The SaaS apocalypse is overblown. AI just raised the bar.

Today Productive is a SaaS platform that helps agencies and service companies stay profitable and organised, covering time tracking, planning, budgets, reporting and task management. The company is bootstrapped, around 150 people strong and has over 40,000 users. Its latest release, Productive 5, adds an AI notetaker and customisable agents.

We got into whether the SaaS apocalypse is real, what changes when you build non-deterministic software, how to get customers to actually use AI agents, and why agency side products rarely make it without a carve-out.

Watch the episode on YouTube

Key takeaways#

Is AI killing SaaS?#

Tomislav doesn't see it. The "SaaS apocalypse" story says companies will vibe code their own Salesforce, and he isn't seeing people go off and build their own Productive. Agencies have always had a version of this: developers on the bench build an internal tool, and years later it has stopped being maintained, there's a bug, and the person who wrote it quit a year ago. Builders will prompt simple tools out of Claude Code, and for a basic CRM that's fine. For complex platforms it's still a cost in time, effort and tokens. The bigger shift is the interface. Business software used to be buttons and input fields. Now you talk or type to it and it does the work, through assistants inside the product or through MCP connections to Claude or ChatGPT. Companies doing nothing there, in his view, are in for an apocalypse of their own.

Try this: list the five things your customers do most in your product, and check whether each one can be done by asking for it in plain language.

Should you build or buy software now that AI can write code?#

The buy-versus-build dilemma has always been around. AI changes the maths, since the unit economics are different, but building is still not free. And the people building tools for a living use the same AI. If a company can now build its own tool faster, a company like Productive can build a better one with more functionality and a better user experience. Tomislav expects the bar to rise on both sides, which is what he's seeing right now.

Try this: before you approve an internal build, price three years of maintenance alongside the first version.

How do you build and test non-deterministic software?#

Expect to change how the organisation works. Classic business software is deterministic: you click a button, you know what happens. Agentic software is non-deterministic on two layers. You don't know what people will type into the box, and the model doesn't always give the same output for the same input. Exact unit and functional tests stop covering everything. Inside Productive, the conversations now sound like "Does the agent feel right? Is the tone of voice good?" Tomislav calls it vibey software. It's a shift for engineers used to precise systems, especially at a company that handles money, where every cent has to be right and you can't ship first and fix later.

Try this: add a qualitative review round to your AI feature releases, where someone reads real conversations and scores tone and judgement alongside the automated tests.

How is AI changing product team roles?#

Execution got much faster, and the pressure moved upstream. What used to take two months now takes two weeks. That puts weight on product strategy, product management and design, who have to work out what to build and why. Engineering deals with far more code and pull requests, so review becomes its own problem. Non-technical people now code up prototypes and test them. Tomislav sees the classic trifecta of engineer, product manager and designer moving closer together. He doesn't think it collapses into one role, because the sensibilities are different. He compares it to the old webmaster who did both design and development (G's reply: that was me).

Try this: give one non-technical person on your team a prototype tool and a real problem this sprint. See what comes back.

What is the difference between a system of record and an AI agent?#

A system of record is where data lives, organised, with the business rules. Productive was that, and still largely is. Agents are AIs that go out and do the work: send emails, synthesise information, build presentations. Early on, everyone said systems of record would become irrelevant. Tomislav points out that agents need data from somewhere, and that any agent can connect to a system of record. How those two kinds of software relate is the market dynamic playing out right now.

Try this: decide which one your product is today, and write down what you'd need to become the other.

How should a SaaS company approach the AI transition?#

First decide whether you want to build AI into the product at all, or stay a system of record. Both are fair answers. Building AI across a platform is a whole new investment cycle, and you need the money, the knowledge and a team willing to learn a new way of building. The agents also have to be really good. A bad agent feels like talking to a very dumb person, and people develop an aversion to using it. If you can't make them good, it may be better to offer an MCP or a CLI and let other agents connect to you. Tomislav sees risk in that route too, but it's a strategy.

Try this: answer the build-or-connect question explicitly at leadership level, with a budget attached. Drifting into half an agent is the worst outcome.

How do you get customers to actually use AI agents?#

Teach them the way you'd brief a new hire. Productive's version 5 added an AI notetaker that attaches meetings to projects and deals, customisable agents that can do anything a person can do in Productive, reusable skills and custom reports. Adoption runs on a scale. The notetaker is easy: it joins a call and gives you notes. Assistants are natural too, but people type five words and expect magic. Tomislav's lesson is that all the problems you have with agents, you have with people. Tell a new hire "fix this" with no context and you get mediocre results. Give them a little direction and they do wonders. When customers ask how to use a feature, his team asks what they do every day that takes 20 minutes.

Try this: run that 20-minute question with three customers. Each answer is a candidate for your first scheduled agent, like a rundown every morning at 8am.

What work do people want to hand off to AI?#

The admin and the boring work. Productive's customers are clear that they don't want to hand off creative work, decision making or client relationships. Some don't think AI can do it, and Tomislav agrees. Others just like doing that part themselves. So Productive aimed its agents at the admin layer.

Try this: map one week of your team's work into "admin" and "craft". Point your first automation only at the first column.

How do you control what an AI agent can access?#

Give every agent an owner and narrow permissions. In Productive, each agent has a manager who handles what it can access in terms of data and what it can do. You can build an agent that only prioritises tasks on one project and touches nothing else. Behaviour is the harder part. Productive has around 10 people called Luka, and "give it to Luka" goes wrong fast. Tune the agent to ask too often and it annoys people. Ask too little and it guesses. The team balances the two, notifies people when it decides on its own, and uses thumbs up and down feedback so the agent learns, for example, which Luka usually handles that type of task.

Try this: for every agent you deploy, write down who manages it, what data it can see and when it must ask before acting.

Why do agency side products rarely become successful companies?#

Because you can't be a car mechanic and a car factory at the same time. Productive started as an in-house tool at Infinum because nothing on the market showed whether a project was profitable in real time, only months later when the bookkeeping closed. Tomislav has seen many agency products since, and few succeed. Products usually start because people are on the bench, and then client work comes back and revenue pulls those people away. His view is that the only way it works long term is a carve-out: spin it off, and someone has to leave or sell their part. The more successful the agency, the more is at stake and the harder that step gets.

Try this: if you run an agency product, give it a dedicated team that client work can't touch. If you can't, be honest that it's a side project.

How do you decide what to build when every customer wants something different?#

Build a system that aggregates feedback instead of following the loudest customer. In the early days there were no customers, so Productive ran on intuition. Tomislav jokes that everything bad in the company is his fault, since he either made the decision or hired the person who did. The turning point came when Productive built an internal system to turn everyone's input into one strategy and stopped building things he personally found interesting. As a bootstrapped company with limited resources, that focus is what got the flywheel going. He keeps the Faster Horses quote in mind too, which he finds true and not true at the same time. It's a spectrum.

Try this: tag every feature request by how many customers raise it, and review the totals before each roadmap cycle.

---

Listen & watch#

• Watch the episode on YouTube
• Listen on Spotify
• Listen on Apple Podcasts
• All episodes at infinum.com/delivered DELIVERED is a production of Infinum and Your Majesty.