Artificial intelligence relies heavily on the cloud and a few dominant players. But I anticipate that emerging technical possibilities will make it possible to use AI outside of this context, since it is now possible to run AI locally on your devices, what we’ll call, for simplicity’s sake, “On-Device AI” or “client-side AI.” Personally, I like the image of “AI in our pocket” This development will be a game-changer for Procurement, at the intersection of sovereignty, cost, and compliance In fact, thanks to the WebGPU API, your browser can run models directly, without any additional infrastructure, by leveraging the GPUs in PCs or smartphones to run them. This means having unprecedented computing capabilities, a major innovation, that can be leveraged directly by your client-side Procurement SaaS application. Your workstation is no longer merely passive: it becomes a computing and processing aid.
What benefits can we expect? Let’s examine this from two perspectives:
1. Users of our applications, 2. The Procurement department, which also consists of users but must also develop a “360-degree” AI strategy. For users, the introduction of on-premises AI opens the door to an enhanced user interface, as AI becomes a native layer of the interface. Runtime locally, without network latency and at no additional cost, new services can be made available, such as: 1. Generating a summary of a screen or a message, 2. Translating on the fly. This is particularly useful in an international context, where a provider can fill in text fields in a language unfamiliar to other users; 3. Improving text quality (grammar, rephrasing, clarification, simplification); 4. Analyzing images… These “micro” services change how users perceive the application by bringing it closer to them. But above all, and this may be the most important point, it prevents uncontrolled copy-and-paste of data into open-access generative AIs without data protection. The infamous “shadow AI”(*):
For Procurement, I see three natural benefits:
1. Reducing dependence on “hyperscalers” Local execution helps limit the transfer of sensitive data to external infrastructures and reduces exposure to the Cloud Act. This is a way to regain control over critical data while facilitating compliance with the GDPR and the AI Act. For many organizations, the issue is no longer technical, it has become strategic. 2. Rethinking the total cost of AI AI running on end-user devices is transforming the business model: fewer server GPUs to provision, lower cloud OPEX, and significantly reduced latency. Every device becomes a computing resource. The marginal cost of inference can approach zero. This is a key factor in managing budgets, trade-offs, and scalability. 3.Better Protect Business Data Purchasing, legal, finance, HR… Certain operations must remain within the company. Local execution ensures that data no longer passes through third-party servers. No external storage, no unintended transfers, and enhanced traceability. High-performance AI and privacy are no longer incompatible Yeah, but that’s all well and good… but it’s not happening anytime soon. True and false. In fact, it’s already here—but not in all browsers… and not yet in SaaS procurement applications. If you’d like to test how it works right now, I invite you to visit this page:https://chrome.dev/web-ai-demos/translation-language-detection-api-playground/ This is an example of an online translation performed locally. 
Once the page has loaded, disconnect from the internet and test the native local translation capabilities. Magic! Note: This requires a recent version of Google Chrome, starting with version 138 released on June 24, 2025
Summary: A Procurement Challenge Before It’s a Technical One
“On-Device AI” opens up a new avenue for optimization. But does that mean it’s a replacement for cloud-based AI? I don’t think so. At least not in the short term. The size of the models and the performance we can expect from them are likely to differ for quite some time. Furthermore, while some models can be natively embedded in browsers, others may require a prior download, which could be slightly less convenient for occasional users.The renewal of the PC fleet, accelerated by the end of support for Windows 10 (October 14, 2025) and the transition to Windows 11, will automatically raise the overall performance level of the fleet and encourage the adoption of more powerful machines capable of running increasingly complex models locally.However, I believe that a hybrid architecture combining “on-device AI” and cloud-based AI should be considered depending on the use case. The choice will likely be made first and foremost by the service provider. At Oalia, we believe this opens up new, complementary possibilities that address challenges that are sometimes difficult to reconcile, such as: • Compliance and governance (AI Act, internal policies, auditability) • Data control (contractual arrangements, localization, reversibility) • Sovereignty and resilience (reducing vendor dependency) • TCO management (direct and indirect costs) Oalia has decided to explore this path in parallel with our work on cloud AI. We’ll discuss this further. (*) Shadow AI refers to the use of artificial intelligence tools and applications by employees or departments without the approval or supervision of the organization’s IT department. 