AI is no longer an experiment at the edge of the organization. 78% of organizations reported using AI in 2024, up from 55% the year before, and by 2025 adoption had reached 88%. Almost all of it runs the same way: a request sent to a hosted frontier model over an API. For a pilot, that made complete sense - no infrastructure to stand up, minimal commitment. The trade-offs were deferred, not examined.
They are worth examining now, because the alternative has quietly become viable. It is the question the sovereignty conversation was always going to arrive at: we have covered where your data lives and under what law - what remains is the model itself.
What sending your data out actually costs
To use a hosted model, your data has to travel to it. For a general question, that’s fine. For a transaction record, a patient file, or a contract, it means routing regulated data through infrastructure you neither own nor govern - often with no clear answer to whether it’s kept or used to train someone else’s model.
That exposure is already showing up in the numbers. In IBM’s 2025 breach study, staff feeding company data into unauthorized AI tools - “shadow AI” - figured in one in five breaches, and most organizations still have no policy to prevent it. When the average breach already runs to millions and lands hardest in the sectors where AI is most useful, that is not a gap worth leaving open.
The point isn’t that hosted models are unsafe. It’s that where the model runs decides where your data has to go - and for a lot of enterprise work, that’s a decision worth making on purpose.
Match the model to the task
The answer is rarely all-or-nothing. Few organizations should run everything on open models, and none should route sensitive work through a public API out of habit. The workable pattern is to match the model to the task: a hosted frontier model for open-ended, low-sensitivity work where capability matters most, and a local open model for the transactional, regulated, or high-volume work where data locality and cost decide the outcome.
This is how Sirma.AI Enterprise is built - a platform that runs proprietary models (OpenAI, Google Gemini, Anthropic Claude) and open models such as Llama and Mistral side by side, routing each task to the right one by cost, performance, and regulatory context. A client can run exactly this split: a hosted model for general queries, and a local model on its own servers for transactional analysis, so the underlying data never leaves the premises.
What this looks like by industry
The reason to run a model in-house differs by sector.
In healthcare, clinical documentation, coding, and patient communication are among the clearest AI use cases and the most sensitive data an organization holds. With breach costs the highest of any sector for more than a decade, keeping both the model and the data inside the hospital’s governance boundary is not a preference but a compliance position.
In high-volume operations - logistics, hospitality, retail - the case is economic. A workflow handling millions of documents, shipments, or guest interactions a month pays a hosted model for every one, indefinitely. Running an open model on infrastructure you already own turns that per-request bill into a fixed cost.
The same logic applies in financial services, where transaction and KYC data cannot casually leave the institution and DORA now scrutinizes third-party dependency, and in the public sector, where sovereignty has moved from a vendor promise into a procurement criterion.
From download to daily use
If open models are capable and cheaper to run, why haven’t more businesses moved? Not the models - the engineering. Turning a download into something a business can rely on means running it securely, connecting it to real systems, and controlling who can use it. That is what an orchestration platform does: open models from Hugging Face can be brought into Sirma.AI Enterprise and run on your own servers or private cloud, connected to your ERP, CRM, and document stores, with all processing kept inside your infrastructure. Adopting an open model becomes a setup decision rather than a six-month project.
Where this leaves you
The question is no longer whether open, self-hosted models are viable for enterprise work. For a growing share of tasks, they are. It’s whether your setup lets you make that choice deliberately - task by task, with cost, sensitivity, and jurisdiction as the inputs - or whether it has quietly settled the matter for you by defaulting everything to an external API.
Sirma has built AI systems for regulated industries since 1992, incorporated in the EU and listed on European exchanges. Sirma.AI Enterprise applies that experience as an architecture: your choice of models, run where you need them, under governance you can demonstrate.
If you want to see which of your workloads could move onto open models you run yourself, we can map that with you.
Learn more about our platform: Sirma.AI
Sources:
Stanford HAI, AI Index Report 2025 (78% of organizations used AI in 2024, up from 55%) https://hai.stanford.edu/ai-index/2025-ai-index-report McKinsey & Company, The State of AI in 2025: Agents, innovation, and transformation (88% adoption in 2025) https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai IBM, Cost of a Data Breach Report 2025 (shadow AI in one in five breaches, most organizations lacking AI governance policies, and healthcare as the costliest sector for over a decade) https://www.ibm.com/think/insights/data-matters/cost-of-a-data-breach Regulation (EU) 2022/2554, Digital Operational Resilience Act (DORA) (third-party dependency scrutiny in financial services) https://eur-lex.europa.eu/eli/reg/2022/2554/oj
