Short answer
Managed AI agents with SLAs means someone is accountable for keeping your agents available, correct and safe after go-live. A good service defines what is covered, how incidents are handled around the clock, how changes are tested and approved and how performance is reported to the business.
On this page
The day after go-live
Launching an AI agent is a project. Running it is a responsibility that never ends. Models are updated by their providers, source documents change, connected systems release new versions, users find edge cases and attackers try prompt injection. An agent that worked perfectly in the pilot can quietly degrade within weeks: answers become less precise, a connector times out, a cost spike goes unnoticed.
Many organizations discover at this point that nobody owns the agent operationally. The project team has moved on, IT operations does not know how to judge whether an answer is "wrong", and the business owner only notices when complaints arrive. A managed service with clear service levels closes that gap.
What a service level for AI agents should cover
Classic SLAs measure availability and response times. For agents, that is necessary but not sufficient. We agree on service levels in four dimensions:
| Dimension | What is defined |
|---|---|
| Availability | Operating hours, maintenance windows, availability of the agent and its connectors |
| Incident response | Severity levels, response and restoration targets per level, communication path |
| Quality | Agreed quality indicators such as escalation rate, human override rate and sampled answer accuracy, with thresholds that trigger review |
| Change | Lead times, testing and approval rules for changes to instructions, tools, models and data sources |
Targets are set per agent, because an internal research assistant and an agent that prepares supplier payments do not need the same level of protection.
Ownership: who does what
Clarity of roles is the heart of the service:
- We own monitoring, incident handling, root-cause analysis, technical maintenance of connectors, model and prompt updates within agreed rules, cost tracking and regular reporting.
- Your business owner owns the agent's mandate, acceptance of quality thresholds, approval of changes that alter what the agent does and decisions on new use cases.
- Your IT owns identity, network access, the connected systems and their release schedules, with a defined interface to our team.
- Your compliance and data protection functions receive the evidence they need and approve changes affecting personal data or risk classification.
This split is written down in a responsibility matrix that is part of the service agreement.
On-call and incidents
Agents fail in unusual ways, so incident handling needs people who understand both the infrastructure and the agent's behaviour. jocapps GmbH offers 24/7 call support in English and German (+49 761 519 751 50). Typical incident types include connector failures, degraded answer quality after a model update, unexpected cost increases, suspected prompt injection and data appearing where it should not. For each, a runbook defines first steps, including the option to pause an agent or switch it to "draft only" mode, where every output requires human approval. After each significant incident we deliver a short post-incident review with causes and measures.
Change management without slowing down
Agents need frequent small improvements. A heavy change board for each prompt adjustment would kill that. We use tiers: minor wording changes are tested against a regression set and deployed with notification; changes to tools, permissions, models or data sources require testing plus business approval; changes that affect risk classification go through a fresh assessment. Every change is versioned so behaviour can be traced back to a specific configuration.
Reporting the business understands
Monthly reports show volumes handled, escalations, human overrides, incidents, changes, costs and open improvement items, plus a short commentary on trends. Business owners see whether the agent is delivering on its mandate; IT sees stability; compliance sees evidence. Detailed monitoring setups are described in AI agent monitoring and observability.
Platform and engineering behind the service
Agents in our managed service run on kalAIDoCreate, with governed agents, audit trails, SSO, integrations into SAP, CRM, SharePoint and Slack and EU data residency. For custom engineering around the agents, such as new connectors, apps or internal tools, OnlineAppCreator provides done-for-you development by AI Code Engineers for B2B and B2G clients with SLAs. That means one partner for operation and further development.
Our managed AI agents service can take over agents we built through AI agent development as well as agents built by your own team, after an onboarding review. If your agent is still a pilot, start with AI agent pilot to production.
FAQ
Can you operate AI agents that you did not build?
Yes. We start with an onboarding review of architecture, permissions, logging and documentation, close the gaps and then take over operations under an agreed SLA.
What happens if an agent behaves incorrectly at night?
On-call staff follow the runbook, which can include pausing the agent or switching it to a mode where every output needs human approval, before investigating the cause.
Do quality metrics really belong in an SLA?
For agents, yes. Availability alone does not show whether answers are still correct. Agreed quality indicators with review thresholds make degradation visible early.
Who approves changes to the agent?
Minor changes follow agreed test rules. Changes to what the agent does, its permissions or data are approved by your business owner.