AI Managed Services
Keep your AI agents running at their best.
Monitoring, maintenance, security and continuous optimisation for the AI systems your business relies on.
Building an AI agent is only the beginning. We help keep your AI systems reliable, secure, observable and effective.
AI application
Agent, assistant or automation
Monitoring & evaluation
Metrics, tests and quality
Incident management
Investigate, fix, roll back
Optimization
Quality, speed and cost
Reliable operations
Predictable business processes
What we do once your AI system is live
AI Managed Services is a monthly service. We keep AI systems running and improve them, whether we built them or another team did.
Business continuity
Catch errors, slowdowns and failed workflows early, before customers feel them.
Controlled change
Every change to prompts, models or integrations goes through testing, approval and a rollback plan.
Measurable quality
Test sets and regression checks show whether answers are getting better or worse.
Security and cost control
Permissions, keys, audit logs and budgets under watch — not a surprise at the end of the month.
How it differs from other services
- One-time AI development
- Development ends at launch. We take over from there: monitoring, quality, security and change.
- Ordinary website maintenance
- A website can run perfectly while the agent gives wrong answers. We watch AI behaviour, not just the server.
- SEO retainers
- The focus is not search visibility but the reliability, quality and cost of your AI systems.
- Chatbot installation
- We do not add yet another tool — we look after the ones you already run, with clear procedures and reporting.
Who this service is for
Businesses that already rely on AI in their daily operations.
E-commerce merchants running AI support agents
Assistants that answer questions about orders, delivery and products.
Small businesses automating operations with LLMs
Repetitive operations handled by language models.
Agencies operating customer-facing AI tools
Teams that deliver AI solutions to clients and need to keep them running.
Companies using AI document processing
Invoices, contracts and requests recognized and processed by AI.
Internal knowledge assistants and RAG systems
Knowledge bases where source quality determines answer quality.
Workflows connected to CRM, ERP or helpdesk
Automations wired into business systems, where mistakes have consequences.
Companies with an agent but no in-house expertise
The solution runs, but nobody is there to monitor, maintain and improve it.
Problems we solve
An AI agent can stop working correctly even while your website and servers are still up.
These problems rarely look like an outage. They usually show up as slower, more expensive or less accurate answers. The most common causes:
Third-party API changes
A provider changes a format, limit or version, and the workflow stops.
Authentication and tokens
An expired key or revoked access breaks an integration without a visible error.
Model changes
A new model version responds differently to the same instructions.
Prompt and workflow regressions
A small change to a prompt or configuration degrades the results.
Tool invocation errors
The agent passes wrong parameters or calls an unsuitable tool.
Knowledge base changes
Outdated or moved documents lead to incomplete or wrong answers.
Unexpected cost increases
More requests, longer prompts or an unsuitable model drive up the bill.
Slow responses
Latency in the model, the tools or the retrieval step.
Retrieval quality degradation
The RAG system finds less relevant sources than it used to.
Incorrect generated output
Confident but wrong answers that reach customers or staff.
Security issues
Prompt-injection attempts, overly broad permissions or data leakage.
Infrastructure failures
Problems in hosting, queues or the application's dependencies.
Configuration drift
Environments diverge over time and changes are no longer traceable.
Not every AI failure can be prevented. The goal is to detect problems earlier, limit their impact and resolve them through a clear procedure.
What operating your AI systems covers
Eight areas we cover for every system we operate. Today the results are delivered through reports, reviews and joint sessions — using data from your systems and providers once access is agreed.
Monitoring and reliability
We watch availability, errors and failed runs and spot degradation before your users notice it.
- What you receive
- A reliability summary and a list of detected issues
- What your system needs
- Logs or metrics from the application and providers
Agent execution visibility
We trace the agent's steps — model calls, tool calls and the decisions between them — so it is clear why it answered or acted the way it did.
- What you receive
- Analysis of failed runs with the cause and a proposed fix
- What your system needs
- Tracing (e.g. OpenTelemetry) with masked content
Cost and token usage
We break token usage and cost down by model, agent and process and track deviations from the agreed budget.
- What you receive
- A cost report and optimisation recommendations
- What your system needs
- Provider usage reports or token data from tracing
Quality evaluation
We maintain a set of test cases and evaluate answers before and after every change to catch regressions early.
