AI agents can take real work off operations teams in banks and government entities, but most pilots never get past the security review. They usually stall on governance questions: where the data goes, who approves what the agent does, and how its decisions are audited. This is the checklist we use to put agents to work inside the boundary and keep them there.
Choose the workflow before the model
The best first candidates for AI agents are operational workflows that consume skilled people's time: a shared mailbox that receives thousands of requests a month, a reconciliation someone rebuilds in a spreadsheet every morning, an asset or contract register that is always slightly out of date.
They share three traits: high volume, mostly known rules, and exceptions that need the judgment rule-based automation handles badly. That is where a well-governed agent earns its place: reading, classifying, matching and drafting, while your existing systems keep doing the calculations and postings they already do reliably.
They also touch sensitive data, regulated records and sometimes money, so governance has to be designed in from day one rather than added the week before go-live.
The checklist
1. Treat data residency and the PDPL as design inputs
Saudi Arabia's Personal Data Protection Law (PDPL) came into force on September 14, 2023, and the one-year transition period ended on September 14, 2024. It is enforced by the Saudi Data and Artificial Intelligence Authority (SDAIA), and transfers of personal data outside the Kingdom are governed by dedicated transfer regulations. For an AI deployment, that makes data flow a design decision to settle at the start.
Map every place data goes when an agent works: the prompt, the documents it retrieves, the embeddings in its search index, the model's output, the logs and the monitoring telemetry. Each copy inherits the classification of its source. Logs and vector indexes are the ones teams forget, and they are often where a full copy of the sensitive data ends up.
Decide early which data classes the agent may see at all. Masking or excluding fields the task does not need is cheaper than defending them later.
2. Deploy inside the boundary
For banks and government entities, the default should be that models, retrieval stores, orchestration and logs run on-premises or in your own cloud tenancy in the Kingdom, under your network zoning and your identity controls. Outbound internet access from the agent runtime should be denied by default and opened only to named, approved endpoints.
It also means the deployment can be assessed against control sets you are already audited on, such as SAMA's Cyber Security Framework or the National Cybersecurity Authority's Essential Cybersecurity Controls, rather than asking your security team to approve an external service they cannot inspect.
3. Scope each agent to a task
A general-purpose agent with broad access is the hardest thing in the building to secure. Give each agent one job, a written description of that job, and the smallest set of tools it needs to do it.
- Read access only to the systems and mailboxes the task requires, through service identities you can audit and revoke.
- An explicit allow-list of actions, such as classify, route, propose a match or draft a reply. Anything not on the list is out of scope.
- Access through your existing APIs and integration layer, not direct database connections or screen scraping that bypass your controls.
- Deterministic logic for anything deterministic. Calculations, postings and rule checks stay in the systems built for them.
Task-scoped agents are also simpler to test and to explain to an auditor, and quicker to switch off.
4. Put maker-checker on every consequential action
Banks already know how to control a capable but fallible actor: separation of duties. Apply it to agents. The agent is the maker; a named person is the checker.
Define "consequential" in writing before the pilot starts. Moving money, changing a record in a system of record, sending anything to a customer or a regulator, and closing a case usually qualify. Routing an email to a queue, or suggesting a match for a person to confirm, usually does not.
Keep the checker independent of whoever configures the agent, and make approval a deliberate act with no default button. A sustained near-100% approval rate on one type of action is evidence for widening autonomy on that action only, with the control kept in place for everything else.
Enforcement orders are a good example of a regulated workflow where this matters: an agent can propose a customer match or draft a response, and a person confirms it. Our note on automating Tanfeeth enforcement orders covers the controls that workflow needs.
5. Keep an audit trail an examiner can follow
For every decision an agent makes, you should be able to answer, months later, what it saw, what it proposed, who approved it and what happened next. That means recording:
- the input item and the context the agent retrieved;
- the model, prompt and configuration version in use;
- the proposed action and the reason given;
- the approver, the decision and the timestamp;
- the result in the target system.
Store it in tamper-evident form, retain it like the business records it relates to, and protect it like the underlying data, because it is sensitive data too.
6. Measure against a baseline before day one
Many AI business cases fall apart because nobody measured the "before." Capture the baseline while the process is still manual: volumes, handling time per item, backlog age, misrouting and rework rates, SLA breaches, and the hours of skilled time consumed.
Agree success thresholds and failure conditions with the business owner and finance up front. Then measure the agent in production-shadow on real traffic instead of a curated test set. The pilot is finished when its numbers are clear enough for your CFO to support a decision either way.
7. Design for where AI regulation is heading
AI governance in the Kingdom has so far been principle-led. SDAIA has published AI Ethics Principles and, in 2026, a National AI Risk Management Framework. In April 2026 it also opened a public consultation on a draft Responsible AI Policy that would introduce risk tiers and binding obligations, including human oversight and record-keeping across the AI lifecycle.
The details may change before anything is final, but the direction is clear: an organization already running task-scoped agents with human approval, documented risk decisions and full audit trails will not need to retrofit them.
A 30-day pilot shape
With disciplined scope, thirty days is enough to prove or disprove value on one workflow:
- Week 1: choose and baseline. Pick one workflow with a clear business owner. Capture the baseline, map the data flows, and start the security and data-protection review in the same week.
- Week 2: deploy in shadow. Stand up the agent inside your boundary with read-only access. It processes live traffic and records what it would have done, while people keep doing the work.
- Week 3: switch on maker-checker. Let the agent propose actions on the lower-risk categories, with every proposal approved or rejected by a checker. Tune it against the rejections.
- Week 4: measure and decide. Compare against the baseline, cost the run rate, and make a go/no-go decision with the business owner, security and finance in the room.
The output is a decision, backed by numbers from your own environment, about whether to take that workflow to production and which one comes next.
The short version
- Data flows mapped, including prompts, retrieval stores and logs, with PDPL implications reviewed
- Models, data and logs deployed on-premises or in your own in-Kingdom tenancy
- One agent, one task, least-privilege access through governed APIs
- Maker-checker on every consequential action, defined in writing
- Tamper-evident audit trail of inputs, versions, proposals and approvals
- Baseline captured and success criteria agreed before day one
- Controls aligned with SDAIA's principles and the direction of its draft Responsible AI Policy
Two of our products apply this checklist to specific workflows: MailTriage for shared operational mailboxes, and AssetSense for asset registers. Both run inside your boundary, with a person approving each consequential action.
If you have a mailbox, a reconciliation or a register that consumes more people than it should, book an executive briefing. We will look at the workflow with you and give you a written view on whether it is a good first candidate and what a pilot would need to pass your security review.