From the chat window to agents in the process
The first wave of AI in business software took the form of a chat window next to the application. Customers now expect the software itself to handle tasks: an agent retrieves data, makes a decision within established rules, and carries out the action in the system.
That sets the bar higher for your architecture than a standalone AI feature. An agent can only act through the interfaces you provide, based on accurate data, with permissions that you can restrict and control. A monolithic application without clear APIs doesn’t offer that. According to SUE, that’s the biggest obstacle to agentic AI in existing software: not the model, but the application surrounding it.
1. From Monolithic to Modular
We break down a monolith into building blocks with clear boundaries that communicate with each other via REST or gRPC APIs. Each building block can be updated, scaled, and tested independently. For an agent, these APIs serve as tools: it calls a well-defined function instead of searching a database itself.
It doesn't have to be a "big bang" approach. We start by separating out the components that change most frequently or offer the greatest value for AI, while the rest continues to run. At CarWise, we used this approach to decouple the user interface from the business logic, laying the foundation for a scalable SaaS platform.
2. Customer data is strictly separated, with a single source of truth
A SaaS application serves multiple customers (tenants) on the same platform. Whenever an agent acts on behalf of a user, it must be ensured that the agent can access only that customer’s data. We therefore design the data structure so that tenants remain strictly separated, both in the database and in every API.
In addition, an agent needs a single, reliable source of truth: a system of record. If the same customer data appears in three different places with three different values, the agent will make a decision based on incorrect information.
3. Safeguards for Every AI Action
B2B customers want to know what an agent does with their data and be able to review what the agent has done afterward. With Retrieval-Augmented Generation (RAG), we link the domain knowledge from your own application to the model, so that responses and actions are based on your own data. Guardrails determine what an agent is allowed to do: which actions, on which data, up to what limit, and when human approval is required first.
Every action is logged so that you can explain what happened. Transparency and logging are also requirements under the EU AI Act. Guardrails should therefore be built into the architecture, not added as a separate layer later on.
4. A usage-based revenue model
When an agent takes over work that was previously done manually, a customer will eventually need fewer user accounts. A per-user licensing model will then shrink accordingly. Software companies are therefore switching to pricing models based on usage or output: per request handled, per document processed, or per transaction.
That’s only possible if your software tracks what’s happening. That’s why metering and telemetry are integral to the architecture: which customer uses which feature, how often, and at what cost. You also need that data to monitor the costs of AI models and cloud services per customer, so that a new AI module doesn’t eat into your margin.
Where do you start?
Using the code you already have. With Re:App, SUE analyzes existing software using AI: architecture, dependencies, and business logic—even when documentation is missing. At Deltares, this involved 6 million lines of code spanning 40 years of history. The analysis reveals which components need to be modularized first and which can remain as they are.
If your application still runs as desktop software on users’ PCs, the transition to a web application is often the first step. We use Multistax as the underlying platform, so that developers can deploy new features themselves and keep track of cloud and AI costs.
To learn how SUE handles this, visit the page Application Modernization and in “Our Approach ” with AI. For software companies, more information is available at Application Modernization for ISVs.
