How to check AI output for accuracy before you send it
Check every fact against the source document, redo totals and percentages yourself, look up each name, date and regulator circular, and rewrite for the customers the answer leaves out. AI sounds just as sure when it is wrong.
This is for staff at Philippine banks and insurers who have started using AI on real work. Emails to customers, summaries for a manager, replies to complaints, figures for the monthly report.
If you need to check AI output for accuracy before it leaves your screen, you do not need a special tool. You need the source documents you already have and a habit of looking at four things: the facts, the figures, the names, and who the answer leaves out.
The checks below come from our AI prompting workshops for banks and insurers. Each one starts with a belief people often hold about AI answers.
A confident AI answer can still be wrong
Say you ask AI to explain the waiting period on a health rider for a customer email. It comes back in clean sentences with no hedging. It reads like someone who knows the product well. That tone is what catches people out. AI writes in the same steady voice whether a fact is right or wrong, so how sure it sounds tells you nothing about what it says.
Go through the answer and mark each fact someone could question. A fee, a waiting period, an eligibility rule, a turnaround time, a line about what the policy covers. Check each one against the policy wording, the core system or the source document. If you cannot find where a fact came from, take the line out or ask the person who owns that product.
A reply that says a claim will be processed within a set number of days is a promise. If that number came from AI and not from your service standard, the customer will still hold you to it.
AI can change the figures you gave it
You paste the monthly branch report into an AI tool and ask for a summary for your boss. The summary is tidy. The figures look right because they look familiar. But AI can round a number, move a figure to the wrong row, or produce a total that was never in your file. It does this even when every number it needed was right in front of it.
Put each figure in the summary next to the original report, one at a time. Do not trust any total or percentage the AI worked out. Redo it yourself on a calculator or in the spreadsheet. AI often gets simple sums wrong, which surprises most people the first time they see it.
If the summary says applications went up by a certain percentage month on month, find both monthly figures in the report and work out the change yourself. If your number and the AI's number differ, yours goes in the email. Then check the rest of the summary, because if one figure is wrong, others may be too.
AI mixes up names, titles, dates and circular numbers
You ask AI to draft a letter to a corporate client that mentions a recent Bangko Sentral or Insurance Commission circular. The draft looks finished. Look closer and the client's job title is out of date, the product has its old name, the date has the day and month swapped, and the circular number looks plausible but does not match any real circular.
Look up every person, product and regulator circular the answer mentions. Check names and titles against your CRM or the client's latest email signature. Check product names against the current product list. Open the circular itself and confirm the number, the date and what it actually says.
This matters more than it seems. A client will forgive a clumsy sentence in an email. A wrong name, or being called by last year's title, is the mistake people remember, and it tells them nobody read the letter properly before it went out.
AI writes for the typical customer and leaves others out
Ask AI for an SMS about a change in ATM fees, or a letter explaining which documents a claim needs. What comes back is clear enough for a confident reader who is comfortable with banking terms and reads English easily. It quietly leaves out the older customer who does not use online banking, the customer who reads more easily in Tagalog, the person using a screen reader, and the customer filing a claim for the first time.
Before you send, ask yourself who the answer misses. Pick one of those customers and read the message as them. Swap words like "rider" or "beneficiary designation" for plain words or a short explanation. Cut long sentences in half. If many of your customers read in Tagalog, write that version too, and have someone who reads Tagalog every day check it rather than trusting the AI translation.
Rewrite for them before it goes out. A letter an older customer can follow on the first read is one a busy customer can follow too.
What to do this week
- Pick one AI-written email, summary or reply you are about to send and run all four checks on it.
- Mark every fact, figure, name and date in that draft, and note where you checked each one.
- Redo any total or percentage yourself before a report goes to your boss.
- Keep the policy wording, the current product list and the latest circulars open whenever you work with AI, so checking is quick.
- Read one customer message as an older customer or a Tagalog reader would, and rewrite the lines they would trip on.
Questions people ask
Can I trust AI figures if they came from my own report?
No. AI can round, swap or make up figures even from your own file, so put each one next to the original and redo any total or percentage yourself.
Which AI mistakes do clients notice most?
Wrong names, old job titles and wrong dates. A client will forgive a clumsy sentence, but a misspelled name or last year's title is the mistake people remember.
Should I use AI to translate customer messages into Tagalog?
You can use it for a first draft, but have someone who reads Tagalog every day check it before it goes out.
On-Off Group trains teams, tests products with real customers, finds where a transformation has stalled and builds what gets it moving, for banks, insurers and enterprises in the Philippines, since 2015. Who we are.
