Skip to main content
CerebraTech AI
/services/managed-ai · MA

Keep an accepted AI system observable, recoverable and ready for change

Maintenance Agreement (MA) is a scoped operating service after handover. We agree what is monitored, who approves access and changes, how incidents are handled, and what evidence is reviewed — without treating maintenance as a model accuracy warranty.

What MA covers

The final scope depends on the system inventory, support window, site and contract. These are the workstreams to confirm during an MA assessment.

System operations

Health checks for service, container, runtime, device, storage, compute, temperature, network, queues and integrations.

Inventory, health signal and incident record

Model operations

Versioning, release notes, drift signals and retraining triggers when data, labels and acceptance criteria exist.

Model change note and evaluation result

Security and continuity

Patch and dependency review, configuration backup, restore or rollback checks, and a recovery path agreed to the architecture.

Approval, backup/restore check and recovery note

Customer enablement

Runbook, escalation matrix, review cadence and training for the named customer owner and decision maker.

Health report, decisions and improvement backlog

The MA operating lifecycle

  1. 01

    Baseline

    Record versions, devices, model, metrics, dependencies and ownership before support starts.

    System inventory and baseline report

  2. 02

    Monitor

    Watch agreed health, latency, queue, storage, integration and drift signals within the data boundary.

    Dashboard/alert and incident record

  3. 03

    Review

    Review monthly or quarterly with the system owner; separate measured facts from open assumptions.

    Health report and risk list

  4. 04

    Change

    Patch, tune or update only through an approved change request with a test and rollback plan.

    Release note, test result and approval

  5. 05

    Recover

    Use the agreed backup, rollback or spare-device path when an incident requires it.

    Recovery evidence and post-incident note

  6. 06

    Improve

    Prioritise the next safe improvement from evidence, operator feedback and an accountable owner.

    Backlog and target review date

Managed AI Operations is an optional scope

Managed AI Operations can add monitoring, review coordination and operational reporting when the team, access model and agreement support it. It is not implied by having an MA, and it does not transfer business decisions or model accountability to us.

  • Monitoring and alert triage within the agreed boundary
  • Recurring health review and risk/backlog facilitation
  • Release, dependency and compatibility coordination
  • Incident communications and recovery exercise support

Data boundary and remote access

The default is to keep business payload inside the customer environment. The exact telemetry and access path are confirmed in the architecture and contract.

Data boundary and remote access
SurfaceDefault handlingApproval / evidence
Business payloadImages, documents, prompts and workflow payload stay on the customer network unless an explicit contract says otherwise.Architecture decision and customer data owner approval
Health telemetryOnly agreed health metrics may leave the site; monitoring can remain inside the customer environment.Allow-list, retention and monthly review
Remote support sessionNo standing access. Each session has an approver, purpose, expiry and least-privilege path.Access approval and audit log
Backup and logsLocation, retention, encryption and restore path follow the customer architecture and contract.Restore check and handover inventory
Model trainingMA does not grant permission to train CerebraTech AI models on customer data.Separate written agreement if any data use is proposed

Example support structure — confirm values in the agreement

This is a template for a scope workshop, not a guaranteed package. Response means acknowledgement and triage; restore and final resolution depend on customer access, hardware, data and third parties.

Example support structure — confirm values in the agreement
SeverityExampleResponse targetRestore / resolution
S1Production workflow stopped or safety-critical integration unavailableTo be agreedRTO/RPO or architecture exception to be agreed
S2Material degradation with a workaround availableTo be agreedTarget and update cadence to be agreed
S3Limited function, report issue or non-urgent defectTo be agreedPlanned change or next review
S4Question, request or improvement backlog itemTo be agreedScoped change request if approved
Service window
Business hours, extended hours or 24×7 — contract value required
Timezone
Asia/Bangkok (ICT), unless the agreement specifies another clock
Channels
Ticket first; phone, chat or on-site only when included in scope
Customer dependencies
Named owner, access approver, network, spare/device plan, data labels and third-party support

Sample monthly AI health report

Template only — the following is a report structure, not a customer result or benchmark.

Executive summary

Green / Amber / Red status, notable changes and decisions needed

System health

Availability, latency, queue/failure, storage and resource signals that were actually measured

Model and data signals

Drift proxy, human review and evaluation context when labels and method exist

Incidents and changes

Response, restore, root-cause notes, patches, approvals and rollback readiness

Risk and next steps

Customer decisions, owner, due date and improvement backlog

Handover and offboarding are part of the service

The agreement records what each side owns. At handover or exit, access and artefacts are returned or revoked according to the contract; this is not a promise of zero lock-in.

Handover and offboarding are part of the service
AreaCustomer ownerService responsibilityHandover / exit evidence
Business data and decisionsData owner and operational approverProcess only within the agreed boundaryData map and decision record
Code, model and configurationAccept versions and change decisionsMaintain only listed components and dependenciesVersion inventory and release notes
Access and operationsApprove sessions and provide site dependenciesUse least privilege and keep an audit trailAccess log, revocation and runbook
MA closureConfirm acceptance and receiving ownerDeliver agreed inventory/log export and remove our accessOffboarding checklist and outstanding-risk note

What MA does not promise

  • Model accuracy without a dataset, method and acceptance criteria
  • New features, retraining or hardware changes without a scoped change request
  • Network or third-party cloud service responsibility unless the contract names it
  • A hardware warranty, zero downtime or zero lock-in
  • Permission to use customer payload for CerebraTech AI model training

Start with a support-scope workshop

Tell us what is already in production, what can stop, what evidence exists and who can approve access. The next step is an MA assessment or support-scope workshop — not an automatic quote.

Request an MA assessment

This site uses very few cookies

We use only the cookies necessary to remember your language choice and save this consent preference (your chosen theme is remembered via browser local storage, not a cookie). Our analytics tool uses no cookies and collects no personal data.

Read the cookie policy