Who Maintains AI After Deployment? Build vs Buy vs Managed Operations

Deploying an AI model is only half the journey. The real challenge starts the day after go-live — when the system has to run stably, deliver business value, and evolve with the data. This article takes apart three models for maintaining AI systems, compares their costs, and helps you decide: build an in-house team, buy a SaaS platform, or outsource to managed operations?

If you are a CTO or COO of a company that has just finished (or is planning) an AI deployment — this guide gives you concrete numbers, a decision tree, and a real 12-month cost comparison.

The problem nobody discusses — what happens AFTER an AI deployment?

The AI consulting industry loves talking about deployments. Case studies, proof-of-concept, MVP, pilots — all of it sounds exciting and looks good in a board presentation. But there is a question consultants rarely raise during the sale: who will maintain this system in six months?

A statistic that should worry you

According to a 2025 Gartner report, 60% of AI projects in mid-size companies die within the first 12 months after deployment. Not because the technology does not work. Not because the business sees no value. They die because nobody planned their maintenance.

The phenomenon has a name — the “AI deployment graveyard”. Companies invest PLN 200–500 thousand in a deployment, enjoy the first results for 3–4 months, and then the model starts losing accuracy, infrastructure generates unplanned costs, and the IT team lacks the skills to retrain. The system is switched off “temporarily” — and never comes back.

Three main causes of death for AI systems after deployment

Model degradation — every machine-learning model is trained on historical data. The world changes, the data change, and the model stays static. Without regular retraining, accuracy falls systematically — on average 10–15% per quarter in predictive models.

Data drift — the input data the model sees in production gradually move away from the training-data distribution. New products, new customers, a change in seasonality, macroeconomic effects — all of this means the model “sees” data it has never processed before.

Uncontrolled infrastructure costs — GPU instances, storage, data transfer, monitoring — monthly cloud bills can grow 20–40% year on year without active optimisation. A company that does not monitor AI infrastructure costs often learns about the problem only from the cloud invoice.

3 models of AI maintenance

Three approaches to maintaining AI systems after deployment have crystallised in the market. Each has its advantages, limits, and a concrete cost range.

Build — an in-house ML/MLOps team

The “Build” model means assembling a dedicated internal team responsible for maintaining, monitoring, and developing AI systems. Organisations choose this when AI is a core product or a key competitive differentiator.

Minimum team composition:

1× ML Engineer (senior) — retraining, feature engineering, model optimisation
1× MLOps Engineer — CI/CD pipelines, infrastructure, monitoring
1× Data Engineer — data quality, ETL, data pipelines
0.5× Security/Compliance — audits, GDPR, model governance
0.5× Engineering Manager — coordination, SLA, communication with the business

Monthly cost: PLN 80,000–150,000 (salaries + tools + infrastructure). In Warsaw a senior ML Engineer is PLN 25,000–40,000 employer gross, an MLOps Engineer — PLN 22,000–35,000.

Time to stand up: 3–6 months (recruitment + onboarding + process build-out).

Buy — a SaaS / managed platform

The “Buy” model means using a ready-made MLOps platform (e.g. Weights & Biases, managed MLflow, Databricks, Vertex AI) to automate monitoring and retraining. The company buys the tool, but still needs someone in-house to operate it.

Monthly cost: PLN 5,000–30,000 (platform licence + 0.5–1 FTE internally to operate it).

Limits: SaaS platforms work well for standard use cases. But when the model is custom, the data have a specific structure, and compliance requirements go beyond the standard — you start fighting the platform’s constraints. Customisation is possible, but expensive and slow.

Time to stand up: 2–4 weeks (platform configuration + integration).

Managed Ops — outsourcing maintenance to a partner

The “Managed Ops” model means handing responsibility for maintaining AI systems to an external partner who provides a dedicated team, processes, and an SLA. It is a hybrid — the company keeps strategic control, but operational maintenance sits with the partner.

Monthly cost: PLN 10,000–50,000 (depending on system complexity and SLA level).

Advantages: full customisation flexibility (the partner works on your stack), no recruitment risk, fast start, contractual penalties for missing SLA.

Time to stand up: 1–2 weeks (onboarding + taking over the system).

