A new class of AI tools bypasses an old problem of process automation: instead of requiring an interface, so-called computer-use agents operate existing software the way a human would — via the interface, mouse and keyboard.Two current reports show both sides of this development: productive use on a large scale and the countermovement at operating system level.
According to reports, OpenAI intends to use the GPT-6 Astra and GPT-6 Luna models to automate administrative tasks in the US healthcare system.The agents operate existing clinic software and insurance portals like human users.For processes such as prior authorisations — i.e. an insurer's prior approval of a treatment — costs of around 8.5 US cents per process are cited for Luna.The corporations CVS Health and Cigna Healthcare are named as users.
Why this is interesting for integration
The real appeal lies less in the model than in the integration path.Classic automation requires an API, i.e. a programmable interface, or at least a data export.This is precisely where many projects in small and medium-sized businesses fail: the central specialist application is old, the manufacturer does not provide an interface, or enabling it is expensive and time-consuming.
Computer-use agents start one level higher.They read the screen interface, navigate through forms and fill in fields.This means systems that were not technically designed for it can also be integrated.In practice this is a big difference — but it shifts the complexity instead of resolving it.
The flip side: access to everything
An agent that operates an application like a human also needs a human's permissions — and often more.This is exactly where Apple comes in: with macOS 27.3, the permission model is being rebuilt.Anyone who wants to grant an app full disk access in future must click through several warnings.The stated reason explicitly includes AI agents that intervene deeply in the system.
These two reports do not contradict each other; they describe the same phenomenon from two directions.One side shows how far productive use has already progressed.The other shows that platform operators take the associated risks seriously enough to deliberately build in friction.More click paths and more warnings are not an oversight, but a deliberate brake before far-reaching permission grants.
What companies can derive from this
For planning their own automation projects, this gives rise to several checkpoints:
- Clarify the scope of permissions: Which accesses does an agent actually need?Full access to a device is something different from a narrowly defined user account in a single specialist application.
- Ensure traceability: If an agent fills in and submits forms, it must be logged who triggered what and when.This is both a technical and an organisational question.
- Factor in platform dependency: Operating systems and portals can change access paths at any time.An automation based on interface operation is directly affected by this.
- Define error cases: What happens if a form changes or a process is aborted?Without a clearly regulated handover to a human, silent backlogs arise.
Unit costs are not the whole calculation
The cited amount of around 8.5 US cents per process seems at first glance like a strong argument.However, it describes only the execution costs of a single process, not the total costs of an implementation.Configuration, test runs, monitoring, error handling and ongoing adaptation to changed interfaces are incurred in addition.No reliable comparative figures on this are available from the reports.
There is also the context: the cited use cases come from the US healthcare system and from very large organisations.Their volume justifies an effort that does not necessarily arise with smaller case numbers.For small and medium-sized businesses, the question is therefore less whether the technology works, but whether the specific process has enough volume and enough uniformity.
Interface or agent?
Computer-use agents should not be understood as a replacement for clean integrations, but as an option for cases in which an integration is not available.Where a stable API exists, this generally remains the more robust, more verifiable and cheaper-to-operate path.Where none exists, the trade-off shifts — then the agent competes against manual processing or against no solution at all.
What makes sense, therefore, is a sober stocktake: Which processes are repetitive, rule-based and well documented?Which systems are involved, and which of them offer an interface?Only then does the question of the tool arise.
Classification for web and software projects
For projects, this above all means: interfaces become an interface, even if they were never intended as such — stable, predictable structures and unambiguous form fields thus gain in value.At the same time, the requirements for the permission and role concept increase, because an agent needs its own, narrowly limited account instead of borrowed administrator rights.
Anyone planning automation should consider the permissions and logging part from the outset, and not only once a platform provider like Apple tightens the reins.