AI in business software: an expert for every role, with the decision left to you

Most of what is called “AI” in business software stops at answering questions: you ask, it writes a paragraph. The real value starts when the assistant knows your data and can do part of the work itself — prepare receipt vouchers, draft an adjusting entry, gather overdue invoices into a follow-up list. That raises a fair question for every business owner: who controls what the assistant does, with what authority, and who owns the decision?

This article sets out the principles we believe any assistant working inside your books must follow, and how we applied them in Arttop 360 AI.

From a general assistant to a specialist

A general assistant answers everything from a single angle. Work in a business, however, is divided into roles: the accountant thinks in accounts and entries, HR in contracts, documents and expiry dates, the collections officer in due dates. So we organised the assistant into personas, each with its own tools, the records it may read and the operations it may propose — and nothing beyond them.

• Business owner: an executive briefing, explanations of why figures changed, a cash forecast and documents nearing expiry.
• Accountant: drafts journal entries pending approval, fills invoices and vouchers, and helps reconcile the bank statement.
• HR: employee records, holidays and shifts, employee requests and expiring documents.
• Medical coordinator: patient files, medical services and appointments.
• Auditor: runs the data audit, reads its findings and records a decision on each.
• Lawyer: reads matters, powers of attorney, hearings, deadlines and judgments, and prepares documents.
• Engineer: reads quantities, variations, certificates, schedules and site reports, and prepares purchase orders and certificates.
• Collections officer: receivables aging and customer statements, and prepares receipt vouchers and follow-up tasks.
• Chief financial officer: budget versus actual, profitability, cash forecasting and receivables aging.
• Procurement officer: payables aging, items at their reorder level and supplier prices, and prepares purchase orders and payment vouchers.
• Sales and POS officer: sales analysis, POS shifts, fast-moving items and active promotions, and prepares quotations and invoices.
• External auditor (CPA): tests journal entries, selects confirmation samples and proposes adjusting entries as drafts pending approval. It never issues an opinion or signs.
• Cost analyst: profitability by cost center, project and item, and proposes allocation entries whose shares add up to one hundred percent.
• Tax and zakat advisor: reconciles the return with the books, flags wrong VAT numbers and rates, and shows a draft zakat base for guidance, not for filing.
• Treasurer: treasury balances, cheques falling due, cash counts and their differences, and prepares transfers, receipts and payments.
• Storekeeper: item balances, count sheets and differences, and prepares stock movements pending approval without issuing more than is available.

Each user sees only the personas whose screens they can open: someone without access to the financial statements will not see the CFO persona, and someone who does not open projects will not find the engineer.

A plan, then a preview, then your confirmation

You write your request in everyday language — for example, “list customer invoices more than sixty days overdue and create a follow-up task for each customer”. The work then moves through steps the assistant cannot skip:

• It reads your data using only the read tools allowed for its persona.
• It proposes a plan of numbered steps, each stating what it will do and why.
• The program checks every step against its own engines and your permissions before showing it, so anything you are not allowed to do appears blocked, with the reason.
• You review the preview card, then press «Execute» or cancel.
• Steps run through the same doors you use: an invoice or voucher opens in its usual window, already filled, and you save it through the normal path — with its numbering, its entry and its submission to ZATCA.
• A result card appears with links that open each record where it lives.

Examples of everyday requests

• To the business owner persona: “Summarise last month against the month before, and explain why net profit fell.”
• To the accountant: “Prepare this month’s office rent accrual on the administration cost centre.”
• To HR: “Which employees have residence permits or contracts expiring within sixty days?”
• To the engineer: “Prepare a purchase order to our usual supplier for the remaining quantity of the concrete item.”
• To the procurement officer: “Which items have reached their reorder level, and who was the cheapest supplier for each on recent invoices?”

In each of these examples the assistant comes back with a plan you read before anything runs, and anything that is purely a question is answered with figures from your own books rather than the model’s estimate.

The program calculates amounts, not the model

Language models are good at understanding and writing, but they are not a calculator you can rely on for your books. So no amount on a document comes from the model’s arithmetic: tax, totals and entry lines come from the same program engines that calculate them when you work by hand. The model proposes, the engine calculates, and you decide.

Approval stays with a person

The assistant never approves a document or posts an entry. Journal entries it prepares are always saved pending approval. If you explicitly ask in your own request to approve specific documents, and you hold the approval permission, all it does is open the decision window for each document — the decision is yours. The segregation of duties you built into your business does not change because a new participant has joined the work.

Every step is logged

Before execution, the program checks with the server how much usage your establishment has available, then logs every step and its result on the server. If the start cannot be confirmed, nothing runs — so you never find half-finished work with no record of where it came from.

Questions to ask any vendor before you switch AI on

• Does the assistant work with the user’s own permissions, or with broader ones?
• Does the user see what will run before it runs, and can they cancel it?
• Can the assistant approve or post on its own?
• Who calculates amounts: the model or the program’s engines?
• Is what it did logged, and where is that log kept?
• Where is your data sent, and does the privacy policy say so explicitly?

In Arttop 360 the assistant is optional and enabled per user, and the privacy policy states what is sent to the model provider to generate the answer. We suggest starting with one persona on a repetitive task you know well — collections follow-up or a weekly briefing — and expanding once you have seen the results yourself.

Related pages

• Privacy policy: what is sent to the model provider
• Security and data protection
• Plans and pricing