How to apply AI in your business: a 6-step roadmap for MENA SMEs

To apply AI in your business, start with two or three costly, repetitive problems, not with a tool. Put the data those problems need in one system, set clear rules for approvals and privacy, run a measured 30-day pilot, train your people, then scale only what proves its value and review it every month.

In ten years of taking products from idea to operation, I have seen the same pattern again and again: the projects that worked started with a narrow problem and a named owner, not with a shiny tool. This guide turns that into a six-step roadmap for owners and general managers of 10–200 person companies in Saudi Arabia, Egypt and the Gulf. No coding is required, and no step needs a data science team.

If you are still deciding where AI could help, read our companion guide on AI use cases for business first. This article is about the how.

Why a roadmap matters more than the tool

Most small and mid-sized companies already have access to AI: a chatbot on someone's phone, an assistant inside email, a feature in the accounting software. Access is not the problem. The problem is that AI used ad hoc produces ad hoc results. One employee saves an hour; another pastes a customer list into a public tool; nobody can say whether the business is better off.

A roadmap fixes three things: it ties AI to business outcomes, it protects money and data, and it gives you a way to stop what does not work.

The 6-step roadmap

1. Pick 2–3 business problems, not tools

Write down every task that feels slow, repetitive or error-prone. Ask each department head for five. Then score each task from 1 to 5 on three factors:

• Frequency: how often it happens (daily = 5, monthly = 1).
• Time: how many staff hours it consumes each time.
• Error cost: what a mistake costs in money, penalties, customer trust or rework.

Multiply the three scores. A task scoring 60 or more out of 125 is a strong candidate. Typical winners in MENA SMEs are chasing overdue invoices, entering supplier invoices, answering repeated customer questions on WhatsApp, preparing weekly sales reports and drafting quotations.

Pick only two or three. For each, write one sentence: "We will know this worked if ___." That sentence becomes your success metric in step 4.

2. Get your data in one place

AI can only work with what it can see. If customers live in a spreadsheet, invoices in an accounting app, stock in another file and orders in WhatsApp chats, any AI tool sees a fragment and gives fragmented answers.

Before the pilot, clean the master data your chosen problems depend on:

• Customers and suppliers: one record each, no duplicates, correct tax numbers and contact details.
• Items and services: consistent names, units and prices.
• Chart of accounts: a structure your accountant agrees with, used consistently.

Ideally, move these into one system of record. That is the core idea behind an ERP; if the term is new, see what is ERP. Companies already issuing e-invoices to ZATCA in Saudi Arabia or the Egyptian Tax Authority have a head start, because those systems already require structured, accurate invoice data.

3. Set rules before you switch anything on

Write a one-page AI policy. It should answer four questions:

1. Who can use which tools, with which data? AI should act only within each user's existing permissions. A sales rep should not be able to ask an assistant for payroll figures.
2. What always needs a human? Anything that moves money, files tax, changes prices, or affects hiring and firing must be approved by a named person. AI drafts; people decide.
3. What data may leave the company? Decide which information may never be pasted into external or consumer AI tools: customer personal data, salaries, contracts, bank details.
4. Which laws apply? In Saudi Arabia, the Personal Data Protection Law is overseen by SDAIA, which publishes guidance through its National Data Governance Platform. In Egypt, Personal Data Protection Law No. 151 of 2020 applies, and the Personal Data Protection Center is the reference body. This is not legal advice: review your specific obligations, especially on cross-border transfers, with a qualified adviser.

For a broader framework, the NIST AI Risk Management Framework and the OECD AI Principles are useful, freely available references.

4. Run a 30-day pilot with a baseline

A pilot without a baseline is a demo. Measure the current process for one or two weeks before you change anything, then run the AI-assisted process for 30 days and compare.

Worked example (illustrative): collections at a Riyadh trading company. Imagine a 40-person trading company whose credit controller spends much of each week chasing overdue invoices.

• Baseline (2 weeks): record the number of overdue invoices, hours spent chasing, reminders sent, and the share of customers who reply with a payment date.
• Pilot (30 days): each morning, AI drafts polite Arabic or English reminders from the overdue list, tailored to the amount and days late. The credit controller reviews and approves every message before it is sent.
• Measure: hours spent chasing, reminders sent per hour, reply rate, how many drafts needed heavy editing, and any customer complaints.
• Decide: compare against the success sentence from step 1. Days sales outstanding moves slowly, so treat it as a later indicator, not a 30-day verdict.

