Where an AI assistant saves time, and where it does not
Seven office tasks an AI assistant handles well, four it should not be trusted with, four team rules, and a ten-minute test in Arabic before you choose.
An AI assistant saves an office time on work made of words, where a person can check the result faster than they could have written it. It wastes time, or does harm, where nobody checks, or where the output is a number or a decision. Below are seven tasks your staff can hand it tomorrow morning, four they should keep, the rules a team needs, and a ten-minute test to run before you choose a tool.
Seven tasks where it reliably saves time
- A first draft of an email or letter, in Arabic or English. You supply the facts and the tone, and correct a draft instead of starting from a blank page.
- Changing the register. A note typed the way you write on WhatsApp becomes a formal letter, and a three-page official letter becomes four lines for a colleague.
- Summarising a long document, contract or email thread, so you know which parts deserve a full reading.
- A first-pass translation between Arabic and English, which someone who knows both languages then checks.
- Pulling fields out of a scan. The supplier, date, invoice number and total from a scanned invoice or form, as a table you can paste into a spreadsheet.
- Turning meeting notes into a task list: who does what, and by when.
- Writing a spreadsheet formula. Describe what you need in words, such as the total of column D for the rows where the branch is Basra, and it writes the formula and explains it.
These tasks have one thing in common: the source is in front of you, so checking the result takes about a minute.
Four tasks where it fails or is risky
- Any number it produces from memory. An exchange rate, a customs duty, the article number of a law, a ministry's phone number, a total it added up itself. The assistant writes what is likely to sound right, and a wrong figure arrives with the same confidence as a correct one. Take numbers from your own documents, or from a source you can open and read.
- Legal or financial wording sent without a person reading it. A contract clause, a reply to the tax office, a letter of guarantee. The draft may be useful, but the signature on it is yours.
- Decisions about people. Ranking CVs, judging an employee, deciding who is given credit or a discount. The assistant can summarise a file. The decision and its reasons belong to someone who can be asked about them.
- Anything confidential pasted into a tool whose terms nobody has read. Client lists, salaries, contracts, copies of identity documents. Before staff paste such material into any assistant, someone in the company should have read what its provider says happens to it.
Good uses, and the part that needs a person
| Good use | Needs a person |
|---|---|
| Drafting an email or letter | Reading it before it is sent |
| Changing the register | Checking names, titles and dates |
| Summarising a document or thread | Reading the important clauses in full |
| First-pass translation | Review by someone who knows both languages |
| Fields from a scanned invoice or form | Comparing every amount with the original |
| Meeting notes into tasks | Confirming owners and dates with the people named |
| A spreadsheet formula | Testing it on rows where you know the answer |
The four risky tasks have no row, because a review alone is not enough for them. A person does that work, and the assistant at most helps with the wording.
Four rules for a team
- A person reads everything before it leaves. Whoever presses send is responsible for the text, whatever wrote it.
- Keep the source next to the output. The summary is filed with the document, the table with the scan, the translation with the original.
- Agree in writing what may and may not be pasted in. One page is enough, with three lists: allowed, allowed once names are removed, and not allowed.
- One person owns the rules. Staff know whom to ask, and that person updates the page when a tool or its terms change.
A ten-minute test with your own material
Product demonstrations use tidy examples, usually in English. Your office works in Arabic, on untidy material. Before choosing a tool, give it three things from your own desk.
- A real customer message in dialect. Copy one from WhatsApp as it is: Iraqi dialect, no punctuation, a spelling mistake or two, such as a customer asking how much the bill came to and when the goods will arrive. Ask what the customer wants, then ask for a polite reply, and see whether the assistant understood or guessed.
- A scanned Arabic document. An invoice or an official letter, with a stamp covering part of the text. Ask for the date, the reference number and the amounts, and compare each with the paper.
- A formal letter to a ministry. Give it the facts in two lines and ask for the letter. Check whether it follows the form Iraqi official letters take (reference number and date at the top, then the addressee, the subject line, the customary greeting and the closing courtesy) or reads like an English business letter in translation.
Judge the Arabic, not the English. The differences between assistants show most in Arabic, in dialect and on scanned pages. Expect mistakes from every tool, and prefer the one whose mistakes are few and easy to spot. The places where Arabic tends to break are described in what to look for in an Arabic AI assistant.
Cost: chatting, or building it into your own system
There are two ways to pay for an assistant.
Chatting. Staff open a window, type or paste, and copy the answer out. Tools of this kind are free or charged per user per month. The cost is easy to estimate and the work stays manual, one document at a time. For the seven tasks above, that is usually enough.
An API. A developer connects the assistant to your own system, so that nobody copies and pastes. Each scanned invoice that arrives is read into the accounting screen, or each website enquiry gets a drafted reply that waits for approval. You generally pay by the amount of text processed, plus the development work to connect it and to handle the cases it gets wrong. This pays when the same task repeats many times a day.
Start with chat. After a few weeks you will know which task your staff repeat most, and that is the candidate for an API project.
Where to start
Pick two tasks from the first list, write the one-page rules, and run the ten-minute test on the tools you are considering. One of them can be Peak AI, the assistant PearlFibers builds: bilingual chat with a right-to-left interface, documents you can download as PDF, Word or Markdown, reading of screenshots and scanned documents, and web search with cited sources. Chat is free with a PearlFibers account, and there is an OpenAI-compatible API for connecting it to a system. Put it through the same three tests as any other tool. If you already know which task you want connected to your own system, tell us about it.