Skip to main content
Operating practice31 Aug 2026 · 8 min read

Adopting AI is a project. Running it is a full-time job.

The expensive part of AI automation is not the pilot. It is the 24 months that follow — and it appears in no quote.

The prototype works. The assistant answers invoicing questions, the agent classifies documents, the model pulls deadlines out of contracts. Management is impressed, IT is proud, the business case adds up on paper.

Then six months pass.

The model provider announces a retirement date for the version you use. A department introduces a new document format the extractor has never seen. Someone notices accuracy has been slipping for weeks — but nobody can say since when, because there is no baseline. Legal asks for the AI Act documentation. And the colleague who built the whole thing is now a team lead with no time left.

This is not an edge case. This is the normal case.

What the pilot costs — and what comes back every year after it

In person-days, mid-sized company running two to three production use cases

Pilot projectone-off20 PDOngoing operationevery year again151–290 PD0100200300 person-days
The pilot impresses people and is quickly paid for. Operation impresses nobody and arrives every year regardless. The lighter block shows the gap between well-run and badly-run operations.

AI is not a software rollout

With conventional software the arithmetic is simple: install once, then it runs. Updates come from the vendor, maintenance is a predictable percentage of licence cost, and if nobody touches it, the system behaves the same in two years as it does today.

AI systems behave differently. They decay even when nobody touches them — because everything around them moves.

  1. The models moveProviders retire versions, change response behaviour, shift safety boundaries. A prompt that reliably returned structured JSON in January produces prose with a preamble in June. Nobody notices automatically — unless someone built regression tests and keeps them current.
  2. The data movesNew contract templates, renamed ERP fields, a different scanning process in the mailroom. Any of these can quietly degrade a retrieval pipeline. Not break it — degrade it. That is the more dangerous version, because the system keeps answering, just worse.
  3. The regulation movesThe EU AI Act transparency obligations have applied since 2 August 2026. Germany's NIS2 implementation act has applied since December 2025 with no transition period, including personal liability for management. Neither asks for a one-off action; both ask for ongoing evidence.
  4. The attackers moveInjected instructions inside uploaded documents, data leaking through over-broad tool permissions, model endpoints without limits. An AI system with access to internal systems is a new attack surface — and one most companies have no hardening playbook for.

An unmaintained AI system does not break. It quietly gets worse.

Schematic view of the typical quality curve after go-live

100 %70 %40 %with ongoing maintenancewithout ongoing maintenanceGo-liveMonth 3Month 6Month 9Month 12
Schematic, not a measurement series. The point is the shape of the curve, not its exact path: there is no outage to trigger an alert — there is a slow decline that nobody notices without measurement.

The regulatory clock runs alongside

Anyone running AI runs it inside a legal framework that changes faster than their own roadmap. The Digital Omnibus pushed the high-risk obligations back in July 2026 — that buys preparation time, but removes no obligation. And the obligations already in force are the ones mid-sized companies most often miss.

What applies now and what is still coming

EU AI Act and NIS2 in order of deadline, not to scale

today02/2025Prohibitions,AI literacy08/2025Obligations forgeneral-purpose AI12/2025NIS2 in force,no transition period08/2026Transparency duties,Art. 5012/2026Marking rules forexisting systems12/2027High-risk,Annex III08/2028High-risk in products,Annex I
To the right of the dashed line is the preparation time created by the Digital Omnibus (Regulation (EU) 2026/1744). To the left are obligations already in force. Generative systems already on the market before 2 August 2026 have until 2 December 2026 for machine-readable marking.

What this actually costs

The breakdown below assumes a realistic mid-market scenario: 150 to 500 employees, two to three production use cases — document extraction, an internal knowledge assistant, a compliance check — operated in-house.

The basis is €800 per person-day for internal work, meaning fully loaded cost rather than gross salary, and €1,400 for external contracting. The range in the cost column reflects exactly those two ends.

Twelve cost blocks that recur every year after go-live

Annual range in thousands of euros

Model and provider changes12–42Prompt and retrieval upkeep20–63Quality measurement16–49Regulatory work and evidence12–42Data protection6–21IT security12–35Operations and incidents16–56Infrastructure9–60Model usage under load6–48Training and AI literacy6–21Rework and technical debt16–56Knowledge continuity4–210204060€k
Staff effortDirect costs
No single block looks dramatic. That is precisely why none of them gets budgeted — and why the total is a surprise at year end.
Operating costs in detail
Cost blockWhat sits behind itEffort p. a.Cost p. a.
Model and provider changesTracking retirement dates, testing version changes, fallback routing, comparison runs15–30 PD€12,000–42,000
Prompt and retrieval upkeepUpdating prompts, adjusting chunking, re-indexing, onboarding new document types25–45 PD€20,000–63,000
Quality measurementBuilding and maintaining a reference set, regression tests, error rates, drift monitoring20–35 PD€16,000–49,000
Regulatory work and evidenceAI Act Art. 4 and Art. 50, Annex III preparation, ISO/IEC 42001, audit documentation15–30 PD€12,000–42,000
Data protectionImpact assessments, processing agreements, technical measures, deletion policies, transfers8–15 PD€6,400–21,000
IT securityHardening against injected input, secrets management, permissions, penetration testing, patching15–25 PD€12,000–35,000
Operations and incidentsLogging, alerting, on-call, incident handling, restore testing20–40 PD€16,000–56,000
InfrastructureCompute and hosting, vector database, observability, staging environmentDirect cost€9,000–60,000
Model usage under loadUsage-based billing or running your own models under sustained loadDirect cost€6,000–48,000
Training and AI literacyObligation under AI Act Art. 4, role-specific training, onboarding new staff8–15 PD€6,400–21,000
Rework and technical debtFramework migrations, architecture changes, paying down prototype shortcuts20–40 PD€16,000–56,000
Knowledge continuityHiring and onboarding, documentation, cover arrangements5–15 PD€4,000–21,000
Total151–290 PD€135,000–515,000

Basis: €800 per internal person-day (fully loaded), €1,400 external. Direct costs estimated separately.

