Most reporting on AI agents describes prototypes. This report looks at what happens after the demo: how teams in Germany, Austria, and Switzerland build, trigger, and operate agents once they carry real traffic. The pattern that emerges is less about smarter models and more about plumbing, reliability, and compliance. Those unglamorous layers decide whether an agent survives contact with production.
The headline numbers
Most agents reach production and stay event-driven
The largest share of agents are live, not experimental. 61% are in production, 28% are still in a prototype or pilot phase, and only 11% have been paused or retired. The teams that get to production quickly share a habit: they wire the agent to a real event source early instead of testing forever behind a manual prompt.
- In production61%
- Prototype / pilot28%
- Paused or retired11%
That habit shows up in how runs start. 72% of production agents are event-driven, woken by an inbound connector. Another 9% are synchronous, a connector in sync mode (or a chat call) that waits for the result, and 19% run on a schedule through the cron connector. In every case a connector drives the run; the split is by trigger shape, and the shape, not the model, sets the agent's delivery guarantees. We break down the trade-offs in our comparison of connector patterns.
- Event-driven72%
- Synchronous (request/response)9%
- Scheduled (cron)19%
How agents get triggered
Webhooks lead by a wide margin, since they are the simplest way for an existing service to call an agent, followed by scheduled cron runs and, for higher-throughput teams, Kafka and Postgres. The long tail of queue, file, and messaging sources is real but smaller. See the full set on the connectors page, or read why teams favor pre-built connectors over custom glue.
The model mix
No single model dominates. Work splits across the Claude, GPT, and Gemini families, with teams routing cheap, fast models to high-volume steps and saving the strongest models for reasoning-heavy ones.
Custom and open models vs the big providers
Most agents run a commercial frontier model as-is, but a meaningful minority do not. 39% run a customized or open model: 16% fine-tune a provider model on their own data, and 23% run open-weight or self-hosted models outright. For a production fleet that customized-and-open share is notably high, and it lines up with the region's data-sovereignty pressure.
- Provider models, off-the-shelf61%
- Provider models, fine-tuned16%
- Open-weight / self-hosted23%
The split comes down to control versus convenience. A big commercial model is the fastest route to a working agent, and most teams take it. The ones that move to a fine-tuned or self-hosted model tend to sit in the most regulated, most German-language-heavy corners of the data, where a smaller tuned model is cheaper at volume, keeps inference on infrastructure the team controls, and removes any dependence on a single US provider. It is more to run, and those teams have decided the data is worth it. For the regulatory backdrop, see running agents in the EU.
What teams actually build
The use cases are firmly operational. Support and ticket triage leads, followed by document processing and data enrichment. These are agents that sit inside an existing workflow and clear a backlog of repetitive work, not customer-facing chatbots.
Adoption by industry
Adoption concentrates in regulated, document-heavy sectors. Insurance, banking, and manufacturing make up the largest share of DACH teams running agents on Connic, with logistics and telecom close behind. These are sectors where a clear audit trail and EU data handling are not optional.
What breaks first
When agents fail in production, the model is rarely the cause. Integration and tool errors top the list, followed by timeouts and malformed output. It is the same failure surface as any distributed system. Prompt-injection attempts are real, but where guardrails are enabled they are caught rather than executed.
Bring the trigger source, governance requirement, and EU data-handling constraint behind your agent. Sales can map it to Connic connectors, guardrails, judges, approvals, and traces.
Discuss your production agentCost and token economics
Model spend per run is low but uneven. The median run draws an estimated €0.038across 7.4k tokens, but most of the spend concentrates in multi-step, tool-calling agents that loop through several model calls. RAG-style retrieval is the next largest bucket. If you are optimizing cost, the lever is the number of steps, not the price per token. Our observability guide covers how to see where the spend goes.
Where the money goes, by agent pattern:
Operational maturity
The teams furthest along treat agents like software. 66% gate deployments on automated tests, and 41% route high-stakes actions through a human-in-the-loop approval before the agent proceeds. Both numbers climb with the sensitivity of the workload: the more an agent can move money or touch customer data, the more likely it is to be tested and gated.
Compliance is the deciding factor
For DACH teams, where an agent runs is a procurement question before it is a technical one. EU AI Act readiness and data residency are the two most-cited drivers, ahead of speed and cost. The pattern is consistent across regulated industries: a capable agent that cannot demonstrate EU data handling does not get deployed. We cover the regulatory side in our guides to EU AI Act compliance for agents and running agents in the EU. If you are still choosing where to run, compare the best EU agent platforms for 2026.
Teams could cite more than one driver, so these do not sum to 100%.
What makes the DACH market different
Two things set DACH apart from the broader agent market: who has to approve an agent, and where its data is allowed to live. In Germany, Austria, and Switzerland an agent is rarely a unilateral engineering decision. It clears data protection, often a works council, and in regulated sectors a sectoral regulator, before it ships. The requirements that come up most often when DACH teams move an agent into production:
More than one applies to most teams, so these do not sum to 100%.
The buyers are mostly Mittelstand: mid-sized firms that weigh reliability and a named, accountable vendor over novelty. That shows up as a preference for a contract with an EU or German entity, EU-region operation, and clear EU AI Act alignment ahead of the lowest price or the newest model. The needs are concrete, and they are where a regional platform earns its place:
What it adds up to
The story of AI agents in DACH in 2026 is a story of operations. The teams shipping agents are not the ones with the cleverest prompts. They are the ones who treat an agent as a production service: triggered by real events, watched by guardrails and judges, tested before deploy, and run where the law requires. The model has become the easy part. Everything around it is where the real work now happens, and where the advantage is won.
See how the pieces fit together in the connectors overview, or start from the quickstart.