Artificial intelligence in web projects has so far almost always been a question of integration: a service is called via an interface, the request leaves the computer, the response comes back.This very pattern is now in motion.Local language models – that is, models that are executed in the users' browsers themselves – move processing to where the data is created in the first place: into the frontend.
The browser thus becomes a kind of runtime environment for AI functions.Contextual support, meaning help directly at the place where someone is currently working, is created without a detour via an external service.In principle, this also works offline, and no inference costs are incurred – that is, no fees per model request.
Why this is relevant for decision-makers
Three points make the difference compared with the previous approach, and all three concern less the technology than the framework conditions of a project.
- Data protection: Content processed in the browser does not have to leave the computer.For forms, internal tools or editorial interfaces with sensitive content, this significantly changes the starting point of a data protection impact assessment.
- Cost structure: Without requests to an external service, the ongoing costs that grow with usage are eliminated.The effort shifts into development and into the delivery of the models.
- Availability: A function that computes locally does not depend on the reachability of a provider.This reduces a dependency that was tacitly accepted in many AI projects of recent years.
What local models can sensibly achieve
The realistic expectation is not to replace a large cloud model one-to-one.Models that run on end devices are smaller, and their strength lies in clearly defined tasks close to the context of the interface.Typical candidates in web projects are:
- Formulation aids and rewording in editorial and input forms
- Summaries or outline suggestions for texts that are currently being edited
- Classification and tagging of content as a suggestion for the editorial team
- Search functions that interpret an input semantically rather than merely literally
- Assistance functions in web apps that structure user input before it is saved
What these cases have in common is that they are short, that they also work without AI if in doubt, and that a failure of the model causes no damage because a human checks the suggestion anyway.
Naming the limits openly
A model that runs in the browser has to get there.This means data volumes that must be loaded and sensibly cached, and it means a dependency on the computing power of the end device.A current desktop computer behaves differently from an older tablet.Anyone planning local AI functions therefore inevitably also plans for a case in which the function is not available.
This leads to a design rule that is familiar from web development anyway: AI functions in the frontend should be conceived as a supplement, not as a foundation.The page, the form, the editorial process must remain fully usable without the model.Anything else leads to an application whose usability depends on the hardware of the visitors.
Consequences for CMS and TYPO3 projects
For content management systems, the development is particularly interesting because two worlds meet there.In the backend, editors work with content that is often still unpublished and sometimes confidential.Especially here, the question of which data an external service may see was often the reason why AI support did not get beyond the prototype.
Local execution shifts this discussion.Instead of an approval decision about data outflows, the focus is on technical questions: How are models delivered and versioned?What are the requirements for the editorial team's desktop computers?Where exactly in the interface does real benefit arise, and where would a suggestion function be merely an additional click?
In the frontend, that is, on the delivered website, the standard is stricter.Load time and accessibility take priority, and the benefit of an AI function must justify the additional resource requirement.For an internal web app with a known user group, this trade-off turns out differently than for a public campaign page.
What a pragmatic start looks like
- Select a concrete work step that costs time today and for which a suggestion is sufficient – not a general assistance function.
- Check which data is processed in the process.The more sensitive, the stronger the case for local execution.
- Clarify the target devices.Internal use with known equipment is the easier start.
- Define the process fully without AI before the function is added.
- Roll out small, measure whether the step actually becomes faster, and only then expand.
For web and software projects, the shift into the frontend above all means a new option in the architecture: AI functions can in future also be implemented where transmission of data to external services is out of the question.This does not automatically make projects simpler, but it decouples them from ongoing usage costs and from the availability of individual providers.
The sensible approach is where the task is clearly defined and the result is checked by a human anyway.
Sources
- LLMs in the browser – Part 1: A new AI runtime — heise developer News