- What you receive
- Evaluation results and a comparison between versions
- What your system needs
- Approved test cases and, for RAG, access to the sources
Human review
We route uncertain or risky results to a person on your team and use their decisions to improve the system.
- What you receive
- Review rules, a case queue and a summary of decisions
- What your system needs
- Named owners and rules for when a review is needed
Incident response
We investigate, contain and fix problems, roll back to the previous version when needed and document the cause.
- What you receive
- An incident report and preventive actions
- What your system needs
- An agreed notification channel and the ability to roll back
Continuous optimisation
We improve prompts, knowledge retrieval, model choice and workflows based on evaluations and reviews.
- What you receive
- A prioritised improvement plan and the measured effect of changes
- What your system needs
- A test environment and an agreed change process
Security and governance
Least-privilege access, secrets in a vault, an activity log, retention periods and approval of changes.
- What you receive
- A risk review and a change log
- What your system needs
- A data processing agreement and clear access roles
The scope depends on the plan and on the data available. Nothing is monitored before it is actually connected and agreed.
What we monitor
What can be monitored depends on the data and integrations available in your system. We confirm the scope during the assessment.
Usually possible from the start
- Availability of the service and its public endpoints
- Request volume and error rates from existing logs
- Failed workflow executions
- Health of the external APIs the agent depends on
- Token consumption and estimated inference costs from provider reports
- Rate-limit errors
Requires an additional integration on your side
- Model and tool latency (tracing)
- Repeated failures by request type or tool
- Unexpected usage patterns
- Answer quality on real conversations, where agreed
- Budget alerts with providers
We do not show invented dashboards, and we do not claim monitoring is running before it has actually been set up and agreed.
AI agent maintenance
Ongoing technical care for the components the system depends on:
- Prompt configuration maintenance
- Model integration maintenance and version upgrades
- API compatibility checks
- Workflow dependency updates
- Knowledge source connection maintenance
- Configuration management and version control
- Deployment support
- Regression testing before every change
- Rollback planning
Ordinary maintenance does not include unlimited new feature development. New capabilities are assessed and agreed separately as a change request or a project.
Continuous improvement
- Periodic operational reviews
- Evaluation-driven prompt improvements
- Workflow optimisation recommendations
- Knowledge base quality reviews
- A monthly improvement backlog
- Performance reporting
- Change requests
- Controlled deployment and rollback
AI quality and reliability
Quality is measured, not assumed:
Test datasets
Real examples that reflect the system's tasks.
Golden test cases
Selected cases with an approved expected result.
Regression evaluation
Checks before and after every change to models, prompts or data.
Correctness and task completion
Whether the answer is right and the task gets finished.
Human review
A sample of results reviewed by a person.
Groundedness checks for RAG
Whether the answer is supported by the retrieved sources.
Tool call validation
The right tools with the right parameters.
Failure categorization
Why the system failed — for targeted improvements.
Quality trends across releases
How quality changes with each new version.
The quality of language model output cannot be guaranteed. Evaluation reveals trends and regressions; it does not rule out mistakes.
Example evaluation criteriaIllustrative examples. Real criteria are defined with you for the specific system.
Customer support AI agent
- The answer matches the current delivery and returns policy.
- The agent hands over to a person for payment questions or complaints.
- It does not promise discounts, delivery dates or refunds.
Product description generator
- Technical specifications match the data provided.
- The tone follows the brand guidelines.
- There are no claims unsupported by the product data.
Document processing assistant
- Key fields (number, date, amount, counterparty) are extracted correctly.
- Uncertain fields are flagged for review.
- Documents in an unrecognised format are never saved automatically.
Internal business knowledge assistant
- The answer cites the source it is based on.
- When information is missing, the assistant says so instead of guessing.
- Staff only see documents they are authorised to access.
Security and governance
AI security is a process, not a one-time setting:
Model output never bypasses permission checks or business rules. AI may propose; the system checks whether an action is permitted, and actions with consequences require human approval.
