[ad_1]
Most AEC firms can run an AI pilot, but few can scale it. Enthusiasm for a first experiment rarely extends beyond the pilot team, leaving the pilot as an isolated win that never changes how the company works. In a recent AEC Business Podcast episode, Adeline Chan, CEO and co-founder of Hong Kong-based AAL Innovation, explained why that gap exists and what separates the firms that close it from those that stall.
The pilot mindset that holds construction back
Adeline draws a sharp contrast between finance and construction. Banks cannot afford to fall behind their competitors, so they pilot constantly, sometimes running hundreds of proof-of-concept teams in parallel and letting them compete internally to identify the approach that works. The winning concept becomes the standard and scales across the organization, and the return on investment is measured in undeniable numbers.
Construction behaves differently. The prevailing preference is to be second, letting someone else prove a technology before committing to it. Adeline calls it a fear of wasting rather than a fear of missing out, and I recognize the pattern. A CEO of a Nordic construction company aptly captures the attitude. He said, “We know the risks of construction, but not the risks of new technology,” emphasizing the current threat over the future opportunity.
Template libraries turn one win into many
When an AI workflow proves itself on a specific task, the firms that scale successfully do one thing right away. They capture it in a template library. Site report generation may be a niche for one engineer, but the underlying steps and building blocks often apply to other functions, and a shared library lets teams reuse what works rather than rebuilding it from scratch each time.
The library also gives teams a channel to compare successes and failures rather than pioneering in isolation. Adeline sees this as the practical middle stage of adoption. The end state, and the most scalable one, is a company where enough people have the AI literacy to build their own workflows.
Governance that guides instead of blocks
Letting people build their own tools raises the obvious worry about ungoverned shadow IT. Adeline’s answer is a traffic light system. Green means low risk and go, red means high risk and stop, and orange means something needs a decision from someone with authority before it moves forward.
This works only when the framework and the AI policy are already in place, and Adeline is clear that putting them there is leadership’s job as much as IT’s. Leadership sets the checkpoints because the IT department often lacks the authority to grant access to sensitive folders on its own. Tools like Microsoft Copilot Studio make the green path easier, since an agent built inside the company’s existing Azure environment stays within boundaries that were agreed on the day the firm adopted Microsoft 365.
Data still decides how far AI gets
The debate over whether modern AI can handle any data or still needs it well-organized has a practical answer. Structuring the data lake still matters. AI can help do the structuring, but at scale that burns a large volume of tokens, so the real decision is a cost comparison between token spend and the human hours it would take to organize the data by hand.
I raised a case from the 2026 AI in AEC conference in Helsinki, where a UK contractor spent two years organizing its safety data before it became useful. My view is that good data at the outset still saves substantial time and money, even once AI is in the loop. Adeline agreed and added that it helps to have someone in-house who knows how to do the work.
Rethinking what ROI means
Adeline shared a concrete result. A partner in Shanghai reduced person-hours by 80 percent on a state grid substation project using auto-generated BIM. The method works because the output is repetitive and compliant rather than expressive, with Revit serving as the interface while specifications and local codes drive the model. A similar rule-based workflow produces construction-ready BIM for KFC outlets in China, where standardized fixtures enable automation, though the script itself can take up to six months to build before it starts paying off across many sites.
The more interesting point is how Adeline frames ROI. Conversations about reduced person-hours make people uncomfortable, and for good reason. Firms with a clear vision repurpose those people rather than releasing them, as IKEA has done with its AI rollout, and the real return should be measured by the new value redeployed staff create.
She adds a warning about measurement that deserves attention. Pre-AI performance metrics should not be carried over to judge post-AI work. How many tokens someone burns says nothing about how well they use AI, because the only measure that counts is the quality of the output.
The bottleneck is operational, not technical
Looking ahead, Adeline expects the technology to continue running well ahead of adoption. She sees three levels at play: strategy, operations, and technology. The technology is far ahead, strategy depends on leaders getting clear in their own minds, and operations, the way people actually work day-to-day, is where the real bottleneck sits.
Her five-year outlook is a world of agentic workflows, in which agents hand work to other agents. The risk she flags is intent drift. When each agent is only 99 percent accurate, small errors compound along the chain until the output no longer matches what anyone asked for. In her view, the role of AEC professionals shifts toward curating and guiding those agents rather than being replaced by them.
The real challenge
The through-line of the conversation is that AEC’s AI challenge is not really about models or tools. It is about the discipline to move from a single working pilot to a shared, governed, and measurable practice. Firms that build that discipline will pull ahead, and the ones waiting to be second may find the gap harder to close than they expect.
You can listen to the full conversation with Adeline on the AEC Business Podcast.