Adding AI to Your App or ERP: Ask-Your-Data in Arabic and English, Smart Search and Summaries, With Guardrails

Most business systems hold the answers managers need. The problem is getting them out: the right report is three menus deep, the filter needs a date range nobody remembers, and the person who knows how to build it is on leave. So people export to Excel and work it out by hand, or ask the developer.

Adding AI to an app or ERP should fix that specific friction, not add a chat bubble for the sake of it. This page covers the AI features that genuinely help inside business software, the guardrails that keep them safe, how we connect to a system you already run, and the cases where a plain report is the better answer.

The features that actually help

"Ask your data": the guardrails that make it safe

Letting people question your sales, stock or HR data in plain language is powerful, and also where things go wrong if it's built carelessly. Our rules:

• Read-only. The assistant can query data but cannot create, edit or delete anything. Actions stay in the normal screens with normal approvals.
• Permission-scoped. It answers within the asking user's role and data scope. A branch manager asking about "all sales" gets their branch, and is told so.
• Shows its working. Every answer shows the numbers, the filters and period used, and the query or the defined measure behind it, so a finance user can verify it.
• Uses agreed definitions. "Revenue", "active customer" and "late order" mean what your business says, defined once. The AI maps the question to those definitions instead of inventing its own maths.
• Says when it can't. If the question is ambiguous or the data isn't there, it asks a clarifying question or says it doesn't know.

You can see this pattern in our own product. In PharmaRiwaq, our commercial platform for pharma companies, "Ask Riwaq" takes questions in English or Egyptian Arabic. Each answer shows its scope (for example, "your permitted territories, current cycle"), the key figures, the sources and a link to the detailed screen, and adds a caveat when part of the explanation is an inference or a data feed is delayed. Role and data-scope permissions are enforced in the database, with an audit trail. The public demo runs on a fictional company in your browser, and there the answers come from rule-based logic rather than a live AI model; it shows the experience and the guardrails, and a client build connects a real model to your data.

Dawawin HR, our HR and payroll system, applies the same idea to staff data. Its assistant answers in Arabic or English, names the records it used, and refuses to show company payroll to an employee-level account. Its guided Journey mode also lets a user type what they need, in Arabic or English, and walks them through the task one question at a time. Its public demo also runs on fictional data without a live AI model.

Smart search, summaries and anomaly flags

These are quieter features, and often more useful day to day than a chat window.

• Smart search understands that "شاشه سامسونج" and "Samsung monitor" are the same item, and that a customer name may be spelled three ways. In Arabic this means handling alef and taa marbuta variants, dialect words and Latin-letter typing.
• Summaries turn a 40-message ticket or a year of visit notes into five lines, with a link to the full record. They are drafts for a person to read, not records in themselves.
• Anomaly flags start with plain rules (an invoice above a limit, a duplicate number) and use AI where rules are hard to write, such as notes that read like a complaint. Each flag says why it was raised, and a person decides.

Forecasting: honest about the data it needs

Forecasts are where AI is most over-sold. A useful demand or cash forecast needs:

• History: generally at least one to two years of clean data if seasonality matters, such as Ramadan or back-to-school.
• Consistency: the same item codes and branch names over time. Renamed products and merged branches break patterns.
• Context: promotions, stock-outs and price changes recorded, or the model learns the wrong lesson.

With that, a forecast gives a range and an explanation, which helps purchasing. Without it, a moving average in a normal report is more honest. We check your data in the proof of concept and tell you which applies.

AI features in mobile apps

Inside a mobile app, AI is most useful where typing is hard: a sales rep dictating a visit note that is turned into structured fields, a technician photographing a part to find its manual page, or a customer asking a question in their own words. Voice and photo features need careful testing with Arabic accents and real lighting, so we pilot them with a small group first. For customer-facing chat on WhatsApp, see our WhatsApp AI agent.

Working with the system you already run

You don't need to rebuild your ERP to add AI. We connect in one of three ways:

1. Through its API, when the system offers one and your plan allows it (common in Odoo, Zoho and many SaaS tools). This is the cleanest route and respects the system's own permissions.
2. Through a read-only database copy (a replica or a scheduled export), when there is no suitable API. The AI never touches the live database, so it can't slow down or change daily work.
3. Built in, when we're building or extending the system ourselves; see our ERP system development page.

We confirm which route is possible before quoting.

Keeping AI costs under control

AI usage is billed per request by the provider, so we design to keep it predictable:

• The smallest model that answers well for each task; bigger models only where tests show they're needed.
• Plain code first: totals, filters and rules run in the database for free; the AI only interprets the question and explains the result.
• Caching common questions and summaries that haven't changed.
• Limits and alerts per user or per month, with the provider account in your name so you see every charge.

When AI is the wrong tool

Often the right answer is not AI at all:

• The same question every week? Build a dashboard or saved report. It's instant, free to run and always consistent.
• A yes/no rule (stock below 10, invoice over a limit) is a filter or alert, not a model.
• Exact figures for filing or audit should come from defined reports, with AI at most pointing to them.
• Little or messy data: fix the data first. AI on bad data gives confident wrong answers.

We'll say so on the first call if a report or a filter solves your problem.

How we work

1. Free call: the questions people ask, the system you run, who may see what.
2. Proof of concept (about 5 days): one feature on a read-only copy of your real data, tested by your team.
3. Review: correct answers, wrong answers, cost per question, and whether it's worth going further.
4. Production (about 3 weeks): permissions, audit logging, cost limits, Arabic and English testing, inside your app.
5. Handover: code and accounts in your name, plus running notes.

Price and timeline

AI usage is paid by you directly to the AI provider. You approve a written scope and fixed price first; nothing is paid before that. Projects up to $500 are paid 50% to start and 50% on delivery. Projects over $500 start with a 30% down payment, then 40% and 30% at milestones, each paid after you see it working.

When we're not the right fit

• Your system has no API and no way to export or copy data. We can't connect safely.
• You want AI to post transactions or approve payments on its own. Our features are read-only or suggest-then-confirm.
• You need an on-site team or a locally registered vendor for a tender.
• The business isn't Sharia-compliant, such as interest-based lending or betting. We don't take that work.

Related services: AI document processing for getting invoices and forms into your system, AI workflow automation for the repetitive steps around it, an internal knowledge assistant for policies and manuals, and AI chatbot development for customer-facing assistants.

FAQ

Can the AI change data in our ERP?

Not in our builds. Ask-your-data features are read-only. Where AI suggests an action, such as drafting a purchase order, a person reviews and confirms it through the normal screens and approvals.

Will it work with Arabic item and customer names?

Yes, with testing. We normalise common spelling variants, handle dialect and Latin-letter typing, and test with real names from your data, because that's where generic tools fail.

How do we know the numbers are right?

Every answer shows the period, filters and the measure or query used, with a link to the matching report. During the proof of concept we compare AI answers with reports your team already trusts.

Does our data leave our system?

Only the minimum needed for each question goes to the AI provider, through a paid tier whose terms say prompts aren't used to train its models. The data itself stays in your database or your read-only copy.

Can you add AI to a system someone else built?

Usually, if it has an API or we can get a read-only copy of the data. We review access and the data model before quoting.

How do we start?

Send us two or three questions your managers ask most, and which system holds the answer, on WhatsApp or through the project planner. You'll get a written proof-of-concept scope within 24 hours.