- API key and secret rotation procedures
- Least-privilege access
- Prompt-injection risk assessment
- Sensitive data handling
- Tool permission restrictions
- Audit logging
- Human approval gates
- Data retention policies
- Third-party provider dependency review
- Incident handling procedures
- Periodic access reviews
AI cost optimisation
We review where model spend goes and how to reduce it without hurting quality:
- Model selection analysis for each task
- Token consumption review
- Prompt length optimisation
- Caching feasibility
- Cheaper models where quality allows
- Usage and budget alerts
- Latency versus cost trade-off evaluation
- Monthly cost reporting
We do not promise a specific percentage of savings. The effect depends on the system and is measured against a baseline. Model API costs, third-party subscriptions and infrastructure are billed separately unless explicitly included in an approved quote.
Let's review your AI system
Request an AI AssessmentIncident response and recovery
When something goes wrong, we follow a procedure agreed in advance:
- 01
Classification
We determine the severity and the affected processes.
- 02
Detection and triage
We confirm the problem and collect the data.
- 03
Customer notification
We inform the contact you have designated.
- 04
Technical investigation
We look for the cause — model, data, integration or infrastructure.
- 05
Remediation
We apply the fix in a controlled way.
- 06
Rollback
When available and safer than a fix.
- 07
Post-incident review
What happened, why, and what we learned.
- 08
Preventive follow-up
Tests or monitoring that will catch the problem next time.
Response times and support hours are agreed in writing in the contract. We do not publish binding response times or availability guarantees on this website.
AI Operations Portal — in development
We are building a secure portal where you will see the state of the AI systems we operate for you. It is not yet available to clients — today you receive the same information in reports.
Overview
Planned- Active agents
- Run volume
- Success and failure rates
- Latency
- Cost
- Open incidents
Agent operations
Planned- Agent registry
- Versions and deployments
- Execution history
- Run status
- Failure analysis
Observability
Planned- Traces
- Tool calls
- Token usage
- Cost by model and provider
- Latency and errors
- Budget thresholds
AI quality
Planned- Evaluations
- Retrieval (RAG) quality
- Feedback
- Regression tracking
- Human review queues
Governance
Planned- Organisations and workspaces
- Tenant isolation
- Roles and permissions
- Audit log
- Data retention
- Secrets are never displayed
Alerts and reporting
Planned- Budget alerts
- Reliability alerts
- Scheduled reports
- Incident notifications
How it will work
- Each client sees only their own organisations and workspaces.
- Access is role-based and every action is recorded in an audit log.
- Telemetry is sent from your system with masked content; your keys never pass through the browser.
- We show only real data from connected systems — sample data only in a clearly labelled demo mode.
How we get started
Onboarding is a separate, clearly bounded phase before the monthly service:
- 01
Inquiry
You describe the system through the form — no passwords, keys or confidential data.
- 02
Call and clarification
We clarify the goals, architecture and risks.
- 03
Technical assessment
With agreed, minimal access we review the system and its instrumentation.
- 04
Proposal
We recommend a plan and scope in a written quote.
- 05
Onboarding
Baseline, monitoring, test set and procedures — after a signed contract.
Proposed service levels
Three levels with a typical scope — from periodic reviews to managed operations. Exact terms, volume and pricing are agreed in writing for each system.
The level names are working proposals, not a commercial commitment. Scope is agreed in writing for each system.
Essential Monitoring & Advisory
For one AI assistant or process that already runs and has an owner.
Periodic reviews, usage and cost visibility and recommendations — no changes made by us.
Custom quote
- Initial technical assessment
- Periodic health review from your logs and dashboards
- Usage and cost visibility from provider data
- Reporting of issues found
- Monthly summary with recommendations
Managed AI Operations
For AI used in daily operations, with several agents or processes.
Monitoring, scheduled evaluations, configuration maintenance and incident triage.
Custom quote
- Everything in Essential Monitoring & Advisory
- More detailed checks on agreed signals
- Scheduled evaluation and regression review before changes
- Maintenance of prompts, configuration and integrations within the agreed volume
- Incident triage and escalation
- Monthly cost review
Advanced AI Reliability & Optimization
For business-critical AI processes with many integrations or sensitive data.
Custom evaluations, managed changes with approval, integration health checks and an improvement plan.
