Early-Stage Planning Software in 2025 - What broke, what worked, what changed

The biggest shift we saw this year wasn’t better AI. It was a change in what people actually want from it.

Early on, the focus was on generation. More options. Faster output. Everything automated. But in practice, that’s not where the bottleneck is. What people are really looking for is control. Not better proposals, but a better way to get there.

Instead of ending 2025 by listing achievements, we’re taking the opportunity to share the lessons shaped by this insight. We’re sharing what broke, what worked, and what changed over the course of 2025.

What broke

A number of assumptions were quietly tested. Some of them didn’t hold up particularly well once constraints, stakeholders, and real decision-making entered the picture.

In practice, several patterns emerged:

• ·Fully automated workflows from input to output didn’t hold up - What's wanted is not a one-step, fully generated model. Its to move through the process, adjust assumptions, and understand the outcomes.
• AI quickly turned into a black box We tested several approaches, including direct GPT-4o input and prompt-based proposal generation. Output quality improved, but transparency did not.

Overall, automation without structure and transparency created fragility rather than clarity.

What worked

Across teams and project contexts, a set of patterns consistently proved useful. These patterns were not tied to specific features or technologies, but to how work was structured and supported in practice.

• Control over outcomes Being able to steer and adjust scenarios, rather than simply generate them, made decision-making more robust and collaborative.
• A strong data model across architectural scales - A shared data model became the foundation for everything else: KPIs, edits, AI assistance, and future integrations. Working consistently across urban, building, block, and unit scales reduced ambiguity and made scenarios easier to compare, iterate on, and explain.

• Typology libraries At the block level, we introduced plans built around core typologies. At the parcel level, we defined building-shape typologies such as L, I, O, and U configurations.

• KPIs grounded in the model We updated the statistics panel to reflect early-stage decision-making in practice. KPIs are now structured hierarchically, linking the model directly to metrics and making trade-offs easier to understand and discuss.

Structured data, clear typologies, and well-designed libraries are what make results understandable, comparable, and usable. Without a solid data model, it doesn’t matter how advanced the generation is.

‍

What changed

Early-stage planning in Hektar quietly shifted away from open-ended design exploration. Instead, it became about control and workflow support across multiple scales, using structured but lightweight representations.

The goal moved from design exploration to design support.

What this means going forward

Much of our work this year focused on building the foundation: the data model, structure, and logic that enable iteration, comparison, and further development.

That foundation enables:

• editable scenarios
• defensible KPIs
• meaningful AI assistance
• seamless integration into existing workflows

This will be our focus in 2026. Without structure, AI adds noise. With structure, it adds leverage.

To learn more, visit our website, Hektar, or contact our team.

‍