Using AI Assistants in Back-Office Work, Safely and Sensibly
Staff are already pasting work into AI chat tools. A few clear rules on where they help, where they mislead and which data stays out will keep that use safe.

In many offices across the region, AI (artificial intelligence) assistants arrived without a decision. Someone in finance started using a chat tool to draft supplier emails, a colleague in human resources (HR) used it to tidy up a job description, and within months half the back office was pasting work into tools that nobody had approved. Some of that use saves real time. Some of it puts confidential data in places it should not be, or produces confident answers that are simply wrong.
A sensible approach sits between banning AI tools and ignoring them. It starts with knowing what these assistants are good at.
What AI assistants do well, and what they do badly
Today’s AI assistants are language tools. They are strong at working with text you give them and weaker at facts they have to recall or calculations they have to get exactly right. That distinction explains most of the good and bad experiences people report.
| Usually a good fit | Use with great care, or not at all |
|---|---|
| Drafting emails, letters and notices from your bullet points | Final figures in financial reports or tax calculations |
| Summarising a long document or meeting notes you provide | Legal or regulatory interpretation you will rely on |
| Translating and tidying text between Arabic and English, with review | Facts, dates or references the tool supplies from memory |
| Turning a messy procedure into clear numbered steps | Decisions about individual employees or customers |
| Writing Excel formulas or explaining how one works | Anything involving data you are not allowed to share |
| Extracting fields from documents into a structured format, with checks | Unsupervised replies sent directly to customers |
Use AI where a person can quickly check the output, and where an error is cheap to catch. Avoid it where errors are expensive and hard to see.
An illustrative example: the finance team
Consider the finance team of a distribution company in Amman. The finance manager notices that one accountant spends hours each week writing payment reminder emails, another summarises supplier contracts for renewal reviews, and a third prepares a monthly commentary on the management accounts.
All three are reasonable uses with the right setup. Reminder emails can be drafted from a template and the overdue list, then checked before sending. Contract summaries are useful as a first pass, but someone must confirm key terms such as notice periods and price clauses against the actual document, because a summary can miss or misstate a clause. The monthly commentary can be drafted from figures the accountant pastes in, but the numbers must come from the ERP (enterprise resource planning) system, not from the assistant, and the explanations of why figures moved must come from people who know the business.
What the team should not do is paste the full customer ledger, including names and balances, into a free consumer chat tool. That is where safety comes in.
Rules that keep it safe
- Use approved tools only. Choose one or two assistants, ideally on business or enterprise plans whose terms state how your data is handled and stored, and whether it is used to train the provider’s models. Free consumer versions often have different terms. Read them, or ask your information technology (IT) lead to.
- Classify what may be shared. A simple three-level rule works: public information can be used freely; internal information only in approved tools; confidential and personal data, such as salaries, customer identities, health information or bank details, not at all unless a specific setup has been approved for it.
- Respect personal data laws. Jordan’s Personal Data Protection Law of 2023, and equivalent laws in the Gulf, set rules on processing personal data and transferring it. Sending customer or employee details to an external AI service may fall under those rules. Get advice before allowing it.
- A person checks every output that leaves the team. Emails, reports and translations are reviewed by the person responsible before they are sent or filed. The tool drafts; the person decides.
- Never trust a figure or reference you did not supply. If the assistant states a number, a law or a date, verify it from the original source.
- Keep a short register of uses. List which teams use AI for what. It helps you spot risky uses early and share good ones across departments.
Moving from individual use to real workflow gains
Individual use of chat tools brings modest gains. Bigger gains come when AI is built into a defined workflow with clear inputs, checks and owners. For example, supplier invoices can be read by a document extraction tool that uses AI, with every field validated against the purchase order before posting. Customer emails can be classified and routed automatically, with replies still written or approved by staff.
These designs follow the same principle as any good automation: the process is understood first, the AI handles a narrow, repetitive step, and a person handles exceptions. A short pilot on one process, with measures agreed in advance such as time saved and error rate, will tell you more than months of general experimentation.
Training matters too. A one-hour session showing staff how to give the assistant clear instructions, how to check its output and which data must never be pasted will prevent more problems than a long policy document nobody reads. Pair it with a one-page summary of the rules, in Arabic and English, kept where people work.
Sensible AI, built into real work
AI assistants in the back office are useful tools with real limits. Treated as fast drafting and summarising aids, governed by simple rules on data and review, they save time with little added risk. Treated as a source of truth, they create errors that are hard to spot. The Automate stage of the KEMLAT Method applies here as much as anywhere: automate only what deserves it, and only once the process is clear.
Our Intelligent Automation service helps organisations identify where AI and automation fit their back-office processes, design the checks around them and measure whether they deliver.



