ENGAGEMENT / 04 · MANAGED AI SERVICES
We run what we built, so it still works in a year.
Monitoring, retraining, tuning and on-call for AI systems in production. Fixed monthly fee, agreed response times, a report every month that shows what it saved you.
- 24/7
- MONITORING
- <1h
- P1 RESPONSE
- 12
- REPORTS A YEAR
- 6.5K
- FROM, PER MONTH, IN EUROS
01 / What this actually is
What this actually is.
AI systems do not stand still. Data shifts, usage patterns change, providers deprecate models, and quality quietly decays — the drift nobody notices until a customer complains. This is a running operations contract for exactly that problem.
Think of it as an operations team for the machine-learning parts of your stack: dashboards and alerts watching behaviour, a scheduled retraining rhythm, continuous cost and latency tuning, and a named on-call rota with a contractual response time.
WHAT IT IS NOT
It is not generic IT support with an AI label. Nobody here restarts servers and calls it a service — the work is model quality, cost per request and failure behaviour.
02 / Who this is for
WHO THIS IS NOT FOR
If nothing is in production yet, this is premature — build first, then hand us the keys.
03 / How it runs
How it runs.
01 · WEEK 1–2
Onboarding and baseline
We instrument the system, capture current quality, latency and cost per request, and write down what 'normal' looks like. Everything after this is measured against that line.
02 · CONTINUOUS
Monitoring and alerting
Dashboards and automated alerts on accuracy, drift, latency, error rate, spend and abuse. Alerts route to a human, not to a mailbox nobody reads.
03 · MONTHLY OR QUARTERLY
Retraining and evaluation
New data in, model re-evaluated against the agreed test set, promoted only if it beats the incumbent. Rollback stays one command away.
04 · CONTINUOUS
Tuning and cost work
Prompt and routing changes, caching, smaller models where they are good enough. Most clients see cost per request fall over the first two quarters.
05 · ON DEMAND
Incident response
Named on-call, severity levels, agreed response times, and a written post-mortem for every priority-one incident.
06 · MONTHLY
Reporting and review
One report and one call: usage, quality, incidents, spend, and hours or euros saved. Plus what we intend to change next month.
03B / Service levels
What we commit to, in writing.
04 / What you hold at the end
What you hold at the end.
05 / Who you work with
Who you work with.
01
Service lead
Your single contact, owns the report, the roadmap and the awkward conversations.
02
ML operations engineer
Monitoring, retraining pipelines, evaluations and releases.
03
On-call rota
Named engineers with the context to fix your system, not a generic first-line desk.
06 / Price
from 6,500 € per month
MONTHLY RETAINER · 12-MONTH TERM · SCOPED PER SYSTEM
INCLUDED
- Monitoring, alerting and dashboards
- Scheduled retraining and evaluation runs
- Tuning work on quality, latency and cost
- On-call cover with agreed response times
- Monthly report and quarterly review
NOT INCLUDED
- Cloud, model and licence usage, billed by your providers
- New features or new systems — those are implementation work
- Support for systems we have never reviewed, until onboarding is complete
07 / Questions
Questions we get.
Can we cancel at any time?
The standard term is twelve months with ninety days' notice, because we hold named engineers for you. Shorter pilot terms are possible and are priced accordingly.
Can you run something you did not build?
Yes. We start with a technical review; if the system is maintainable we onboard it, and if it is not we tell you what has to change first. Systems we built ourselves onboard faster and cost less.
How do we know it is worth the money?
The monthly report carries the numbers that matter: cost per request, error rate, resolution time and hours saved. If those do not move in your favour, the review exists to say so.
What counts as an incident?
Severity one is the system down or producing harmful output; severity two is degraded quality or latency; severity three is everything else. Each has its own response window written into the contract.
Who keeps the knowledge in-house?
Your team keeps access to everything and joins the monthly reviews. Many clients take operations back after a year — that is a fair outcome, not a failure.
08 / Where this sits
Where this sits.
An AI system is not a product you buy once. It is a thing that has to be kept honest, every week.