A freelancer sends a client a product description set for a Lagos skincare shop. The text reads well. The client then spots a claim that the cream "is NAFDAC approved", which nobody ever told the writer, and which the AI invented to sound convincing. The client doesn't just reject that line. They start doubting everything else you've delivered.
This is the real risk of AI-assisted work. The tool produces fluent text whether or not it's true, so the polish hides the mistakes. Your job, if you're being paid, is the checking. This guide gives you a repeatable review process you can run in 15 to 30 minutes on most deliveries, even on a phone with limited data.
Why AI drafts fail in predictable ways
Text generators predict likely wording. They don't look things up the way you'd check a source, and they don't know your client. That produces a small number of error types you can learn to hunt for:
- Invented facts: statistics, dates, prices, laws, quotes, company details, certifications, and "studies show" claims.
- Invented sources: links, book titles, or article names that look real but don't exist.
- Wrong context: American spellings and examples, dollar prices when the client sells in naira, or advice that assumes stable power and card payments.
- Drift from the brief: wrong tone, wrong length, missing a requirement you were given.
- Calculation slips: totals, percentages and conversions that don't add up.
- Generic filler: sentences that could appear in any business's copy and say nothing about this one.
- Language errors in local languages: unnatural Yoruba, Igbo, Hausa, Pidgin or Swahili, or the wrong register.
Step 1: Re-read the brief before you read the draft
Start with what the client asked for, not with what the AI wrote. Copy the brief into a short checklist: audience, purpose, length, tone, must-include items, must-avoid items, deadline, file format.
Then read the draft against that list. Most rejected work fails here, not on grammar. If the client said "no discounts mentioned" and the draft offers "limited-time savings", that's a miss.
Step 2: Mark every checkable claim
Read once, slowly, and highlight anything a reader could prove wrong. These are:
- Numbers, percentages, prices and dates.
- Names of people, companies, places, products and institutions.
- Rules, laws, regulatory bodies and requirements.
- Claims about results, such as "reduces costs by half" or "approved by".
- Anything described as a study, report or ranking.
Everything highlighted needs one of three outcomes: verified against a reliable source, replaced with information from the client, or deleted. There's no fourth option called "sounds right".
Step 3: Verify against the right source
The best source is the client. Their price list, product sheet, past invoices and brand notes outrank anything a tool produced. For public facts, go to the original: the official website of the agency, company or institution involved, not a blog repeating it.
A few habits that save time:
- If the draft cites a link or source, open it. If it doesn't load or doesn't say what the draft claims, remove the claim.
- Search for a quoted statistic by its exact wording plus the supposed source. No trace usually means it was made up.
- For rules and fees that change (tax, bank charges, visa costs, regulations), send the reader to the official page rather than stating a figure from memory.
- Never ask the same AI tool to confirm its own facts and treat that as verification. It can repeat the mistake confidently.
Step 4: Check the numbers by hand
Recalculate every total, average and percentage with a calculator or a spreadsheet, even when the maths looks simple. Be especially careful with currency conversions, since rates move and the tool has no live feed unless it's connected to one. Get the rate from the client or a bank or official source on the day, and write the date beside it.
For spreadsheet deliverables, spot-check three rows manually, then check that totals match a quick independent sum. If you want a fuller routine for sheets, the reconciliation guide shows how to match records line by line.
Step 5: Make it sound like the client, not like a tool
Clients can often tell. Look for these signs and rewrite them:
- Openings such as "In today's fast-paced world" and closings that restate the whole text.
- Every paragraph the same length and shape.
- Vague praise words with no specifics: "high-quality", "innovative", "tailored solutions".
- Examples that don't fit the market: snow, Black Friday sales in dollars, or "pop to the shop" in a Nairobi pitch.
Replace generic lines with a specific detail from the client: the neighbourhood they deliver to, how long a job takes, what customers usually ask. Read it aloud. If you'd never say a sentence to a customer, change it.
Keep a client glossary of preferred words and names so the voice stays consistent. This glossary guide explains how, particularly for local-language work.
