Microsoft is shifting the focus of its platform strategy. At its Ignite 2026 conference, AI agents took centre stage: Azure is to be expanded into the central platform for their development and operation, flanked by new AI models. At the same time, Microsoft and Salesforce announced that they would integrate CRM functions and business processes into Copilot. The obvious question is whether the AI assistant will replace the classic user interface of enterprise software in the medium term.
An agent is understood here as an AI system that not only generates text but executes tasks in multiple steps – retrieving data, preparing decisions and triggering actions in connected systems. The interface in which a person fills in fields and presses buttons is no longer necessarily the main point of access.
Two announcements, one direction
The two announcements describe the same development from different angles. One concerns the infrastructure level: where agents run, how they are built, which models are available. The other concerns the application level: which business processes move from the respective specialist application into a cross-cutting assistant.
Taken together, the two form a pattern that extends beyond Microsoft. If an assistant becomes the primary access layer, individual application interfaces lose importance. The domain logic remains, but it is accessed via interfaces rather than operated through screens. Reliable statements on the timetable, functional depth or availability in individual markets cannot be derived from the information available – caution is advisable here until concrete product details are available.
What changes for enterprise systems
For organisations that operate their own portals, intranets, shops or specialist applications, the planning perspective shifts. Previously the question was: how do we design the interface so that people can work efficiently? In future a second question is added: how do we provide the same functions in such a way that an agent can also use them reliably?
This is less a design question than an architecture question. The decisive factors are:
- Clean interfaces. Functions that are only reachable through the interface remain invisible to agents. A documented API – i.e. a machine-readable programming interface – becomes a first-class access route, not a by-product.
- Structured content. Information that exists only as running text or in PDF attachments is difficult to use. Clearly typed data fields and consistent terminology significantly increase the hit rate.
- Permissions and traceability. When an agent acts on behalf of a person, it must be clear which rights apply, what is logged and where human approval remains mandatory.
- Fault tolerance. Agents interpret. Systems should reject implausible input rather than processing it silently.
Keeping dependencies in view
A provider's agent platform brings convenience, but also lock-in. Anyone who places processes deep within a vendor-specific assistance layer shifts part of the control over user guidance, data flows and further development to that layer. This can make economic sense – provided the decision is made deliberately and not as a side effect of individual pilot projects.
A two-track approach is usually practical: the domain logic remains in your own system and is accessible via open interfaces. The assistant – whether Copilot or another solution – is then one access channel among several, interchangeable and replaceable. This separation costs somewhat more effort initially but preserves room for manoeuvre.
Websites are affected too
The discussion does not only concern internal specialist applications. Public websites are increasingly being read by agents that research, compare and prepare on behalf of users. What is findable, structured and clearly formulated there has a better chance of appearing in such answers.
Content management systems such as TYPO3 are fundamentally well positioned for this, because content there is already maintained in separate fields and records. The lever lies in consistent use: clean content models, robust metadata, structured markup of key information such as services, locations or contact persons – and an API layer through which content is also available outside the page view.
Assessment
For web and software projects, the priority shifts from the interface to the API: what is created today should be usable by both humans and machines from the outset. This does not require an immediate overhaul, but it does require an architecture that consistently separates domain logic, data and presentation. Anyone who establishes this separation now can add agent scenarios later without rebuilding the system – and remains independent of which provider prevails.
Sources
- Microsoft Ignite 26: Azure wird agentischer und Windows rückt in die Cloud — heise developer News
- Copilotforce: Wie KI-Agenten Unternehmenssoftware-Oberflächen ablösen — heise developer News