An AI agent earns its keep in a Saudi business on exactly one kind of work: a multi-step, repetitive process a person can still check at the end. Vendor-pitch autonomy is decoration until you have named the process, and that choice decides whether the project returns hours or burns them. Here is what an agent does inside your systems, a three-question gate for picking the first process to hand over, the PDPL lines it cannot cross, and a ninety-day path to a measured result.
What an AI agent does once it is inside your systems
A chatbot answers. An agent finishes. You give it a goal in plain language; it plans the steps, reads and writes inside the systems you connected, and stops when the work is done or when it hits a case it was told to escalate. The difference from every AI tool you have bought is the actions: an agent does not describe the work, it performs it.
Three parts of the design decide whether you have a business tool or a science project:
- Tools. The systems it may touch, your ERP, your ticketing system, your email, each through a dedicated role rather than a shared admin login.
- Limits. Hard rules it cannot talk its way past: a spending cap, a fixed list of permitted actions, a ban on anything irreversible.
- Escalation. The defined moment it stops and hands the case to a human, with a summary of what it did and why.
Miss one of the three and you do not have an employee that never sleeps; you have a liability with a demo.
Which process should get your first AI agent
Most teams pick the agent before the process; that ordering is why pilots die. Run every candidate through the same gate:
flowchart TD
A(["A candidate process"]) --> B{"Multiple steps across systems?"}
B -->|No| Z(["Automate it or leave it alone"])
B -->|Yes| C{"Repeats often enough to matter?"}
C -->|No| Z
C -->|Yes| D{"Can a person check the exceptions?"}
D -->|No| Y(["Keep a human on it"])
D -->|Yes| E(["Scope the pilot here"])
- Multiple steps across systems. The value is in stitching work that lives in email, spreadsheets and an ERP nobody enjoys. A single-step task needs a script, not an agent.
- Repeats often enough to matter. In our experience a process that runs a few times a month rarely justifies the build plus the supervision; weekly or daily is where the math works.
- A person can check the exceptions. If every case is high-stakes, an agent adds review work instead of removing it. The best first targets fail loudly and cheaply.
If several processes clear the gate, our ranking of what to automate first breaks the tie.
Where agents already work in Saudi companies
AI is projected to contribute about $135bn to Saudi Arabia’s economy by 2030, the largest gain in the Middle East (PwC, via US ITA), and agents are one place that value shows up: hours returned from work nobody was hired to do. The processes where we see them working clear the gate above.
- Customer requests that span systems. From “where is my order” through to resolution: find the shipment, update the customer in Arabic or English, issue the return, escalate the angry ones. See our piece on AI customer service in Arabic and English.
- Government tender monitoring. An agent watches Etimad for tenders matching your trades, extracts requirements and deadlines into a checklist, and drafts a first response pack for a bid manager to shape.
- Incoming invoices. Read the invoice, match it to the purchase order, check the VAT number against the vendor record, post the clean ones, and flag the mismatches with a one-line reason.
- Coordination after a signature. A deal or hire closes: the agent creates the accounts, books the sessions, orders the kit, and reports what is waiting on whom.
A worked example: the tender-watching agent
Take a Riyadh contractor that bids on public work through Etimad. The numbers are made up to show the shape. A bid manager spends about five hours a week scanning notices and matching them to the company’s trades, and about ten hours per bid assembling the first-draft compliance pack of requirements, deadlines and forms. At six bids a quarter that is roughly 125 hours a quarter of a senior person’s time.
An agent takes both halves, with a person approving everything it produces. It shortlists the relevant notices, extracts their requirements into a checklist, and drafts the pack. The manager reviews the shortlist in half an hour a week and spends about three hours shaping each pack instead of ten: about 25 hours of human time a quarter against 125, a gross saving of around 100 hours before build, run cost and review.
The lesson is not the number but the shape of the decision: the agent was given drafting work, never submission work, which is why one manager could supervise it without becoming its bottleneck.
PDPL lines your agent cannot cross
An agent reads customer records, so it inherits every obligation attached to them. The one that bites hardest in the Kingdom is route: an agent that sends personal data to a model hosted outside Saudi Arabia is making a cross-border transfer, and under the PDPL cross-border transfer is governed by Article 29 and SDAIA’s Transfer Regulations (trade.gov, 2024). Its data path needs the same review as any system touching customer data, and the answer is often to keep personal data in-Kingdom and send the model only what the task needs.
Three habits keep it inside the lines:
- Map the data path before the pilot. Where does each field live, and where does it go when the model processes it? Write it down; compliance will ask for it first.
- Minimum data per task. An agent matching invoices does not need customer contact history. Scope what each task reads as tightly as what it may do.
- Keep the log data-aware. Action logs are essential, but they must not become a second copy of the personal data you worked to protect.
Where AI agent projects go wrong
Gartner predicted that at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs, or unclear business value (Gartner, 2024). All four hit agents harder than chatbots, because an agent acts on the bad data and needs the risk controls a chatbot could skip. The traps we keep seeing:
- The shared admin login. The agent gets root because scoping a dedicated role took an afternoon nobody had, and from then on one wrong conclusion becomes a wrong action everywhere it can reach. Broadening permissions is a decision, never a convenience.
- Compounding errors. A chatbot gives you one wrong answer. An agent gives you a wrong record, then an email built on it, then a follow-up built on that. Log every action and make the log readable by a non-engineer, or you find out a month late.
- Prompts smuggled in through the documents. Anything the agent reads from outside, emails, PDFs, web pages, can carry text that tries to steer it. Treat external content as untrusted, and never let an irreversible action depend on it alone.
- A demo with clean inputs. Ask what happens when the invoice is a photo of a screen, the vendor’s name is spelled two ways, or the customer writes in mixed Arabic and English. If the answer is a pause, the demo was the product.
How to start: ninety days to a measured agent
The path is boring on purpose, and matches the first 90 days of any AI programme:
flowchart LR
A(["Measure the baseline in hours"]) --> B(["Shadow mode"])
B --> C(["Human approves every action"])
C --> D{"Clean log after a month?"}
D -->|No| B
D -->|Yes| E(["Automate the clean cases within hard limits"])
E --> F(["Measure net saving, then scale or stop"])
- Measure first. A real week of timestamps on the candidate process. No baseline, no project: you cannot tell saving from noise later.
- Run shadow mode. The agent works the real cases next to the human; its output is compared but never actioned, and you learn its failure modes at zero risk.
- Promote slowly. Once a month of logs is clean, let the clean cases run on their own inside hard limits: fixed tools, a spending cap, irreversible actions still gated. This staged promotion is how we scope our AI consulting engagements.
- Judge on net hours. Gross saving minus build, run and review. If it pays, give the agent the next process through the gate. If not, you bought knowledge cheaply instead of shelfware.
Have a process in mind that an agent could take off your team’s plate? SDCG scopes and builds business agents with hard limits, readable logs and PDPL-aware data paths, and because we are independent we will tell you when an agent is the wrong answer. Book a free 30-minute review.