Comparison table of AI maintenance models

Criterion	Build (in-house)	Buy (SaaS)	Managed Ops
Monthly cost	PLN 80,000–150,000	PLN 5,000–30,000	PLN 10,000–50,000
Control	Full	Limited to the platform’s capabilities	Full (your stack)
Flexibility	Maximum	Low–medium	High
Time to start	3–6 months	2–4 weeks	1–2 weeks
Recruitment risk	High	None	None
SLA with contractual penalty	None (internal KPIs)	Depends on the vendor	Yes
Scalability	Limited by the HR budget	Limited by the plan	Flexible

What goes into maintaining an AI system? (Checklist)

Maintaining an AI system is not just “watching whether it works”. It is a set of regular, repeatable activities that have to be done systematically — often on a 24/7 regime. Below is the full list of tasks that make up professional managed operations.

Model-metric monitoring

Tracking accuracy, precision, recall, F1-score in real time
Alerting when a metric drops below a set threshold (e.g. accuracy < 85%)
Business dashboards — translating technical metrics into KPIs the board understands
A/B testing new model versions before production deployment

Retraining pipeline

Automatic or semi-automatic retraining on new data (weekly/monthly)
Validation of the new model before deployment (automated testing, canary deployment)
Model versioning — ability to roll back to the previous version in < 15 minutes
Feature store management — updating and validating input features

Infrastructure

Managing compute resources (GPU/CPU scaling)
Cloud cost optimisation (right-sizing, spot instances, reserved capacity)
Backup and disaster recovery — RPO < 1h, RTO < 4h
Secrets, credentials, and API-key rotation

Security

Vulnerability monitoring in dependencies (CVE scanning)
Penetration testing of inference endpoints
Model security — protection against adversarial attacks and prompt injection (for LLMs)
Audit trail — logging every change to the model and the data

Compliance and governance

Model documentation (model cards, data sheets)
GDPR compliance — the right to an explanation of automated decisions (Art. 22)
Bias and fairness audits — quarterly validation of equal treatment
Reporting to the regulator (EU AI Act — mandatory for high-risk from August 2026)

SLA and incident management

Defined response times (P1: 15 min, P2: 1h, P3: 4h, P4: 24h)
Runbooks for typical incidents
Post-mortem after every serious incident
Monthly SLA reports with availability metrics

Model degradation — why AI stops working after 6 months

Model degradation is the biggest silent killer of AI projects. Unlike traditional software, which works equally well (or equally badly) for years — ML models have a built-in “expiry date”. Understanding this mechanism is essential for every CTO running AI systems.

Data drift — the data change, the model does not

Data drift occurs when the production data distribution moves away from the training data distribution. Causes can be subtle:

Seasonality — a model trained on winter data struggles in summer
Change in customer behaviour — a pandemic, inflation, new buying trends
New products/services — the model never “saw” them during training
Changes in data sources — a new ERP, a CRM migration, a change of data format

Concept drift — what the model is supposed to predict changes

Concept drift is more dangerous than data drift, because the relationship between input data and the outcome itself changes. Example: a model predicting customer churn risk trained before a price increase — after the increase, entirely different factors drive customers to leave.

Feedback-loop degradation

In systems where the model’s decisions affect future training data, a feedback loop forms. A model that rejects a certain category of candidates never sees data about their possible success — and becomes even more convinced of a wrong decision.

No monitoring — “you don’t know what you don’t know”

The worst scenario is a model that degrades in silence. Without active monitoring the company learns about the problem only when the business starts losing money — and that can take months.

An example from practice

A demand-forecasting model deployed in a manufacturing company. Training data: 2 years of sales history. Result after deployment:

Months 1–3: accuracy 92%, warehouse savings PLN 180,000/quarter
Months 4–6: accuracy 84% (−8pp), savings fall to PLN 120,000/quarter
Months 7–9: accuracy 71% (−13pp from start), the model generates more wrong forecasts than correct ones
Months 10–12: system switched off, return to manual planning

Accuracy drop: ~15 percentage points per quarter without retraining. This is not an exception — it is the norm for predictive models operating on dynamic business data.

With regular retraining (every 4–6 weeks) the same model holds accuracy > 88% throughout the period — confirmed in our QCare deployments.

How much does a LACK of maintenance cost? (Hidden losses)

The paradox of AI maintenance is that the cost of no maintenance is many times higher than the cost of professional managed operations. Companies that “save” on maintenance pay for it in ways that often never show up directly in the IT budget.

Downtime and lost revenue

When an AI system stops working correctly, the business processes that depend on it fall back to manual mode. For a company with automated customer service (chatbot, query routing) every hour of downtime means:

Longer response times — NPS down 5–15 points
Extra load on the team — overtime, burnout
Lost conversions — customers who leave because they did not get an answer in 5 minutes

Estimated cost of 1 hour of AI-system downtime in a 50–200 person company: PLN 2,000–15,000 (depending on process criticality).