Step 6: Handle local languages with extra care
If you deliver in Pidgin, Yoruba, Igbo, Hausa or Swahili, don't assume the draft is fine because you can't see an error. Tone marks, dialect choice and formal versus street register are easy to get wrong. Ask a fluent native speaker, ideally from the client's audience region, to read the final version. Pay them a small fee if needed and build it into your price. Mistakes in names, greetings and religious or cultural references cost the most.
Step 7: Protect the client's data and your own honesty
Before you ever paste material into a tool, check that you aren't exposing customer phone numbers, bank details, or confidential documents. Remove or replace them. The post on using AI at work without leaking customer data may exist in our archive, but the rule is simple: if you wouldn't hand it to a stranger, don't paste it.
Also read your contract or the client's brief for rules about AI use. Some clients ban it, some require you to disclose it, and some don't care about method as long as the result is accurate. If a rule exists, follow it. If you're unsure, ask in one line: "I use AI tools to draft and then edit and fact-check everything. Is that fine for this project?"
A worked example (hypothetical)
Say you're paid ₦25,000 to write a 600-word blog post for a Kumasi catering business about planning food for a wedding of 200 guests. The AI draft includes:
- "A 200-guest wedding needs exactly 40 kg of rice." That's a specific claim with no source. You check with the client, who says they usually cook about 30 kg of jollof for that size. You replace the figure and say "based on our own events".
- A price of "$5 per head". The client charges in cedis. You delete it and ask for the current rate card.
- A line saying "certified by the Food and Drugs Authority". You can't confirm it, so you ask the client. They say a certificate is pending. You delete the line.
- A closing paragraph that repeats the intro. You cut it and add a clear booking instruction with the client's WhatsApp number.
The review takes about 25 minutes, and the client receives something they can publish without embarrassment. That's what the fee is for.
A quick comparison of review methods
| Method | Catches | Misses | Time |
|---|---|---|---|
| Read it once for flow | Awkward wording | Fake facts, wrong numbers | 5 min |
| Brief checklist | Missed requirements, wrong tone | Factual errors | 5 min |
| Highlight and verify claims | Invented facts and sources | Tone problems | 10-20 min |
| Native-speaker read | Language and cultural errors | Factual errors | Varies |
| Read aloud next day | Stiff or generic phrasing | Facts | 10 min |
No single method is enough. For paid work, combine the brief checklist and claim verification at minimum.
Common mistakes
- Checking only spelling. Fluent text with a fake fact is the dangerous kind.
- Trusting a confident tone. Tools sound equally sure when right and when wrong.
- Skipping the review under deadline pressure. Tell the client you need an extra few hours instead.
- Not saving your changes. When you find the same error twice, note it. A simple mistakes log helps you adjust your prompts.
- Pasting the output straight into the client's file. Always paste into your own document first.
Make it a checklist you can reuse
- Brief matched: audience, tone, length, must-haves.
- Every number, name, date and claim verified or removed.
- Sources opened, not assumed.
- Maths and currency rechecked, with dates.
- Local examples, spelling and currency fit the client's market.
- Generic filler replaced with client specifics.
- Local-language text read by a native speaker where relevant.
- No private data exposed; client's AI rules followed.
- Final read aloud, file name and format correct.
Frequently asked questions
Do I have to tell clients I used AI?
Follow your contract and the client's stated rules. If none exist and you're unsure, a short, honest question before you start avoids trouble later.
Can I use another AI tool to check the first one?
It can help spot awkward wording or inconsistencies, but it isn't verification. Facts still need a real source.
How long should a review take?
Roughly one-third of the time you spent producing the draft is a reasonable starting point for factual work. Short social posts can take less; reports and data work take more.
What if I find an error after delivery?
Tell the client quickly, send the corrected version, and say what changed. Early honesty protects the relationship far better than hoping nobody notices.
Your next step
Take the last piece of AI-assisted work you did and run Steps 2 and 3 on it today. Count how many claims you can't verify. That number tells you how much review time to build into your next quote.
Results from any side hustle or online work vary widely, and nothing here is a promise of income. Be wary of anyone who asks you to pay upfront to start earning.