tool.

Three rules cover most of the risk:

  • Anonymise first: replace client names, PINs, account numbers and staff names with labels before anything goes into an AI tool.
  • Use a work account the firm controls: not a personal free account, and with model training on your chats switched off where the tool allows it.
  • Test every number: AI explains well and counts badly. Recompute anything it calculates.

Which seven tasks can you hand over to AI?#

Hand over the tasks where the risk is in the wording, not in the number. AI writes a solid first draft of commentary, notes and summaries in minutes. You then check the facts and add the judgement. Here are the seven we cover with finance teams.

  • 1. Variance commentary: paste the budget-versus-actual lines and your notes on the reasons. AI turns them into clear management commentary.
  • 2. Review notes: give it your rough points from a file review. It turns them into numbered, polite review notes for the preparer.
  • 3. Audit programme drafts: describe the client's sector and risk areas. It drafts procedures for you to edit against your firm's methodology and the standards you apply.
  • 4. Management report narrative: it turns a page of figures you have already checked into a readable summary for a board that does not read schedules.
  • 5. Policy summaries: upload a long internal policy or a public standard. It summarises the parts that apply to a given transaction, with page references to check.
  • 6. Client emails: requests for missing documents, explanations of a finding, polite chasers for a late trial balance.
  • 7. Data cleaning: messy exports from M-Pesa statements, bank downloads or an old accounting system. AI writes the Excel formulas or steps to clean them, which you run yourself.

Here is a prompt for the first task that participants find useful straight away.

text
You are a management accountant writing for a non-finance board. Below are
budget and actual figures for the quarter, already checked, with my notes
on the main causes. Write variance commentary of no more than 200 words.
Cover only variances above 10% or Ksh [amount]. For each one: what happened,
why, and what management is doing about it. Do not calculate any new
figures. If my notes do not explain a variance, write [EXPLANATION NEEDED].

[paste figures and notes, with client names removed]

The last two lines matter most. They stop the AI from inventing numbers or reasons it does not have.

A worked example: review notes for Pwani Freight#

Pwani Freight is an invented freight forwarder we use in demonstrations. Say you are reviewing its year-end file and your rough notes read: "debtors ageing doesnt tie to TB by 340k, no support for fuel accrual, FX rate on USD payables looks old, revenue cut-off not tested for Dec shipments". That is fine for you, and unhelpful for the preparer.

text
Turn my rough review points into numbered review notes for the preparer of
this file. For each note: what I found, why it matters, and what evidence or
correction I need, with a column for their response. Polite and direct.
Keep my figures exactly as written. Do not add findings I did not raise.

[paste your rough points]

The draft comes back as four clear notes, each with the evidence requested. You check that the meaning has not changed, adjust the tone for the person receiving it, and send. Ten minutes of typing becomes two minutes of reading. The judgement in the notes is still entirely yours: the AI only formatted what you found.

This is the pattern for all seven tasks. You bring the facts and the judgement, the AI brings speed and structure, and you read every word before it goes out.

Which three tasks must you never hand over to AI?#

Never let AI sign off a figure, form an opinion, or hold client data you have no right to share. These are the tasks where the professional is accountable by law and by the standards they practise under, and where an AI mistake cannot be caught by reading the wording.

  • 1. Final figures and calculations nobody tested: AI language tools predict words. They can add up a column wrongly and present the answer with complete confidence. Any number that goes into a return, a set of accounts or a pack gets recomputed from source.
  • 2. Audit opinions and professional judgement: materiality, going concern, the opinion on the financial statements, whether a weakness is significant. AI can list the factors. The decision, and the signature, stay with the engagement partner.
  • 3. Identifiable client or staff data on a tool the firm does not control: a client's full ledger with names, KRA PINs, bank details or payroll on a personal ChatGPT or Claude account is a confidentiality breach waiting to happen.

