Anyone starting a business has to make early decisions about software. Which CRM should we use? Where will we handle our accounting? Where will we store our documents? Should we use Microsoft 365 or Google Workspace? Do we already need an ERP system?

Most people answer these questions based on today's requirements: price, feature set, usability, and perhaps integration with other applications. Artificial intelligence adds another question that is likely to become at least as important in the future:

Will AI systems be able to effectively access the data and processes of this software in the future?

Any company that currently spreads its business data across systems that are difficult to integrate will limit its own ability to use AI in the future. That’s why startups need to adopt a new principle: Their software infrastructure should not only be cloud-ready, but also AI-ready—that is, capable of supporting AI.

Those who choose software today are shaping the AI possibilities of tomorrow

Imagine a young company. Customer data is stored in the CRM. Invoices and payments are in the accounting software. Quotes and orders might be in an ERP system, emails in yet another system, and contracts, presentations, and project documents in a fifth location.

For humans, this distinction is already a hassle. For AI, it becomes an architectural issue. An AI assistant could, for example, answer this question:

What outstanding orders do we have with customers with whom we've had relevant email communication in the last three months and who still have unpaid invoices?

To do this, the system needs access to CRM, email, order management, and accounting. If each of these systems offers a suitable interface, this can generally be accomplished. If a key system is largely complete from a technical standpoint, it is precisely the information that makes all the difference that is missing.

This means that the quality of a future AI solution does not depend solely on the model. It also depends on whether the AI has access to the right data in the first place.

The problem isn't on-premises

It would be too simplistic to label cloud software as modern and on-premises software as problematic. An on-premises application can be highly integrable. The key factor is whether it has, for example, a documented API or other controllable interfaces.

A few examples:

  • Odoo can be hosted or run on your own infrastructure. The official documentation explicitly describes on-premises installations.
  • ERPNext can be self-hosted. The underlying Frappe framework provides REST interfaces for data objects by default.
  • SAP continues to describe SAP S/4HANA as an ERP system that can also run on the company's infrastructure.

So on-premises isn't the actual problem. The problem is a closed architecture. The issue arises when an application runs on an internal server but offers neither a usable interface nor controlled access to its data. An external AI system then cannot access this information.

That might even be intentional. A completely isolated system sometimes makes sense for security or operational reasons. In that case, however, you’ll need to determine early on in the architecture phase how AI should still be implemented: via controlled middleware, an internal interface, or a locally operated AI model.

So the question isn't "Cloud or on-premises?" It is: To what extent do other systems have controlled access to our data and functions?

The API is becoming a purchasing criterion

An API—put simply, a standardized interface between software systems—has long been a topic of interest for developers. For founders today, it is a strategic feature of software.

HubSpot is a good example. Its APIs allow you to integrate contacts, companies, deals, and tickets, among other things. It also offers webhooks for event-driven integrations and an MCP server that enables compatible AI systems to access parts of the CRM. That doesn’t mean HubSpot is the right solution for every company. But it does show which factors matter when making a modern software decision.

Not just, “Can the CRM manage my leads?” but also: Can I later automatically read, link, and—where appropriate—modify my data in a controlled manner?

CRM: Customer Data Must Not Become a Dead End

For many companies, the CRM is likely to become one of the most important sources of data for AI. It contains customer data, sales opportunities, activities, notes, and often the entire history of a customer relationship. An AI assistant could, for example:

  • Preparing for Sales Meetings
  • Summarize Customer Histories
  • Identifying Follow-ups
  • Prepare quotes
  • Analyze Sales Opportunities
  • Link information from other systems to customer data

To do this, the CRM must make its data accessible. HubSpot offers extensive integration options, while Odoo connects CRM and ERP and can also be run on your own infrastructure. Here, too, the key factor is not the hosting model, but the system’s openness.

ERP: This is where a large part of the reality of business lies

This issue becomes even more important when it comes to ERP. It encompasses orders, products, inventory, suppliers, projects, services, invoices, and other key business processes. Anyone who truly wants to integrate AI into business processes will find it nearly impossible to do so in the long run without this data. For example, an AI system could recognize:

  • Which orders are delayed?
  • which customers have orders on hold
  • how the order volume is trending
  • which projects exceed their budget
  • what information needs to be compiled to make a business decision

Odoo and ERPNext demonstrate that self-hosted systems can be API-enabled. SAP shows that even large enterprise systems can run on-premises. While SAP would be overkill for a startup in most cases, it still serves as a good example: on-premises deployment and integration capabilities are not mutually exclusive.

Accounting: Swiss Companies Should Take a Closer Look

In Switzerland, bexio and Abacus are widely used. bexio provides a documented REST API that allows external applications to exchange data with the system. Abacus also offers REST APIs and web services. In addition, there is a cloud-based version called AbaWeb, while Abacus installations can also run on in-house servers.

This shows that it’s not enough to simply ask the vendor whether the solution is “cloud-based” or “on-premise.” You need to look at the specific product version and its interfaces. It’s also worth asking which data is actually available via the API. Just because an API exists doesn’t mean that every feature of the software is accessible through it.

Email Is Becoming an Important Source of Knowledge

A significant portion of a company’s knowledge is still stored in emails. Google provides an API for Gmail that allows authorized applications to read, organize, or send messages, among other things. In addition to its cloud offerings, Microsoft continues to offer Exchange Server for on-premises deployment. Microsoft explicitly refers to the current Exchange Server Subscription Edition as an on-premises email solution.

Architecture matters here, too. If emails are to become part of an AI-powered knowledge system later on, you need to consider access, permissions, and interfaces from the very beginning.

