Managed AI Platforms & LLMOps Keep AI platforms healthy, measured and within budget

It depends on the scope you choose, and both are written into the contract. Platform operations cover monitoring, patching and updates, backup and recovery drills, access reviews, incident handling, capacity and cost reporting, and a regression check on every change before release. With evaluation and improvement added, we also extend the regression sets with your reviewers and propose and test changes that improve quality or lower cost. The contract sets support hours, priorities and escalation for each platform, and cover outside working hours is included only where the contract provides for it. Your business applications stay with the teams that own them: we run the AI platform they use, not the applications themselves.

No. LLMOps means operating platforms built on large language models: the models, prompts, indexes, evaluation, capacity and cost. AIOps means using AI to run IT operations, for example to correlate alerts or spot anomalies in infrastructure monitoring, and it is not part of this service. Monitoring the availability and performance of your networks, servers and links belongs to our network operations centre (NOC) service, and day-to-day IT operations to Managed IT Services. The two meet at the infrastructure under the AI platform: if the NOC already watches those servers and links, the responsibility matrix records which alerts go to which team.

Against the regression set, in a test environment, before anything reaches users. Every change to a model, prompt, index or configuration runs on the same Arabic and English questions and tasks, built with your reviewers, and is scored against the baseline on the measures you agreed, such as grounded-answer rate, retrieval quality, refusals, output format, latency and cost per request. The results go into the change record, your named owner approves the release, and the rollback step is ready before it starts. Because a regression set only covers the cases it contains, we keep adding cases from incidents and user feedback. Larger changes, such as a new model or new tools for the system, may also call for independent testing under AI Security Assessment & Red Teaming.

It goes through the same change procedure, so the replacement is evaluated before users see it. Commercial model providers retire some versions and update others; we follow their announcements, pin versions where the provider allows it, run the candidate replacement on the regression set, adjust prompts where needed and plan the release. If the embedding model changes, the document indexes must be rebuilt, which takes capacity and is scheduled as a change of its own. Open-weight models you run locally follow the same path for new versions, and your legal team reviews any licence change. If a provider changes its terms or where it processes data, we raise it with you and your data protection officer: sending personal data to a model running outside the Kingdom is a transfer question under the Regulation on Personal Data Transfer outside the Kingdom.

By restoring the platform in drills, not by checking that backup jobs completed. We restore into an isolated environment on the schedule set in the contract, run the regression set to check that the restored platform answers as it did before, and record the time taken against the recovery objectives your organisation sets. Rebuilding a large index from the source documents can be faster or more reliable than restoring it, and the drill shows which approach meets your objectives. Each drill produces a report of what worked, what failed and what is being fixed, which you can use as evidence for business continuity and audit.

Through monthly reports by application and use case, and thresholds that alert you before a budget is reached. We measure requests, tokens or compute, GPU use and storage from the platform and the gateway, and reconcile our estimates with provider invoices, which remain the reference for commercial models. Hard spending limits are enforced at the gateway where one is in place; designing that gateway belongs to Secure AI Gateway & Guardrails. Recommendations to lower cost, such as caching answers, a smaller model for simpler tasks or scaling down idle capacity, are evaluated on the regression set first, because a cheaper option can also lower quality. For platforms you run locally, capacity forecasts from usage trends show when more hardware will be needed and why. Your organisation sets the budget, and we report against it.

You do, apart from third-party models and software, which stay under their own licences. Your data stays yours; configurations, documentation and deliverables are handed over as set in the contract. Where possible we keep configuration, prompts and policies as code in a repository you control, together with the regression sets, dashboard definitions and runbooks, so another team can take over. On exit, the exit plan agreed in the contract covers knowledge transfer to your team or another provider, and the accounts we used are removed or have their credentials changed. If your organisation is within ECC scope, ECC-2:2024 requires its contracts with third parties, managed services included, to provide for secure removal of its data at the end of the service, and we write that step into the exit plan.

Through a joint procedure agreed during takeover: we handle the AI platform, and your teams keep their roles. Platform faults are handled from the runbooks and escalated as the contract sets out; faults in the infrastructure underneath, such as networks, servers or storage, go to whoever runs it, whether your IT team, the NOC or Managed IT Services. Security events, such as suspected prompt injection, possible data exposure, unusual access or a leaked key, are sent to your SOC through your SIEM, and we carry out the containment steps on the platform that your SOC directs, such as revoking a key, rolling back a change or taking a feature offline. Where personal data may be affected, the decision to notify under PDPL rests with your organisation and its data protection officer; we provide the facts and the logs. This service does not replace your SOC.