I didn't add client facing AI to my previous product

To clarify: I am proud that it was not prioritised over the deterministic tools we were able to bring to life. I will later mention some items we were thinking about that have validity. The main decision was to not build our product around agentic, conversational paradigms.

In short, this was a decision to forgo conversational control and grandiose AI at the earliest stages of product development.

When making the decision I wanted to do one thing: be deliberate and make a weighted choice. We were an ultra small bootstrapped team with no resources. Every decision counts and ambitious experiments were unaffordable.

The core build was happening in December 2025. The tools were out there and there were valid use cases.

There were two facets in the decision making process. One is demand, or potential value add. The other was the engineering cost of some of the possible solutions. I will start with some thoughts relating to the product domain and follow them with the cost of engineering.

Our product exposed some of the novel tools that would be best fit for a power-realtor to the client, on a premise that the client can do most of the realtor’s work themselves. Examples are getting full broker-side listing information, scheduling showings, and generating all kinds of real estate offer documents from the agreement of purchase and sale to mutual release forms.

I frequently told our clients our core premise: if you’ve ever bought a used car, you can buy a house yourself. What this means is that with the right tooling and educational resources you are in full capacity to do the legwork, and as per our USP keep the commission that is in five figures.

When thinking about how the clients might use our software, and via some observations and conversations, they did a lot of research via YouTube resources, our blogs and other content to get up to speed. It’s not a stretch to assume that Claude or ChatGPT was their realtor replacement too.

This was the critical point that led to an unspoken positioning: the product is a deal facilitator to decisions happening in Claude/ChatGPT. I deeply believe that no matter what you introduce, the users will fall back to their tool of choice for end to end intelligence, and keeping up with the capabilities of frontier labs’ products was a lost battle.

In simple terms: our product needed to become a surface to act on the decisions you make in your chatbot.

Importantly, I do not disregard some of our roadmap items such as an AI powered Comparable Market Analysis, clarifications on listing details, and even agentic showing scheduling. What I immediately discarded is system control and actioning via an agentic interface.

The first reason regarding control was the observation that the clients had no notion of the steps they needed to take. Secondly, the steps they needed to take were (1) highly rigid and transactional, (2) rich in data inputs and (3) mostly high stakes and involving legally binding contracts.

Hence, the solution needed to showcase and guide the user through the steps and expose the tools in a digestible way. Once they get hold of what is possible, the final set of capabilities is finite and rigid.

Some users may not have a good notion of the operations necessary to build software. Most in the professional context however do. But most importantly: once they get the hang of the capabilities, the set of operations is infinite and non-deterministic.

Our product wasn’t. It would be at most 10 operations you could do. Furthermore, all of those 10 operations with the exception of a few are very meaty, such as inputting 20–30 parameters to your offer composer. Additionally, while there may be many permutations of the order of actions, I would still argue that there existed a sequential nature to actions.

The choice was simple: this was a case for deterministic UI. We leverage the strength of data presentation, confirmation, control and guardrails to ensure the client is in full control and can trust the system with the high stakes operations. Furthermore, subsequent actions can be done blindly in seconds, as all are highly defined.

While contemplating an opposite path for product development, cost of engineering was also important. Even if there was a solid idea for an agentic HCI in this product, developing evals, guardrails and engineering a resilient and delightful system would be an expensive endeavour. Pairing high cost with a fuzzy value add projection was the deciding factor to push the ideas into the backlog. Trying to compete with the memory and intelligence capabilities of frontier tools would introduce staggering variable costs too, on what is often a 90–120 day sales cycle with a high chance of abandonment for exogenous reasons.

Instead, our team and I focused on iteratively building out amazing deterministic capabilities and infrastructure to support the previously B2C platform turning into a B2B2C ready solution. With the resources and time we had, I am confident that the solutions we brought bring more strategic value to the company than betting on a conversationally driven real estate process after a flashy demo subject to the 80/20 rule. We were able to navigate estimation and prioritisation accurately and hit the goals we set out to hit.

Now that the infrastructure is ready and the product performs as per client feedback and the metrics we keep an eye on, it was the time to contemplate revisiting this subject.

The internal pilot I made was great: quickly building a product API and MCP, and creating a slackbot to empower the customer support and real estate team. The problems an internal agent solves are different from the problems the client has, and a non-deterministic interface to the platform from the place of inter-team communication (Slack) was a great productivity enhancement.

What was resurfaced on the roadmap is thinking harder on a symbiosis of deterministic UI and conversational guidance. A copilot that is able to control the UI and surface/resurface when necessary could be fun. A WhatsApp agent that would allow the client to react to certain events or perform highly surgical actions, perhaps? Or maybe validating conversational usage via an MCP that would connect to the client’s frontier product of choice?

I love some of the amazing choices out there on the market. My statement is simple: deterministic UI is still a strong candidate in many cases, and prioritisation in the roadmap is more critical than ever. We chose what NOT to build, at least at that point in time.

Was I wrong?