On-premises AI means running AI models inside an organisation's own data centre or private cloud, so sensitive data never leaves its control. After years of cloud-first AI, banks and insurers are moving their most sensitive AI work back in-house: the share of enterprises running production AI mainly on public cloud fell from 56% to 41% in a single year. Y Combinator's latest Requests for Startups point the same way.
Of enterprises run production AI mainly on public cloud, down from 56% in one year
Would choose on-prem, private or hybrid for AI on sensitive data
Of custom enterprise AI tools reach production
For most of the last few years, enterprise AI meant sending data to a model in someone else's cloud. That works for drafting emails. It doesn't work for credit files, medical underwriting records or claims, which most regulated firms can't send outside their own environment.
The numbers show the shift:
| Finding | Number | Source |
|---|---|---|
| Enterprises using public cloud as their main home for production AI inference | 56% → 41% in one year | Broadcom Private Cloud Outlook 2026, via LLM Configurator |
| Enterprises that factor an AI vendor's country of origin into selection | 77% | Deloitte State of AI in the Enterprise 2026, via LLM Configurator |
| Would choose on-prem, private or hybrid for AI on sensitive data | 91% | Cloudian (2026) |
| Hit new cross-border data restrictions in the past two years | 45% | Cloudian (2026) |
| Firms with AI repatriation initiatives | 8% → 16% in a year | BARC Data Sovereignty 2026 |
The pattern is consistent: sensitive AI work is moving into private infrastructure, while low-risk tasks stay in the cloud.
| YC request | What it asks for | Why it points to private AI |
|---|---|---|
| Startups that want to sell to huge companies (Summer 2026) | Small teams landing Fortune 10 pilots and large deals in year one | Big-company deals pass through security, data-residency and procurement review |
| AI-native service companies (Summer 2026) | AI that delivers finished work in insurance, audit, compliance and healthcare admin | Every one of these sectors is regulated and handles customer data |
| Company brain (Summer 2026) | A structured map of a company's know-how that AI agents can use | A company's know-how is its most sensitive data |
| AI-native compliance infrastructure (Fall 2026) | Regulatory monitoring, anomaly flagging and reporting | Compliance tools can't leak the data they protect |
| AI open source and inference infrastructure (Spring 2025) | Open-source AI businesses and inference infrastructure | Open models are what make self-hosted AI possible |
YC's message to founders is to build AI that does real work inside big, regulated companies. In those companies, the first question is always where the model runs and where the data goes.
| Problem | What it looks like |
|---|---|
| Data sovereignty | Risk and legal teams block sending customer data, credit files or claims to a third-party AI service. |
| Regulatory accountability | Regulators expect an AI inventory, testing, human oversight and monitoring, including for AI bought from vendors. |
| Pilots that never ship | Demos impress, then stall on integration, accuracy and security sign-off. MIT research found only 5% of custom enterprise AI tools reach production. |
| AI that doesn't know the business | General chatbots can't apply the firm's own policies or learn from corrections. |
| Skills gaps | 44% of firms cite a lack of staff to run AI infrastructure (BARC). |
| Unpredictable cost | Usage-based bills grow quickly once AI agents run every step through a large model. |
| Shadow AI | Staff at over 90% of firms use personal AI tools, but only 40% of firms provide an official one (MIT NANDA). |
Most institutions pick one of four routes, and each trades control for effort:
| Route | Data control | Effort to run | Where it falls short |
|---|---|---|---|
| Cloud AI services with no-training terms | Contractual | Low | Data still leaves the institution, which fails strict residency reviews |
| Self-hosting open models | High | High | Needs scarce ML and operations staff; a model on its own has no audit trail or workflow |
| General private AI platforms | High | Medium | Broad search and chat tools, not the underwriting or claims workflow itself |
| Custom builds by systems integrators | High | High | Slow and bespoke, often a year or more to production |
Many large firms end up hybrid: sensitive steps on private models, low-risk tasks on cloud services. What's usually missing is the last step: an auditable workflow that runs on the institution's own infrastructure and reaches production in weeks.
We've been building private, auditable AI for banks and insurers since 2021, long before on-prem AI became a talking point. Our platform, Autothought, runs AI agents inside the client's own cloud or data centre and links every decision to its source evidence, so it can be replayed for auditors and regulators.
| Problem | How Autothought answers it |
|---|---|
| Data sovereignty | Runs inside your cloud or on your own servers. Nothing leaves your environment, and you own your data, models and outputs. |
| Regulatory accountability | Every decision is linked to its source documents and rules, with human review built in and a replayable audit trail. |
| Pilots that never ship | Pre-built applications for document review, research, onboarding and reporting, and go-live in as little as 21 days. |
| AI that doesn't know the business | Agents apply your own business rules and policies, not generic ones. |
| Skills gaps | Our team deploys, monitors and supports the platform with you, so you don't need to build an AI operations team first. |
| Unpredictable cost | Each task is routed to the right model, including our own small language models, which is how clients have cut token costs by 83%. |
The results come from production, not pilots:
| Measure | Result |
|---|---|
| Agent runs completed | 154 million |
| Live institutional deployments | More than 12 |
| Time to an underwriting decision | 90% less |
| Token cost | 83% lower |
Our small-model research is public too. In 2024 we open-sourced 1.5-Pints, a compact language model trained in 9 days that outperformed Apple's OpenELM and Microsoft's Phi-1.5 on the MT-Bench benchmark.
MAS issued its Guidelines on AI Risk Management on 7 October 2026. The first expectations apply from 7 October 2027, and AI life cycle controls such as human oversight, testing and monitoring follow by 7 October 2028. Institutions remain responsible for AI they buy from vendors.
That makes evidence the deciding factor. Autothought produces it by default, inside your own environment. For the full requirements and deadlines, read our explainer, MAS AI Risk Management Guidelines Explained.
Experience Autothought through co-funding of up to S$200,000 from MAS under the FSTI 4.0 AI Pathfinder track (only for FIs onboarded to MAS’s PathFin.ai programme).
On-premises AI runs AI models inside an organisation's own data centre or private cloud, so sensitive data is processed without leaving its control.
Data residency rules, regulatory accountability and vendor risk. Broadcom found the share of enterprises running production AI mainly on public cloud fell from 56% to 41% in a year.
Upfront costs are higher, but at high volumes on-prem can be cheaper. Routing tasks to smaller models also cuts running costs; Pints AI clients report 83% lower token costs.
Yes. Autothought runs inside your own cloud or on your own servers, and you keep ownership of your data, models and outputs.
Pints AI performance figures (agent runs, deployments, underwriting time and token cost) are as reported by Pints AI.