Custom AI Solutions · 8 min read

Before an AI agent touches your company: scope, residency, approval

A chatbot answers; an agent acts. Before an AI agent acts inside a Saudi company, three things must be written down: what it may touch, where its data lives, and who approved it. Here is the pattern, in plain language, with the Saudi rules that now sit behind each one.

Octopus Mena · 13 September 2026

ShareX LinkedIn

A glowing central node with eight luminous arms reaching outward, an agent connected to every system it is allowed to touch

Key takeaways

A chatbot answers; an agent acts. Before an agent acts inside your company, write down its scope, its data residency, and who approved it.
Scope is three columns (read, write, never); residency is three data tiers; approval is a named owner plus the line where a human checks.
The PDPL, its transfer regulation, and SDAIA's new accreditation certificates (two years, licensed assessors) are the Saudi references behind each part.

Contents

Scope: what the agent may touch
Data residency: where the agent's data lives
Approval: who said yes
Put together, it fits on one page
Frequently asked questions

A chatbot answers. An agent acts. That is the whole difference, and it changes the question you should be asking.

For the last two years the question in most Saudi boardrooms was "which AI tool should we use?" Agents move that question. An AI agent is software that can carry out a task on its own: read the inbox, draft the reply, update the CRM, raise the invoice, send it, without a person touching each step. That is the value. It is also why "which tool" is no longer the right question. The right question is about permissions: what is this thing allowed to do inside my company, and on whose say-so?

We answer that with three words, in this order: scope, residency, approval. None of them is technical. All three are management decisions, and each one now has a Saudi reference sitting behind it.

Scope: what the agent may touch

Scope is the written list of what an agent can see, what it can change, and what it can never do. It sounds obvious. Most of the AI agent deployments we come across do not have one. Someone connected the agent to "everything" because that was easiest, and nobody wrote down what "everything" meant.

Think of a sales agent. It may read the enquiry inbox and the product sheets. It may draft replies and prepare quotes. It may not send a quote above a set value without a person approving it. It may never touch contracts or payment details. Or a finance agent: it reads incoming invoices, matches them to purchase orders, flags what does not match, and never moves money.

The principle is the same one you use for people. You do not hand a new hire the company chequebook on day one. You give them the smallest job that is useful, and you widen it as they earn trust. An agent is no different, except that it never gets tired and never asks whether it should. So the limits have to be yours.

Practically: scope lives in a document a manager can read, not only in the agent's code. Three columns are enough: read, write, never. Review it every quarter, because agents get extended and nobody remembers to update the page.

Data residency: where the agent's data lives

Residency is where your data sits and where the model does its thinking. This is the part most people skip, because it happens out of sight. When an agent reads a file, that file goes somewhere to be processed. If that somewhere is a server outside the Kingdom, you have just made a data transfer, whether you meant to or not.

The Saudi reference here is the Personal Data Protection Law (PDPL), fully in force since September 2024, with penalties up to SAR 5 million. It comes with its own regulation for transferring personal data outside the Kingdom. Personal data means information about people: your customers, your employees, your patients, your tenants. Moving it abroad for processing is regulated, not forbidden. It needs a legal basis and it needs care, and "the agent did it automatically" is not a legal basis.

Here is the part that is changing. Until now, keeping data in the Kingdom often meant running smaller models on your own servers and accepting less capability. That trade-off is about to loosen. In-Kingdom compute from the large global cloud providers is dated for late 2026, with at least one major region announced for November 2026, on top of the national AI capacity the Kingdom is building. Residency is becoming a choice you can make without giving up much.

The pattern we use is three tiers. Tier one, public or non-sensitive data: any capable model, anywhere. Tier two, confidential business data: models running in the Kingdom, or data encrypted before anything reaches a model. Tier three, personal data and secrets: in the Kingdom or on your own infrastructure, full stop. Write the tier next to each item in the agent's scope. Now scope and residency are one page.

Approval: who said yes

The approval layer means two things, and companies tend to think of only one.

The first is who signed off. A named person in your company approved this agent, its scope, and its residency tier, and that person reviews it. Not "IT set it up." A name.

The second is what a human checks before an output counts. Money, contracts, anything customer-facing, anything that touches personal data: a person looks before it goes. Everything below that line, the agent runs unattended. The line is where the cost of a mistake becomes high. You draw it, and you move it as trust builds.

The Saudi reference here is the pair of words the PDPL is built on: controller and processor. The controller decides why and how personal data is processed. That is you, the company. The processor handles the data on your behalf: a vendor, a cloud, or the partner running your agent. SDAIA, the authority responsible for the PDPL (and the author of the national AI Adoption Framework we explained for executives), has now published rules for accreditation certificates for controllers and processors. An independent assessor, licensed by SDAIA and accredited by the Saudi Accreditation Center (SAAC), checks that your data practices conform to the PDPL and its regulations, and the certificate runs for two years. SDAIA has also set the rules for licensing those assessors. In plain terms: "are we compliant?" is moving from a feeling to a document.

What that means for agents is direct. An agent that processes personal data is processing it on your behalf. You are the controller. If a partner runs the agent, they are a processor, and your contract with them should say so. Your scope document, your residency tier, and your named approver are exactly what an assessor will ask to see. Companies that wrote them down before the first agent went live will find that conversation short.

Put together, it fits on one page

Agent name and purpose. Scope: read, write, never. Data tier and where it runs. Owner. Reviewer. Which outputs a human checks. Next review date.

If you cannot fill that page, the agent is not ready, whatever the demo looked like. If you can, most of the hard part is done, and the agent can run for months without anyone lying awake over it.

This is the discipline we build agents on at Octopus Mena. It is not paperwork for its own sake. It is what lets us put AI to work inside your company and keep it running, and "keep it running" starts with scope, residency, and approval written down before day one.

We run the same rules on our own agents. The roles, the job ticket, the reserved list, and what the first runs got wrong are all in the Octopus case study.

Once the page is written, the next question is which job to give the agent first. What AI agents can actually run in a Saudi company today goes job by job: what runs alone, where a person approves, and what has to be connected.

If you want to know where your company stands before any agent touches it, the readiness assessment takes minutes.

Frequently asked questions

What permissions should an AI agent have inside a company?

The smallest set that makes it useful, written down in three columns: what it may read, what it may write or change, and what it may never touch. Widen the list as the agent earns trust, the way you would with a new hire, and review it every quarter, because agents get extended and nobody updates the page. That written scope, plus a named owner, is the core of AI agent governance in a Saudi company.

Does the Saudi PDPL apply to AI agents?

Yes, whenever an agent processes personal data about customers, employees, patients, or tenants. Your company is the controller, and a partner running the agent is a processor, so the contract should say so. Sending that data outside the Kingdom for processing is regulated under the PDPL's transfer regulation and needs a legal basis; "the agent did it automatically" is not one.

Can an AI agent keep company data inside Saudi Arabia?

Increasingly, yes. In-Kingdom cloud regions from the global providers are dated for late 2026, with one announced for November 2026, alongside the national AI capacity being built. Until then the practical pattern is tiers: public data anywhere, confidential business data in the Kingdom or encrypted before it reaches a model, personal data and secrets in the Kingdom or on your own infrastructure.

Our service