Wrong predictions — decisions on bad data

Worse than a switched-off model is a model that runs but returns wrong results — and the organisation does not know. Real-loss examples:

A pricing model that underprices by 12% for 3 months — margin loss: PLN 340,000
A scoring model that over-rejects leads — lost revenue: PLN 180,000/quarter
A demand-forecasting model that over-orders — excess stock: PLN 200,000 of frozen capital

Compliance violations

From August 2026 the EU AI Act imposes documentation and monitoring duties on high-risk AI systems. A lack of compliance maintenance means:

Administrative fines of up to 3% of global turnover
An obligation to withdraw the system from the market
Personal liability of the board (Art. 49 AI Act)

Loss of team trust

This is the hardest cost to measure, but it is real. When an AI system fails — operations teams lose trust in the technology. The next AI deployment will meet resistance: “last time they promised it would work too”. Rebuilding an organisation’s trust in AI after a failed project takes 12–18 months.

Total cost of no AI-system maintenance over 12 months in a typical SME: PLN 300,000–800,000 (sum of downtime, bad decisions, compliance, and opportunity cost).

QCare from QA10 — managed operations with an SLA

QCare is a managed AI operations service designed specifically for SMEs that have deployed AI systems and need professional, cost-predictable maintenance — without building an in-house MLOps team.

5 plans matched to scale

Plan	Monthly price	Models covered	SLA uptime	P1 response time
Lite	PLN 1,800	1 model	99.0%	4h
Standard	PLN 4,500	up to 3 models	99.3%	2h
Professional	PLN 8,500	up to 5 models	99.5%	1h
Business	PLN 13,000	up to 10 models	99.5%	30 min
Enterprise	PLN 18,500	no limit	99.9%	15 min

What is included in every QCare plan?

24/7 monitoring — continuous tracking of model metrics, alerting, dashboards available to your team. You do not wait for a report — you see system state in real time.

Retraining pipeline — automatic or managed model retraining on new data. Frequency matched to the dynamics of your data — from weekly to monthly.

99.5% uptime SLA (Professional plan and above) — guaranteed system availability with a contractual penalty for missing it. This is not a marketing promise — it is a contractual commitment.

Incident response — a dedicated team responding to incidents within the times set in the SLA. Runbooks prepared for your system, escalation to ML Engineers in < 15 minutes.

Monthly reports — model performance, data-drift trends, optimisation recommendations, infrastructure-cost forecast for the next quarter.

Compliance support — model documentation aligned with the EU AI Act, support in audits, model-card updates at every retraining.

Contractual penalty for missing SLA

QCare is the only managed AI operations service on the Polish market with a contractual penalty for missing SLA. If in a given month uptime falls below the guaranteed level:

Drop of 0.1–0.5pp → credit of 10% of the monthly fee
Drop of 0.5–1.0pp → credit of 25% of the monthly fee
Drop of > 1.0pp → credit of 50% of the monthly fee

That is real skin in the game — our interest is aligned with yours. We maintain your system because we pay out of our own pocket when it fails.

When to build an in-house team, and when to outsource

The “Build vs Managed Ops” decision is not binary — it depends on the organisation’s context. Below is the decision tree we use in consultations with QA10 clients.

Build an in-house team when:

AI is your product — you sell an AI-based solution and need full control over IP
Budget > PLN 200,000/month for the AI team — you have the resources for recruitment, tools, and development
You have > 10 models in production — the scale justifies a dedicated team
Regulatory requirements demand full internalisation — finance, defence, medical sectors with restrictions on external vendors
AI skills are your competitive moat — keeping know-how inside the organisation is strategically critical

Outsource to managed ops when:

AI is an operational tool — it supports processes but is not the product
Budget < PLN 100,000/month — you cannot afford a full ML/MLOps team
You have 1–5 models in production — the scale does not justify a dedicated team
You want to start fast — you do not have 3–6 months for recruitment
You need predictable costs — a fixed monthly fee instead of variable HR costs
You want an SLA with a contractual penalty — an in-house team will not give you a contractual commitment

The grey zone — a hybrid model

In practice many companies choose a hybrid: an in-house Data Scientist (1 FTE) owns strategy and prioritisation, and operational maintenance sits with the managed-ops partner. That gives strategic control at a low operating cost — a typical budget: PLN 30,000–60,000/month (1 FTE + QCare Professional/Business plan).

Case study: a manufacturing company — QCare vs an in-house team

To illustrate the cost difference between the models, let’s walk through a real scenario of a manufacturing company (150 people, 3 AI models in production: demand forecasting, production optimisation, quality control).

