Why AI coding agents won’t kill SaaS

Accelerating the software lifecycle

Coding is definitely faster with AI, according to Varsavsky, and SiteMinder is seeing a massive acceleration of what it can do with fewer people. “As a SaaS provider, we’re taking full advantage of that. We’re not sitting still waiting for somebody to out-develop us. We’ve got a good head start and we’re leveraging those assets to do even more, even faster.”

In particular, he has seen a massive acceleration in AI coding since the release of Claude Opus 4.5 last November. “Over the last six months, the amount of code that we’re using is doubling every month.”

“We’re finding utility in all parts of the software lifecycle,” he added. Some initiatives that would previously have taken months are now being done in weeks, and others that would have required a full team can now be completed by two or three people.

When it comes to innovation, SiteMinder product managers are using AI tools to code prototypes before the ideas go to engineering, so customer feedback can be collected earlier in the process. “That's really been useful,” he said.

I’m not spending any time thinking about how to replace Salesforce with my own vibe-coded solution. I’m spending all my time trying to figure out how my engineers can now go faster with AI to deliver our roadmap so we can grow the business

CTOs often struggle with tech debt and managing complexity, Varsavsky observed. A lot of the repetitive tasks that need to happen across a big fleet of software – such as version upgrades or refactoring – are much easier to do with AI because one person can code a recipe that can be easily applied to every codebase.

Normally, asking every team to update to the latest version of a particular library takes them away from providing new value to customers and other important activities. Instead, a principal engineer in the developer experience team spent a couple of weeks building and testing a Claude skill for that task for the rest of the company to use. The provision of that skill allowed the upgrade to be completed an order of magnitude faster than would otherwise have been possible.

Because the developer experience team owns the tooling and processes used by the other engineering teams, a dollar invested there yields $10 in benefit among the other teams, thanks to the increased safety and speed, he noted.

“It means that our hygiene factors will get a lot easier to keep up with, [and] we all know that a clean kitchen makes good meals,” he said.

Another example of AI-powered development is that one of SiteMinder’s engineering managers has used Claude Code to build a dashboard that combines data from sources including GitHub, Jira, Confluence, email and Slack to reveal their team’s progress, where they are blocked, and who needs help.

The rise of agentic coding

While some companies proudly announce that parts of their operation have stopped doing manual coding, Varsavsky regards this as a “vanity metric” largely used by self-interested companies that see AI coding as the panacea for software development.

Developers have always used tools to build software, and AI is just one of those tools, he said. Twelve months ago, people used it as an assistant: they would write the code, and the AI would suggest changes. But “over the last six months, the workload’s definitely switched to more agentic coding and less hand coding.” 

With Claude Code, the developer mostly interacts with the chatbot, while the code is generated in the background and inspected infrequently. “I find that when I'm coding with Claude, I'm spending less than 10% of the time looking at the code, and 90% of the time I'm interacting with the agent.”

But at the end of the day, the engineer driving the AI is responsible for the outcome. SiteMinder still has quality assurance (QA) processes including code reviews and other hurdles, but those processes are being accelerated by applying AI, saving half an hour per day per person.

“One of our key value propositions is the robustness and reliability of the platform; we just can’t afford to lose reservations so we take a more cautious approach to that than maybe you would if you were a startup with no customers,” Varsavsky said.

Mastering token economics

The rising cost of using models such as Claude has been a hot topic recently. “I think token economics will be a big topic for CTOs going forward,” Varsavsky said, drawing a parallel to how tech leaders had to learn to manage the use of cloud computing after decades of on-premises hardware. “We got better at managing variable costs and consumption costs. I think the same lessons, successes, and patterns will apply with tokens.”

“Right now, we’re still very much focused on adoption and getting the most out of AI while sustaining our quality and security standards. This is currently a priority over the costs involved. We’re seeing the numbers double every month, making Claude one of our top five vendors, and in many ways, we’re just getting started. While these costs continue to grow, our focus moving forward will soon transition to optimisation and measurement to ensure we're maximising value from every token,” said Varsavsky.

Next year will be about getting value for money, he predicted. “We know that some of our tokens are not going to the most valuable outcome, but I can't tell you which ones.”

The focus will shift to measuring productivity – something that has historically been difficult because software engineering involves a lot of one-offs, unlike a manufacturing environment that produces widgets by the thousand. “The industry hasn’t come up with a universal measure of engineering productivity just yet, and I’m not sure we will with AI either. So, it will rely on leadership judgment in the end, but we need some data to back that up.”

“We’re getting better at the metrics side of it... we’re investing in the analytics and the metrics that will show us the difference between various teams, or people that use or don't use the tools in different ways,” he said. This is partly for cost control reasons but also serves as a way to amplify successes across the wider company.

Fostering AI adoption and setting guardrails

Returning to the matter of AI adoption, “technology people have always been agile learners,” he observed, meaning teams are naturally adopting AI tools. “My job as a CTO is to amplify that internally,” for example through the developer experience team, which has the task of triaging the latest tools, figuring out what works for the company, deploying them, and then encouraging widespread use.

Other ways SiteMinder is promoting adoption include a guild of champions that share successful recipes, an internal Slack channel to highlight the latest developments, and a company-wide AI hackathon held in 2025.

Some people are less receptive to the change to agentic coding, whether that’s because they believe that AI isn’t going to give the best result or because they are complacent, he observed, but “we focus very much on getting everyone there,” dealing with the laggards on an individual basis.

The “feature factory” part of the team that delivers new features to customers has already recorded examples of working 50% faster with AI, and Varsavsky expects to see smaller, faster teams delivering higher throughput in the future.

But to be successful, they need a robust platform. “That means consistent architecture, easy to deploy, easy to develop, easy to keep secure, and QA guardrails” to deliver quality at speed.

“That part of our team [responsible for the platform] is growing because that becomes a force multiplier for the rest. If we can do all those things consistently and well, then we can let the feature teams go really fast with less supervision, knowing that we’ve got the guardrails in place that mean we’re not going to get into trouble,” he said.

“In the end, businesses that provide good value for the customers at a reasonable cost will win. That's what we’re focused on. In technology shifts, there’s always risk and opportunity, and the companies that come out of this well will have taken up the opportunity.”