Running three to five use cases will not triple the person-days, but it will not keep them flat either. Realistically the factor is 1.8 to 2.2.

Where those days go is just as telling. Half of them go not into compliance or security, but into the plain work of catching the technology up with changes that happened somewhere else.

Where the person-days go

Midpoint of the range, roughly 220 person-days per year

220person-daysTechnical upkeep: models, prompts, testing, rework115 PD, 52 %Operations and security50 PD, 23 %Regulation, data protection, training46 PD, 21 %Knowledge continuity10 PD, 5 %
The largest block is the least remarkable one: adapting systems to changes somebody else set in motion.

The line items that appear in no table

  1. The bus factorIn most mid-sized companies building AI in-house there is exactly one person who understands why the system is built the way it is. When that person resigns, goes on leave or gets promoted, what begins is not maintenance but archaeology. Rebuilding typically costs 40 to 60 percent of the original development effort.
  2. The opportunity costThe person spending 200 person-days a year maintaining AI systems is usually the best engineer or architect in the building. What they would have produced in product or process work over the same period appears on no cost centre — and is often the most expensive item in the whole calculation.
  3. The evidence gapReconstructing model versions, prompt states, test results and decision records once the audit has started costs a multiple of what continuous logging would have cost. Under section 65 of the German BSIG, late NIS2 registration alone can carry a fine of up to €500,000; the AI Act penalty range reaches €15 million or 3 percent of global annual turnover in the standard case.
  4. Write-offs on technology shiftsWhat was built in 2024 on the framework of the day is often not the obvious architecture in 2026. Building in-house means paying for those shifts yourself — and deciding for yourself when to follow them.

In-house or specialist provider

The honest comparison, including the rows where in-house wins.

Two routes, the same job
CriterionBuilt in-houseSpecialist provider
Time to production4 to 9 months4 to 10 weeks
Ongoing operating cost€135,000–515,000 per yearpredictable service fee
Model updatesyour risk, your effortincluded in the service
Regulatory monitoringhas to be builtpart of the product
Audit evidenceproduced by yougenerated continuously
Knowledge riskhigh, often one persondistributed at the provider
Data sovereigntymaximumdepends on hosting model, the key selection criterion
Domain controlmaximumhigh where processes are configurable
Handling edge casesunlimitedbounded by product scope
Scaling to new use caseseach one a new projectincremental

The two highlighted rows are why the answer is not always to outsource.

When building in-house is the right call

This calculation is not an argument against internal development. It is an argument against unexamined internal development. Building it yourself is right when at least two of these four apply.

Four-point check
  1. The use case is core to the business model and therefore a competitive advantage in itself.
  2. At least three people in-house have real operating experience with AI systems — not one person with an interest in the topic.
  3. No market product covers the domain.
  4. The data situation is specialised enough that a standard product would need more adaptation than a build.

If none or only one applies, in-house is usually the most expensive option — and it only reveals itself as such after 18 months, when unwinding it already hurts.

Three questions before any AI rollout

  1. Who is named on the org chart as responsible for model changes, quality measurement and audit evidence?And how many days a year are they allocated?
  2. How would we notice quality dropping?If the answer is “when somebody complains”, there is no measurement.
  3. What happens if that person leaves?If the answer is uncomfortable, the risk has already been taken.

Our approach

This is exactly where Agentic360 comes in. We run a compliance automation platform for German mid-sized companies in which the operating tasks listed above do not sit with the customer: model changes, quality assurance, regulatory monitoring and evidence generation run on our side.

Two things matter to us. Data sovereignty stays with the customer — the platform runs locally hosted models on German infrastructure, and on-premise and air-gapped installations are part of the product rather than an exception. And decisions stay traceable: rules are evaluated in a deterministic policy engine, not hidden inside a model prompt. Every assessment traces back to the underlying regulation and the processing step that triggered it. That is the difference between “the system says X” and “the system says X because”.

If you are currently working out what your own AI operation costs, or realising you have never done that calculation: talk to us. Even if the answer turns out to be building it yourself — at least let it be a deliberate decision with a budget behind it.

Key takeaway

This calculation is not an argument against internal development. It is an argument against unexamined internal development. Even if the answer turns out to be building it yourself — at least let it be a deliberate decision with a budget behind it.

As of August 2026. Regulatory deadlines change; the details reflect the position at publication. This article is not legal advice. The cost figures are estimates based on the stated day rates, not measurements.

Sources
Book a demo

Turn the regulation into a running system.

30 minutes, scoped to your frameworks and integrations. You leave with a concrete plan — not a sales loop.