A second illustration: supplier invoice capture at a Cairo distributor. Measure minutes per invoice and errors found in a weekly sample before and during the pilot, and check that captured data matches the supplier's e-invoice on the Egyptian Tax Authority system.

We are not quoting results here, because results depend on your data and process. The point is the method: same metrics, before and after, with a human approving every output.

5. Train people and change the process

AI that sits on top of an unchanged process rarely sticks. Update the procedure itself: who receives the draft, who approves it, what happens when the AI is wrong.

• Name a champion in each pilot team: someone curious, respected, and given time to help colleagues.
• Build a prompt library in Arabic (and English where needed): tested instructions for recurring tasks, such as "draft a second reminder for an invoice 30 days overdue, formal Saudi tone." Store it where everyone can find it.
• Clarify roles: the AI prepares, the employee checks and owns the result. Make it clear that the goal is to remove repetitive work, not to monitor people.

Short, hands-on sessions on real company tasks work better than long generic courses.

6. Scale and review monthly

When a pilot meets its success metric, extend it to more users or a second use case. When it does not, stop it and record why. Retiring a pilot is a valid outcome.

Hold a 30-minute monthly review covering:

• Cost: subscriptions and usage-based AI charges against the time saved.
• Quality: a sample of AI outputs checked by a person; error trends.
• Risk: any data or permission incidents.
• Next: which scored problem from step 1 comes next.

A 90-day plan

One-page readiness checklist

• We have 2–3 written business problems with a score and a success sentence.
• Each problem has a named owner.
• Customer, supplier and item records are de-duplicated and current.
• Our chart of accounts is agreed and used consistently.
• Core data lives in one system, not in personal spreadsheets.
• User permissions are defined by role.
• We have listed the decisions that always need human approval.
• We have a written rule on what data may go to external AI tools.
• We have read the official PDPL guidance for our country.
• We have measured a baseline for each pilot.
• Each pilot team has a champion with time allocated.
• A monthly review date is in the calendar.

Common mistakes

• Buying a tool first and then searching for a problem it can solve.
• Skipping the baseline, so nobody can prove the pilot helped.
• Automating a broken process: AI makes a messy process faster, not better.
• Pasting sensitive data into consumer chatbots without a policy.
• Letting AI post to the accounts or send payments without human approval.
• Measuring logins instead of outcomes: usage is not value.
• Never stopping anything: zombie pilots consume budget and attention.

How to think about budget

We will not quote prices, because they vary by vendor, users and volume. Think in three buckets:

• Software: per-user subscriptions plus, often, usage-based AI charges that grow with volume. Set usage caps and alerts from day one.
• People time: usually the largest and most hidden cost, covering data clean-up, training and pilot reviews.
• Data clean-up: a one-off effort that pays back across every future use case.

During pilots, prefer monthly plans and free tiers over annual commitments, and check whether tools you already pay for include AI features. Compare plans on the pricing page or any vendor's equivalent with your three buckets in mind.

How national AI strategies shape the environment

Both Saudi Arabia and Egypt treat AI as a national priority, which affects the tools, talent and rules available to you. In Saudi Arabia, the Saudi Data and AI Authority (SDAIA) leads the national data and AI agenda and oversees personal data protection, and has published AI ethics principles. In Egypt, the National Council for Artificial Intelligence and the Ministry of Communications and Information Technology lead the National AI Strategy, now in its second edition covering 2025–2030.

For an SME, the practical effects are more Arabic-capable tools, more training programmes, and clearer expectations on data handling. Following official guidance from the start is easier than retrofitting later.

Where an AI-native ERP fits

Steps 2 to 4 are easier when your data and your AI assistant live in the same place. Dopave is one option: an Arabic-first, AI-native cloud ERP for Saudi Arabia, Egypt and the Gulf, covering accounting with ZATCA Phase 2 and Egypt ETA e-invoicing, CRM, inventory, POS, HR and payroll and more. Its AI copilot can create leads, invoices and tasks from Arabic or English text or voice, answer questions about your business data, and acts only within each user's permissions. You can read how this works in AI ERP explained, or start a free plan for up to 3 members, with a 14-day trial and no card required.