In short: AI agents rarely replace established enterprise software. They build on top of it. That raises the value of the data stored in those systems. But that value goes to whoever owns the day-to-day relationship with the customer. If a third party takes over that relationship with its agent, the software underneath becomes interchangeable infrastructure. Every software company therefore has to decide deliberately where it wants to stand in this new architecture.
What’s the difference between systems of record and systems of action?
Systems of record are the systems where a company keeps its authoritative data: the ERP with transactions and orders, the CRM with customers and contracts, the service system with the equipment history. They capture what applies, who is allowed to do what and what happened when. Systems of action are the systems that act on that foundation: agents that plan, decide and execute.
The result is a three-layer architecture.
Bottom: system of recordwith structured data, audit trail and compliance.
Middle: context layerwith users, permissions and business rules.
Top: agentic layeragents that plan, decide and execute.
For vendors of strong systems of record, that’s good news at first.
Their data is the foundation agents need to work reliably.
But there’s a catch. Just because a foundation is needed doesn’t mean its owner profits from the building.
What does “headless SaaS” mean for software vendors?
In investor circles, the danger now has a name: “headless SaaS.” It means software without a face of its own. Its data is still valuable, but it’s only accessed through its API, by systems that belong to another vendor. To the person responsible for the work, it’s invisible.
The mechanism behind it is simple. If you’re invisible, you get paid for your costs and hardly for your value. The customer sees the benefit the agent delivers and credits the agent with it. What sits underneath looks like a technical prerequisite to be bought as cheaply as possible. Gartner describes the same movement from the customer’s point of view: agentic systems deliver results directly, the software behind them becomes invisible, and the user interface is no longer a differentiator (Gartner, July 1, 2026).
The industry has seen this pattern before. In the early 2000s, many vendors of installed software allowed SaaS challengers to establish themselves on top of their products. Most of them faded in importance without ever having consciously decided to go down that road. Back then, doing nothing was a decision nobody made. It is again today.
How can an executive team tell that the customer relationship is slipping away?
I find three signals revealing, and none of them requires an elaborate analysis:
- The share of access through APIs rather than the interface, and above all how it develops quarter over quarter.
- Requests for API documentation with no apparent reason, especially from large customers.
- The customer’s language: When an IT director talks about her processes and names another vendor, the relationship has already moved, even if the contract is still running.
Does the user interface still have a future?
Yes, but with a different job. The interface is changing from a workstation into a control room. Companies won’t trust fully autonomous agents from day one. People need to review, approve or correct suggestions before they take effect, and they need a place where they can see what an agent is proposing, why and how confident it is. The Bitkom study “Softwarewelt 2036” reports that the executives surveyed expect agent actions to be logged in a traceable, audit-proof way.
Vendors whose software is already their customers’ everyday work tool have a head start here. They can embed agents into existing approval steps, including other vendors’ agents. A control room where suggestions from different agents come together creates a kind of lock-in no outside agent can replace, because it sits where the data and the accountability are.
What do software companies need to rebuild now?
APIs. Most APIs were built for integrations, not for agents. An agent asks many small questions in an unpredictable order and needs not just the data but its meaning. The yardstick: an agent must be able to do everything through the API that a person can do through the interface. Open standards such as the Model Context Protocol (MCP) are becoming table stakes.
Contracts. Many standard contracts define a user as a natural person and say nothing about agents. Vendors that deny agent access or price it out of reach will be screened out of the next tender. The first to offer fair, understandable rules gains negotiating leverage. A plausible model keeps access itself cheap or free and charges for the services actually delivered through it.
Visibility in AI search. The search for software is shifting, too. According to the SaaS & AI Pricing Report 2027 by hy, OMR Reviews and Appinio, a quarter of software buyers in the German-speaking region already use AI systems for research, and 92 percent of B2B buyers end up choosing a product that was already on their first shortlist (report at OMR Reviews). If you’re missing from an AI system’s first answer, you often don’t even get to negotiate.
Infrastructure, your own agent layer or orchestration?
Sooner or later, every established software company reaches the same fork in the road. I distinguish three options:
| Criterion | A: Infrastructure | B: Own agent layer | C: Orchestration |
|---|---|---|---|
| Investment required | low | medium | high |
| Short-term impact on earnings | positive | negative | strongly negative |
| Control over the customer relationship | low | high | very high |
| Share of the value of work done | low | medium to high | high |
| Demands on the organization | low | considerable | very high |
| Fits when … | your own layer dominates the market and the data is irreplaceable | your own data and industry expertise make the agent better | a company wants to shape an industry as a platform |
Option A is not a failure. Payment processors and data center operators are highly profitable infrastructure. But they are because they set the standards in their layer. That rarely applies to a mid-sized niche vendor. Option B costs earnings and requires an organization that can do two things at once. It explicitly does not mean locking out other vendors’ agents. Option C is a multi-year goal for most mid-sized companies, not a measure for next quarter.
No option comes free. A costs the future, B costs earnings, C temporarily costs both. In practice, a combination works best. Your own agent layer pays off where your own data provides a real advantage. Outside agents get fair access where customers want to use them, and the next steps build toward orchestration.
Entering from the edge deserves special attention. Instead of rebuilding the existing software, a vendor launches a standalone agent that needs little integration, shows results quickly and gradually connects to other systems. That way, a vendor can gain a foothold even with customers whose core system comes from a competitor.
What should executive teams review now?
- Think about who owns the relationship with the people in charge if your three largest customers put an outside agent on top of your data tomorrow.
- Track what share of access runs through APIs rather than the interface, per key account and quarter over quarter.
- Run a test and ask three AI assistants about the leading vendors in your category. Do you show up?
Frequently asked questions
What is headless SaaS?
Headless SaaS refers to software that is only accessed through APIs by outside systems, such as AI agents. The data remains valuable, but the relationship with the user sits with the agent’s operator, which leads to price pressure and weaker customer retention.
Are AI agents replacing existing enterprise software?
Mostly not. Agents need reliable data, permissions and business rules, and those live in existing systems of record. But they change who uses the interface and who owns the customer relationship.
Should software vendors lock out outside AI agents?
No. When choosing software, customers increasingly check whether agents can work without restrictions. A fair, contractually governed access model in which the services delivered are paid for makes more sense, combined with your own agents where your own data provides an advantage.
Further reading: Is SaaS dead? Why that’s the wrong question · From tool budget to headcount budget: pricing for AI agents · Managing two curves at once: leading the transformation
Sources: Gartner press release, July 1, 2026 · Bitkom (2026): Softwarewelt 2036 · SaaS & AI Pricing Report 2027 (hy, OMR Reviews, Appinio), report at OMR Reviews