Scenario A: In-house team (Build)

Item	Monthly cost	Annual cost
ML Engineer (senior)	PLN 35,000	PLN 420,000
MLOps Engineer	PLN 28,000	PLN 336,000
Data Engineer (0.5 FTE)	PLN 14,000	PLN 168,000
Tools and licences	PLN 8,000	PLN 96,000
Infrastructure (GPU, storage)	PLN 12,000	PLN 144,000
Recruitment (one-off, amortised over 12 months)	PLN 7,500	PLN 90,000
TOTAL	PLN 104,500	PLN 1,254,000

Additional risks: employee turnover (average ML Engineer tenure in Poland — 18 months), time to onboard a new team member (2–3 months), no SLA (an in-house team does not sign a contract with a penalty clause).

Realistic 12-month cost including turnover and idle time: ~PLN 480,000–600,000 (assuming 1 turnover and a 2-month vacancy gap).

Scenario B: QCare Professional

Item	Monthly cost	Annual cost
QCare Professional (up to 5 models)	PLN 8,500	PLN 102,000
Infrastructure (passes through to the client)	PLN 10,000	PLN 120,000
TOTAL	PLN 18,500	PLN 222,000

Annual saving: PLN 258,000–378,000 compared with an in-house team.

What does the company gain on QCare vs in-house?

No recruitment risk — you do not have to hunt an ML Engineer in a market where there are 0.3 candidates per opening
SLA with a contractual penalty — a contractual commitment to 99.5% uptime
Immediate start — taking over the system in 1–2 weeks vs 3–6 months of recruitment
Predictable budget — a fixed amount, no surprises from raises, sick leave, turnover
Access to a team of experts — not 1 ML Engineer, but a team with experience across dozens of deployments

When does in-house pay off?

At a budget > PLN 200,000/month and > 10 models in production — an in-house team starts to be more cost-efficient. But for a typical SME with 1–5 AI models, managed ops is the rational financial choice.

FAQ — the most common questions about AI maintenance

Can I maintain an AI system myself, without a dedicated team?

In theory yes — if you have an in-house Data Scientist who deployed the model, they can run basic monitoring. In practice this works for 3–6 months, after which lack of time (the Data Scientist has other duties) and lack of process (ad-hoc retraining, no runbooks) lead to system degradation. Maintaining AI requires discipline and process — not only technical skill.

How long does QCare onboarding take?

Standard onboarding takes 5–10 working days. In that time we take over infrastructure access, document the system architecture, configure monitoring, prepare runbooks, and run knowledge transfer with your team. From day 11 the system is under our care with a full SLA.

What happens if I want to return to in-house maintenance?

QCare has no lock-in. The contract provides a 30-day notice period, during which we run a full knowledge transfer — documentation, runbooks, access, incident history. We leave cleanly, leaving you a complete operational package.

Does QCare cover infrastructure costs (GPU, cloud)?

No — cloud infrastructure costs are billed directly between the client and the cloud provider (AWS, GCP, Azure). QCare covers managing that infrastructure — cost optimisation, scaling, monitoring — but the compute bill itself stays with the client. That way you keep full cost transparency and control over the provider.

Can I start on the Lite plan and move up?

Yes — a plan upgrade is possible at any time during the contract, with SLA continuity. A downgrade requires 30 days’ notice (because we reorganise resources on our side). Most clients start on Standard or Professional and scale as they bring more models online.

Summary — do not let your AI die after deployment

Deploying an AI system is an investment — often on the order of PLN 200,000–500,000. Letting that investment lose value through a lack of maintenance is like buying a car for a million zloty and never changing the oil. Technically possible — economically absurd.

Three things you should do now:

Check whether your current AI system has professional maintenance — monitoring, retraining, SLA. If the answer to any of those is “no” — your system is probably already degrading.
Count the real cost — compare an in-house team (PLN 80,000–150,000/month) with managed operations (PLN 8,500–18,500/month). For most SMEs, outsourcing is 3–5× cheaper.
Plan maintenance BEFORE deployment — if you are only now planning an AI deployment, include the maintenance cost in the budget from day zero. This is not an optional add-on — it is an integral part of the project.

QCare from QA10 — managed AI operations with a contractual penalty for SLA.

From PLN 1,800/month. 24/7 monitoring. Retraining pipeline. Incident response. No lock-in.

See QCare plans →

Questions about maintaining your AI system? Book a free consultation — we will analyse your case and propose the optimal maintenance model.