Delivery
Operational AI Implementation in Saudi Arabia: What Must Be True Before You Accept the Handover
The handover terms an operations owner should insist on before an AI system touches live records, written from the side that has to run it.
5 sources · Evidence review
In this guide6 sections
- Who signs for the system, and what exactly are they accepting?
- Which records may it change, and where do unusual requests go?
- What does the log show when a write times out?
- Can you roll it back, and what will rollback not undo?
- Are the five data terms in writing before the first live record?
- What should be in the handover packet your team receives?
- Sources
- Frequently asked
Operational AI implementation is finished when your own team can run the system on an ordinary Sunday morning without the builder in the room, not when a URL goes live. In our Custom AI Implementation, Handover is the last of five stages, after Engineer, Integrate, Test and Deploy, and it starts only once the AI Solution Blueprint's acceptance tests pass. This guide is written from the receiving side: what has to be true about the owner, permissions, data terms, logs, rollback and runbooks before you sign. The fictional example stages a service-request classification for an existing business system; it does not describe Ting's production stack or a delivered client implementation.
Who signs for the system, and what exactly are they accepting?
Name one operating owner in the business, a person rather than a department, who is accountable for what the system does each working day, plus a technical owner who maintains it, and write both names into the acceptance note before go-live. Then freeze what is being accepted: the permitted task, the excluded actions, the evaluated version and the exact cases that passed or failed, with unresolved defects left visible instead of folded into a later phase. NIST's AI Risk Management Framework covers design, development, use and evaluation as voluntary guidance; it does not replace your own operating decision or a Saudi legal review.1
Which records may it change, and where do unusual requests go?
In the proposed request classifier, the model returns an allowed queue and a short explanation; a deterministic validator checks the schema and the queue list, and a separate authorized integration performs any write. Credentials and write permissions stay outside the generated text, so an instruction inside a customer message cannot create a new permission or widen the destination-system scope. OWASP's LLM01:2025 guidance recommends giving the application its own API tokens and handling privileged functions in code rather than providing them to the model, and its LLM06:2025 entry lists permissions on downstream systems beyond the intended operation as a cause of excessive agency.23 Anything the validator cannot place, such as an unknown queue, a restricted account or a customer asking for an exception, goes to a named supervisor's queue instead of being guessed. Treat this as a design requirement to test, not a claim that a prompt alone enforces security.
What does the log show when a write times out?
Every action should leave a timestamped entry that shows what was proposed, what passed validation, what was submitted, what the destination confirmed and what triggered it. Use explicit states such as proposed, validated, submitted, confirmed and outcome_unknown in the proposed workflow. A timeout after submission does not establish whether a change occurred, so it is logged as outcome_unknown, never as a success or a failure. Before another attempt, inspect the destination's supported status or deduplication mechanism, and never invent a success receipt. Define duplicate handling against the real system contract; repeated extraction or classification must not automatically create repeated business operations.
Can you roll it back, and what will rollback not undo?
Fictional drill: request REQ-DEMO-12 was submitted, the response timed out and an operator disables new writes. Expected behavior: retain its unknown outcome, keep read-only diagnostics available and reconcile it against the destination before a retry. The implementation standard is that any change can be rolled back, yet rolling back application code is not proof that an already-written record was undone. Before you accept, watch the switch that stops new writes being used once, and get a written list of which changes are reversible, which need a compensating action in the destination system and who is authorized to perform it.
ExhibitRollback drill
What happens when a write times out? An operator stops new writes, and the work order keeps its unknown outcome until it is checked against the maintenance system.

Rollback and reconciliation drillRasikh Industries · WO-1047
- ProposedThe model proposes an allowed queue
proposed - ValidatedA deterministic check passes
validated - SubmittedThe work order is sent to the maintenance system
submitted - Outcome unknownThe response times out; new writes are disabled
outcome_unknown - ReconciledChecked against the maintenance system before any retry
reconciled
New writesDisabled
- Rolling back code does not undo a record that was already written.
- Diagnostics stay available, read-only.
Illustrative exampleRasikh Holding is fictional
The decisionHand over only when this drill runs end to end, with a named owner for every reversible and compensating action.
Are the five data terms in writing before the first live record?
Five data terms should be settled in writing before the build and checked again at handover: what data the system reads, where it is processed and stored, who can access it, what is logged and how long it is kept. The log itself needs correlation IDs, release versions, failed validations, unknown outcomes, duplicate attempts and resolution times, without unnecessary customer content; Saudi Arabia's Personal Data Protection Law limits collected personal data to the minimum necessary for its purpose (Article 11).4 Name who receives each alert and what they may do about it. OpenAI's evaluation guide recommends continuous evaluation as the application changes, so rerun a regression set whenever models, rules, prompts, data sources or integration mappings change; a healthy server alone does not prove correct business behavior.5
What should be in the handover packet your team receives?
The packet should hold the release scope, system boundaries, access owners, the data terms, evaluation evidence, alert routes, runbooks for the common failures, the recovery procedure and known limitations. Knowledge transfer means your operators run a safe failure drill from the runbook themselves while the builder watches, not a walkthrough of slides. Check that hosting, model usage and software subscriptions sit in accounts your company controls, as scoped in the Blueprint, and that support hours and escalation terms are written down rather than assumed. If your team cannot stop writes, identify an unknown outcome or reach the responsible owner, record the handover as incomplete.
Practical decision tool
Four handover checks to run before you sign
- Invalid queue → block the write, keep the failed check in the log and send the request to the supervisor's queue.
- Duplicate request → apply the agreed duplicate policy without an extra business action.
- Unknown submission outcome → reconcile with the destination before any retry.
- Disable switch → stop new writes while diagnostics and the recovery owner stay available.
These are proposed controls and an unexecuted fictional drill, not a penetration test, certification, SLA or client result.
Sources
Evidence review ·
Primary sources: regulators, standards and official documentation
- AI Risk Management Framework: voluntary use and scopeNISTRetrieved: November 6, 2026
- LLM01:2025 Prompt InjectionOWASP Gen AI Security ProjectRetrieved: November 6, 2026
- LLM06:2025 Excessive AgencyOWASP Gen AI Security ProjectRetrieved: November 6, 2026
- Personal Data Protection Law (amended, English text)SDAIARetrieved: November 6, 2026Official Saudi government source
- Evaluation best practices: objectives, datasets and continuous evaluationOpenAIRetrieved: November 6, 2026
Frequently asked
When can we accept the handover of an AI system?
When the agreed scope, evaluation evidence, permissions, data terms, monitoring, recovery and named owners have been shown working by your own team, not simply when a URL is live.
Can the model decide which records it may edit?
In this proposed design, no. A separate authorization layer enforces the allowed actions independently of generated text, and anything outside them goes to a supervisor.
What should happen after a submission timeout?
Log an unknown outcome with its timestamp, then use the destination's supported status or reconciliation process before deciding whether a retry is safe.
Does rollback undo every completed transaction?
No. Reversing code and reversing business data are separate procedures. Get the supported compensating actions, and who may perform them, in writing.
Related guidance
Reviewed on November 7, 2026