Custom quote
- Everything in Managed AI Operations
- Custom evaluations for your tasks
- Managed changes with approval and a rollback plan
- Integration health checks
- In-depth cost and model reviews
- Quarterly reliability and improvement plan
- Custom support arrangements subject to contract
| Capability | Essential Monitoring & Advisory | Managed AI Operations | Advanced AI Reliability & Optimization |
|---|---|---|---|
| Initial technical assessment | Included | Included | Included |
| Monitoring (with your tools) | Periodic review | Agreed signals | Extended |
| Usage and cost visibility | Included | Included | Included |
| Quality regression evaluation | Not included | Scheduled | Custom |
| Prompt and configuration maintenance | Not included | Included | Included |
| Changes to the system | Recommendations only | With your approval | Managed, with a rollback plan |
| Incident handling | Reporting | Triage and escalation | Agreed procedure |
| Cost reviews | Monthly summary | Monthly review | In-depth |
| Integration health checks | Not included | Not included | Included |
| Response targets | Agreed per contract | Agreed per contract | Agreed per contract |
Model API costs, third-party subscriptions and infrastructure are billed separately unless explicitly included in an approved quote.
No plan includes unlimited usage, unlimited incidents, 24/7 support or a guaranteed service level (SLA) unless explicitly agreed in writing.
The levels in detail
What each level includes and excludes, what it needs from you and what is agreed commercially. Open a level to see every dimension.
Working names — not a commercial commitment. Nothing here is a price, a guaranteed service level or an availability guarantee.
Essential Monitoring & AdvisoryFor one AI assistant or process that already runs and has an owner.
- Ideal for
- One AI assistant or process in production or pilot
- An independent view without handing over operations
- Entry requirements
- A named technical owner on your side
- A written description of the system: providers, models, data sources
- Read-only access to logs, dashboards or usage reports
- Included
- Initial technical assessment
- Periodic review of health and errors
- Usage and cost summary from provider data
- Monthly report with recommendations
- Not included
- Changes to the system made by us
- Incident response beyond reporting
- New development
- Monitoring and evaluation
- Review of your existing logs and dashboards on an agreed schedule
- Checking a sample of answers against agreed criteria
- Changes and human approval
- We recommend; you decide and implement
- No change without your written decision
- Incidents and escalation
- We report issues we find through the agreed channel
- Escalation and resolution stay with your team
- Cost and usage
- Monthly usage and cost summary with its source
- Recommendations without promised savings
- Security and privacy
- Read-only, through your own access management
- No personal data where technical metrics are enough
- A data processing agreement before any production data access
- Onboarding
- Assessment of the architecture and the available data
- Agreeing review criteria and the report format
- Offboarding
- Handover of the last report and the recommendations
- Removal of our access, confirmed in writing
- Dependencies on your access and integrations
- Logs or dashboards you can grant read access to
- Usage reports from the model provider
- Agreed commercially
- Review frequency
- Number of systems in scope
- Report format and language
Managed AI OperationsFor AI used in daily operations, with several agents or processes.
- Ideal for
- AI in daily operations — several agents or processes
- A team that wants to hand over part of operations
- Entry requirements
- Everything in the previous level
- A non-production environment or an agreed change window
- Example cases for an evaluation set (anonymised where they contain personal data)
- An agreed incident channel and named contacts
- Included
- Monitoring of agreed signals
- Scheduled evaluation and regression review before changes
- Maintenance of prompts, configuration and integrations within the agreed volume
- Incident triage and escalation
- Monthly cost review
- Not included
- Unlimited changes or new development
- Business-rule decisions without you
- Model and infrastructure costs
- Monitoring and evaluation
- Checks on the signals the system supports or that we add together
- An evaluation set run before every change and on a schedule
- Changes and human approval
- Every change is tested in a non-production environment
- Changes that affect customers are approved by you before release
- A change log linked to the version
- Incidents and escalation
- Classification by impact and initial analysis
- Escalation to your contact through the agreed channel
- Post-incident review with preventive actions
- Cost and usage
- Monthly review of usage and cost by model and process
- Alerts on deviations where the provider supports them
- Security and privacy
- A separate least-privilege account per environment
- Keys stay in your secrets store
- A log of our actions
- Onboarding
- Quality and cost baseline
- First evaluation set and incident procedure
- Agreeing the volume and the channels
- Offboarding
- Handover of the evaluation set, change log and procedures
- Removal of the access and keys created for us
- Dependencies on your access and integrations
- Access to logs and metrics, or permission to add instrumentation
- A test environment or another safe way to check changes
- Access to the agents' configuration
- Agreed commercially
- An agreed monthly volume of hours or tasks
- Business hours and the incident channel
- Response targets — only when they match real staffing and are written into the contract
Advanced AI Reliability & OptimizationFor business-critical AI processes with many integrations or sensitive data.
