AI is changing what companies buy from SaaS vendors

From dashboards to decisions

Ir mantas Beržauskas, Director of Legal Affairs and Organisational Development at Lithuanian Railways Group, has used business-intelligence dashboards for years. LTG’s Power BI dashboards combine information from multiple systems and provide a useful overview.

The limitation is what happens after the chart appears. A manager reviewing financial figures may still need to determine:

Beržauskas used to compare report lines and periods himself. Now he gives the data to AI and uses simple prompts to identify changes and trends. The process has become part of his normal work with numbers rather than a separate AI application.

The dashboard has not disappeared. It has become a starting point for questions it was never designed to answer.

Google is moving in the same direction. In June, the company introduced conversational agents for Looker in preview, allowing users to ask questions inside a dashboard and follow up on the answers. Google is adding a conversational layer to an existing product rather than asking customers to replace it.

That distinction matters. A dashboard can show monthly revenue. A system that investigates an unexpected change in revenue can help decide what to do next.

The cheaper route to a missing feature

When an application lacked a feature, companies traditionally had three choices: ask the vendor to build it, hire developers, or use a spreadsheet. AI has made a fourth option more practical: create a specialised internal tool.

GoDaddy said in April that it was testing internal replacements for small SaaS tools from third-party developers, including in its corporate divisions. The company’s stated goals were to reduce costs and simplify operations. It did not say that it had replaced its entire software stack.

A February survey by enterprise-software developer Retool offers some additional context:

The sample describes people already involved in software development. It does not show how many companies worldwide have cancelled subscriptions.

Helmes sees the same demand from another angle. Sharunas Putrius, CEO of Helmes Lithuania in Vilnius, said clients often have purchased applications that cover nearly all their needs but fail at one specific task or connect poorly to another system.

Those gaps are increasingly filled with AI-generated scripts and spreadsheets. Such fixes can work at first, then become liabilities when the underlying system changes, responsibility moves to another employee, or the script receives data its author never anticipated.

Helmes therefore starts with the business process rather than the code. Its engineers establish:

The code may be easier to produce, but reliability still has to be designed.

In one project, Helmes built an AI assistant that searches more than 50,000 service manuals and answers questions using the organisation’s technical documentation. In another, an assistant prepares structured audit findings from documents while employees review the output. Neither project requires a universal application. Both require a narrow tool fitted to work that already exists.

That is the more important change in the buying decision. Companies do not have to choose between a large SaaS subscription and a fully custom system. They can keep the database, security controls and core business application they already trust, then build a small AI tool on top.

More software, same bottlenecks

TransferGo shows the limit of the “software is now cheap” story. The company operates a regulated payments business serving people who move money across countries, often in different languages and circumstances.

CEO Daumantas Dvilinskas said AI now handles 80% of incoming support requests, with customer satisfaction among people receiving AI responses at 94%. TransferGo has also introduced AI into development, marketing and other functions. Dvilinskas said the amount of software-building work doubled over the past year with roughly the same number of employees.

The company can now put more ideas into the market and test them. One experiment is an application built on Anthropic Claude and connected to TransferGo’s data warehouse. It acts as a personal revenue director, analysing activity across about 1,600 international payment corridors and producing a weekly decision log.

The application recently detected unexpected growth in transfers from Iceland, a market without a dedicated manager. Nobody had instructed the system to inspect Iceland. It noticed the change and identified a possible connection to seasonal employment and work permits, after which TransferGo began investigating the hypothesis.

TransferGo already uses Tableau for business intelligence. The difference is that the AI application can search for meaningful changes without someone first deciding which market deserves attention. The data was already available; the new capability was the ability to investigate it without a predefined question.

The constraint has moved elsewhere. Dvilinskas said infrastructure and the server side are now the bottleneck.

That is why I think the immediate threat to SaaS is narrower than the headline suggests. AI can make a small internal application cheap enough to attempt, but it does not automatically make that application secure, maintainable or suitable for a regulated process. Helmes’ warnings about ownership and changing source systems, and TransferGo’s infrastructure constraint, point to the same missing layer: building is easier than operating.

Where the savings become real

LTG’s less glamorous use cases may be more commercially important than its financial analysis. Its freight division receives about 7,000 customer requests a month. The group uses AI to check responses against 17 quality criteria and identify messages that fail to meet the standard.

LTG is also testing voice-based incident reporting for train conductors. After a journey, conductors must document incidents and work-related situations. In the experiment, they would describe events aloud while the system helps complete the forms.

The company estimates that this could reduce the task from about ten hours to three hours per conductor per week. That is a forecast, not a delivered result.

LTG is not handing every decision to an AI agent. Beržauskas stressed that humans must remain involved when an error could affect railway safety. Drafting a report and controlling train movements carry fundamentally different risks.

The same principle applies to software procurement. A manager who can investigate a financial discrepancy with one prompt may need fewer custom dashboards. A support team using AI to check thousands of responses may spend less time on manual quality control. A developer who can build a small internal application in a few days may no longer need a subscription for a single missing feature.

None of the companies described here has confirmed reducing its SaaS subscriptions. Still, their examples show where the pressure will appear first: at renewal, when a customer can compare a recurring fee with an internal tool built on data and systems it already owns.

My guess is that SaaS vendors will have to defend not only whether their products work, but whether the value they provide is difficult to reproduce. AI is making the interface cheaper to build; the durable premium will sit in the infrastructure, reliability and processes that customers cannot safely improvise.