The third one has a legal edge. Under Kenya's Data Protection Act, 2019, the Data Commissioner can issue a penalty notice of up to Ksh 5 million or 1% of the previous year's annual turnover, whichever is lower (section 63, checked October 2026). For a practical list of what not to paste, read our guide on AI and Kenya's Data Protection Act.

How do you test an AI number before it leaves your desk?#

Treat every AI figure as unaudited until you have tied it to source. The rule we teach finance teams is simple: no AI number leaves your desk untested. That means recomputing it, tracing it back, and checking it makes sense.

Use these four checks, the same way you would review a junior's schedule:

  • Recompute: redo any total, percentage or ratio in Excel yourself. Do not ask the AI to "double-check" its own sum.
  • Trace: every figure in the commentary must match a line in the trial balance or the schedule. If you cannot find it, delete it.
  • Reasonableness: does the movement make sense against what you know about the client this year? AI does not know that the client lost its biggest customer in March.
  • Quote the source: ask the AI to cite the row, page or cell for each fact it uses. Missing citations are a warning sign.

A better pattern is to let AI write the method, not the answer. Ask it for the Excel formula or the steps, then run them on the real data in your own spreadsheet. The calculation happens in a tool you trust, and the AI never needs the confidential file.

text
I have an Excel sheet with columns: Date, Description, Debit, Credit, Balance.
The descriptions come from a bank export and are inconsistent. Write the
formulas to (1) flag duplicate transactions, (2) group descriptions into
five categories I will name, and (3) total each category by month.
Explain each formula in one line. Do not ask me for the data.

This is the data-to-report sequence we use with finance teams: clean the data with formulas you run yourself, analyse it in Excel, let AI draft the insight in words, then add the recommendation from your own judgement.

How should a firm roll out AI safely?#

Start small, write the rules down, and train people on real tasks. A firm that bans AI usually finds staff using it anyway on personal phones. A firm that allows it without rules finds client data in places it should not be.

A rollout that works for an audit firm or a finance team of ten to fifty people:

  • Week 1, rules: a one-page AI policy covering which tools, which accounts, what may never be pasted, and who reviews AI-assisted work. Keep it short enough that people read it.
  • Week 2, tools: company accounts for an approved tool, with training on chats switched off where available. No client work on personal accounts.
  • Week 3, training: a hands-on session on three of the seven tasks above, using the firm's own anonymised documents.
  • Week 4, review: the partner or finance manager reviews a sample of AI-assisted work and adjusts the rules.

Keep a short note on the file when AI helped draft a working paper. It costs one line and answers the reviewer's first question.

The "AI for Auditors" programme Cloudwise runs follows this sequence, built on the firm's own templates.

Cloudwise Limited, an ICT company in Westlands, Nairobi, runs hands-on AI training on Claude for individuals and teams. Our AI training for finance and audit teams is built on your own documents, with the safe-use rules above included.

Frequently asked questions#

Will AI replace accountants?#

No, but it will change the work. AI takes over first drafts, commentary and data cleaning, the hours that used to go to juniors. Judgement, sign-off, client relationships and professional accountability stay with qualified people. Accountants who learn to direct and check AI will produce more in less time than those who do not.

Can I upload a client's ledger to ChatGPT or Claude?#

Not as it is. Remove names, KRA PINs, account numbers and anything that identifies a person, and use a company account the firm controls with training switched off. Better still, ask the AI for the formulas or steps and run them on the real ledger in Excel. Check your engagement letter for confidentiality terms first.

Does ICPAK have guidance on AI?#

We did not find a standalone ICPAK guideline or practice note on AI use when we checked icpak.com in October 2026. ICPAK does run AI training for members, such as a webinar in August 2026 that covered verifying AI outputs, confidentiality and data privacy. Your existing duties of competence, confidentiality and professional scepticism still apply in full.

Training a team of 10 or more? Ksh 13,500 per person, built on your own documents. Book a 30-minute call at cloudwise.co.ke/ai-training/teams.

Share
G

Written by

Gavin

Comments

Be the first to comment.

Leave a comment

Related articles