Adobe has announced that it will open up Photoshop and Premiere to external AI assistants. Systems such as ChatGPT or Claude will in future be able to operate the programs directly, instead of users carrying out every step manually in the interface. A detail explicitly mentioned in the announcement is decisive here: the project files remain editable.

What is new about the announcement

Generative AI in image and video tools is nothing new in itself. Until now, it has generally run as a function within the respective application – you click a button, describe the desired result, and get a result. Control remained with the person operating the software.

The announced step reverses this direction. An AI agent – that is, an assistance system that not only generates text but independently calls tools – accesses the application from outside and carries out work steps there. Photoshop and Premiere thus become tools that a model can operate, similar to how it already addresses file systems, search services or development environments today.

Why editability is the key point

For day-to-day production, the reference to editable project files is more important than the automation itself. Pure image generators deliver a finished pixel result: if the colour is wrong, the logo is too small or the cut-out edge causes problems, often the only option is a new attempt with a changed prompt.

If, on the other hand, the project file with layers, masks or edit tracks is retained, a machine-generated intermediate state can be processed further just as before: corrected, replaced, approved. The AI then takes over the mechanical part, while content control remains in the familiar tool. For agencies and marketing departments that work with correction loops and approvals, this is a relevant difference.

Typical use cases in the web environment

Especially with web and campaign assets, a lot of work arises that is technically undemanding but time-consuming. This includes, for example:

  • deriving an image idea in many formats for different placements
  • recurring preparation steps for product images according to a fixed scheme
  • video edits in several lengths or aspect ratios for different channels
  • replacing individual elements such as texts or price information in existing templates

These are precisely the tasks where an agent can sensibly start, because the rules are clear and the result remains verifiable. Specific information on which functions will be accessible from outside and to what extent is not yet available – here it is worth waiting for the technical documentation before restructuring processes.

Open questions before productive use

Opening up to the outside also shifts responsibilities. Who checks whether an automatically generated asset complies with brand guidelines? What data leaves the company when an external assistant gains access to project files? And how is it documented which work step was machine-generated and which was manual?

These questions are not new; they arise with every agent integration. However, they gain weight as soon as a model not only makes suggestions but actually changes files. A step-by-step approach is therefore sensible: first clearly defined, repeatable tasks, with human approval before publication.

Assessment for web and software projects

The announcement fits into a development that has been emerging for some time: applications are becoming operable not only for people but also for machines – via interfaces that agents can use. For projects, this means that asset production should in future be conceived more strongly as a process chain, with defined handover points between automated and manual steps. Those who have already structured templates, naming conventions and approval paths cleanly will benefit fastest – because automation reinforces existing order just as reliably as existing chaos.

Sources