- Ideal for
- Business-critical AI processes
- Many integrations, actions in systems or sensitive data
- Entry requirements
- Everything in the previous levels
- Documented business rules and agent permissions
- A change approver and a deputy
- A written incident procedure accepted by both sides
- Included
- Custom evaluations for your tasks
- Managed changes with approval and a rollback plan
- Integration health checks
- In-depth cost and model-choice reviews
- Quarterly reliability and improvement plan
- Not included
- A guarantee that every model answer is correct
- Work without a written contract and explicit permission
- Decisions that are yours (legal, financial, HR)
- Monitoring and evaluation
- Extended monitoring of workflows and integrations with your tools
- Custom evaluations — for example accuracy of extracted fields or the share of cases handed to a person
- Changes and human approval
- Changes go through your approval, the evaluation set and a rollback plan
- Changes to permissions, models and integrations only with explicit written approval
- Model output never bypasses deterministic checks and permissions
- Incidents and escalation
- An agreed procedure: classify, notify, mitigate or roll back
- Post-incident review and tracking of preventive actions
- Cost and usage
- In-depth reviews: models, prompts, caching, volume
- The effect is measured against a baseline — no promised percentages
- Security and privacy
- Least privilege for every integration
- Key rotation under your procedure
- A separate risk assessment for sensitive data
- Onboarding
- Full assessment of the architecture, integrations and permissions
- A custom evaluation set and baseline
- An accepted incident and change procedure
- Offboarding
- Handover of all procedures, evaluations and logs
- A joint check that access was removed
- A transition plan to your team or another provider
- Dependencies on your access and integrations
- Access to every integration in scope with separate accounts
- A change-testing environment close to production
- Availability of your approvers
- Agreed commercially
- Scope of systems, integrations and evaluations
- Business hours, channels and response targets per contract
- Review frequency and report format
Available today and planned
We keep what we can do today clearly apart from what is still being built.
Available today
- Technical assessment of the AI system
- Reviews from your logs, dashboards and reports, with agreed read-only access
- Evaluation sets and regression reviews before changes
- Maintenance of prompts and configuration within the agreed volume
- Monthly reports and recommendations
- Incident triage through an agreed channel during business hours
Planned (not available yet) Planned
- AI Operations Portal — a dashboard with data from connected systems
- Automatic telemetry collection from your system
- Alerts on quality or cost deviations
- Self-service for reports and approvals
The website shows no live data, does not monitor your systems and promises no specific integrations. Every integration is confirmed after the technical assessment.
Service scope
What is included
- Agreed monitoring based on the integrations available
- Maintenance of configuration, prompts and integrations
- Quality evaluation against an agreed test set
- Incident response under an agreed procedure
- A monthly summary and recommendations
What is excluded
- Unlimited new feature development
- Third-party model, subscription and infrastructure costs
- Guaranteed correctness of every model answer
- Guaranteed response times or availability without a signed contract
- Access to your systems without a contract and explicit permission
Frequently asked questions
Does the AI system have to be built by you?
No. We also take over systems built by other teams, after a technical assessment of the architecture, integrations and access.
What access do you need?
The minimum required, agreed after the assessment — for example read access to logs and usage reports. We never ask for passwords or keys through the inquiry form.
Can you monitor the system without changing it?
Partly. Availability, errors and costs can often be tracked from existing logs and reports. More detailed monitoring requires an additional integration that we agree on together.
Do you guarantee a response time?
Response times and support hours are agreed in writing for each system. This website contains no binding guarantees.
Will you reduce our AI costs?
There are often opportunities — a more suitable model, shorter prompts, caching. We do not promise a specific percentage; we measure the effect against a baseline.
How do you handle personal data in the AI system?
As a processor under a separate data processing agreement: minimal access, technical metrics where that is sufficient, and access to content only where agreed.
How do we start?
With a short inquiry and a call. We then propose a technical assessment and a suitable plan in a written quote.
Let's review your AI system
Briefly describe the system and its problems. We will propose a technical assessment and a suitable scope.
AI assessment request
Answer a few questions about the AI system. Do not enter passwords, keys or confidential data.