Documents are often the largest archive of unstructured data

Proposals, presentations, concepts, contracts, meeting minutes, technical documentation, and internal guidelines contain a large portion of a company’s knowledge. That’s exactly what makes them so interesting to AI. Imagine if your employees could ask:

"What did we promise this customer in our last three proposals?"

or:

"What decisions have been made so far regarding this project?"

For this to work, the AI must be able to find relevant documents and, depending on the individual’s permissions, read them. Here, too, there are different architectures. In addition to the major cloud platforms, Nextcloud, for example, can be run on your own infrastructure. Nextcloud explicitly describes its offering as on-premises-capable and supports the integration of various external storage solutions.

The real problem: Corporate knowledge is scattered

Even if every individual system works well on its own, another problem arises as the business grows: knowledge becomes scattered. Some of it is stored in the CRM, some in emails, some in documents, in the ERP system, and in accounting. Later on, support systems, project management tools, a data warehouse, and specialized applications may be added.

A person switches back and forth between these applications. An AI agent, on the other hand, needs technically controlled access to every source of information. That is why you should view the software architecture as a cohesive system from the very beginning.

The goal is not to move all data into a single application. The goal is for systems to be able to communicate with one another and to ensure that data does not end up permanently trapped in isolated silos.

MCP is interesting, but a good API is more important

When it comes to AI, the Model Context Protocol—or MCP for short—is a term that comes up frequently these days. It can simplify the connection between AI systems and external applications. HubSpot already offers its own MCP server, and other providers are working on similar solutions.

Still, you shouldn't make MCP the sole deciding factor when choosing software. Technologies change. A well-documented API, a good authorization model, and an open data architecture are more fundamental. If you have this foundation, you'll be able to integrate future AI technologies as well.

Beware of Vendor Lock-in

When a company is first founded, many software decisions seem reversible. Later on, they often are no longer. After five years, there may be tens of thousands of contacts in the CRM, hundreds of thousands of transactions in the ERP, years’ worth of correspondence in email systems, and large volumes of documents on a single platform. At that point, switching becomes expensive.

That’s why every major software decision should include the question: How do we get out of this later? A good export feature is not the same as a good API, and both are important. The API enables ongoing integration and automation. The export feature ensures that your data isn’t permanently locked into the system. Check early on whether you can export your data completely, in a structured manner, and in a reusable format.

An API alone does not make software AI-capable

There's more to the phrase "API available" than meets the eye. Before making a decision, you should consider at least these seven points:

  1. What data is available through the interface? It's not much use if the API exists but the relevant information is missing.
  2. Is the API read-only, or does it also support writing? Read access is sufficient for analysis. Agents that are supposed to perform tasks may need controlled write access.
  3. Is there a reasonable authorization system? An AI shouldn't automatically have access to all data just because a technical interface exists.
  4. Does the platform support automation? Webhooks and similar mechanisms allow applications to respond to events without constantly retrieving all the data.
  5. Are there any limits? APIs may have different volume limits depending on the provider, license, and pricing plan.
  6. Is the interface documented and actively maintained? In practice, a poorly documented interface results in high integration costs.
  7. Can data be exported in its entirety? This question is relevant regardless of AI and reduces dependence on the provider.

Checklist for Entrepreneurs

Before you implement a CRM, ERP, accounting, email, or document management system, I would ask at least these questions today:

  • Is there a documented and maintained API?
  • What key data is available through this API?
  • Can data be read and, if necessary, written?
  • Are there webhooks or similar automation options?
  • Is there a finely tunable role- and permission-based system?
  • Can technical users or applications be granted their own access rights?
  • Can all company data be exported, organized, and reused later?
  • What restrictions apply to the API?
  • Does API access depend on a specific license?
  • Can the solution be integrated with automation platforms or middleware?
  • Are there already integrations with major AI platforms?
  • Does the manufacturer support open standards?
  • How much work would it take to switch to a different system later on?
  • Can you continue to use your data regardless of the provider?

And finally, the question it all comes down to: Could a future AI agent gain controlled access to the information you’re currently storing in this system? If you can’t answer that, make sure to clarify it before you buy.

"Cloud-first" is no longer enough

In recent years, many startups have followed the “cloud-first” principle. This made sense: Cloud software reduced infrastructure costs, simplified updates, and allowed young companies to get off to a quick start. With AI, the requirements are changing once again. Today, the better question is: Is our infrastructure AI-ready?

That doesn’t mean every company has to start building AI agents right away. Nor does it mean that every system belongs in the cloud. It means that when making decisions today, you need to consider how your data will be used in the future. Data doesn’t become valuable just because you collect it. Its value arises when it is discoverable, linkable, and accessible for the right applications.

The software decision becomes a data decision

It’s tempting to choose software based on what seems easiest or most affordable today. But such a decision can have consequences that last for ten years. When it comes to core business systems, therefore, it’s not enough to ask about the current range of features. Just as important—if not more so—is the question: What can we do tomorrow with the data stored in the system?

In a nutshell: Companies whose CRM, ERP, accounting, email, and document management systems can be seamlessly integrated don’t just have a good IT infrastructure. They lay the groundwork for future AI systems to truly work with the company’s knowledge and processes.

What software are you using today, and do you know if an AI could access your data tomorrow? Tell me about your experiences—I'm curious to hear where you've already come across closed systems.

Privacy Notice: swissAI addresses the types of corporate and personal data you are permitted to transmit to external AI systems, as well as the applicable requirements, in a separate article titled “Using AI Safely Without Unnecessarily Limiting Its Capabilities.”