Sam Zolfagharian returns to the AEC Business Podcast for her second appearance. She is an AI strategist, keynote speaker, and the author of two books, Disrupt It and Future by Design. She spoke at the AI in AEC conference in Helsinki in March, and we picked up the conversation where that keynote left off.
The starting point was a question her clients keep raising. Executives are being asked by their boards about the return on their AI investments, and many of them do not have an answer they trust. This episode is about what to measure instead, and when the dollar figure finally becomes the right number to look at.
The ROI question arrives too early
Sam’s argument is that asking for AI’s return on investment after one or two years is the wrong test. She compares it to education. Parents spend 15 or 18 years funding a child’s schooling, and if you audit that investment at year 3, there is no return to be found. The return arrives later, and nobody concludes the spending was a mistake in the meantime.
The reason the analogy holds is that AI is not ordinary software. A conventional tool is deployed, adopted, and measured on a predictable curve. A transformational tool changes what the work is, which means the payback period stretches. Sam puts the realistic horizon at three to five years.
That does not mean measuring nothing. Her position is that firms need metrics from day one, but that those metrics should not start with dollar value. Progress has to be visible before it is monetizable, and the wrong metric early on will kill an initiative that was working.
The metric one engineering firm uses instead
Her client Wade Trim, a top ENR engineering firm, tracks RONI rather than ROI. RONI stands for Risk of Not Investing. It asks what the firm stands to lose over the next three to five years if it stays where it is.
The value of the reframe is that it forces a second number into a conversation that usually has only one. An AI investment looks expensive when measured against its immediate return and cheap when measured against the cost of falling behind competitors who moved earlier. Wade Trim uses the metric to keep those two pressures in balance rather than letting caution win by default.
Sam also noted a motivation that rarely gets stated out loud. Some executives are two or three years from retirement and do not want to hand over a company mid-transformation. Understanding that incentive matters if you are the one making the case internally.
Metrics mature in three stages
Sam divides the AEC market into three stages and argues that each requires its own metrics. Inception is about activity. You count licenses, active users, and what people are actually doing with the tools, whether that is drafting emails or getting real work done faster.
Integration is where workflows change. The metric shifts from how many people use AI to what percentage of work is now done differently. She described an architecture firm of about 250 employees where every single person was using AI, and the executive’s frustration was that many were using it to do the same job the same way. Full adoption at the inception stage can appear successful while producing nothing.
Transformation is where ROI finally belongs. At that point you are measuring new business models, new markets, new services, client satisfaction, and quality, not hours saved. Her summary of the arc: ten percent efficiency at the beginning turning into ten times the value later. You cannot skip to the third stage, because the first two are what make it possible.
Token costs against employee effort
Cost came up repeatedly, and Sam’s answer was to stop looking at the token bill in isolation. She described an architecture and engineering firm in which two employees spent 40% of their time on market research, connecting external market intelligence to internal project work. They built an agent that now does it in half a day.
The tokens are a real cost, and she expects pressure on pricing as investors eventually ask AI companies for their money back. Her point is that the comparison has to include what the work used to cost in salaried hours, and what the quality of the output is now worth. A cost figure with no counterweight will always look bad.
She also sees the competitive field as a form of protection. Multiple credible model providers keep pricing honest, and firms are increasingly building wrappers to access several models at once rather than committing to a single platform.
Do not boil the data ocean
We discussed data quality, and here Sam was blunt about a common failure. Firms hear that their data is messy, conclude they need a data lake or warehouse, and treat that as a prerequisite for everything else. She described a medium-sized engineering firm that spent two million dollars building one and still has the same data problem.
Her alternative is to reverse the sequence. Start with a specific opportunity, determine which data it actually needs, and clean only that. Then move to the next opportunity. The problem becomes a series of manageable pieces instead of one project that never finishes.
This connects to a shift I see in Finland, where the effort is going into standardizing data from the design phase so it remains useful through procurement, construction, and maintenance. Sam’s caution applies there too. As one recent guest put it, if you wait for perfect data, you will never start.
What has changed since March
On technology, the conversation has moved from chatbots to agents. Sam is working with small and medium-sized companies running five or six hundred agents across the organization, a scale that would have been unthinkable two years ago. She expects those numbers to roughly double by next year.
Silicon Valley’s current preoccupation is moving humans from in-the-loop to on-the-loop. When AI runs faster than a person can approve each step, the human becomes the bottleneck. The proposed answer is oversight of the process and validation of outputs rather than step-by-step approval. I raised the counterpoint that managing large numbers of agents is proving exhausting for the people responsible for them.
On people, the model wars have cooled. Practitioners have stopped asking which single AI platform to standardize on and started picking tools by task. Sam compares the models to the Avengers, each with a different strength. Small firms have changed their minds too, and now see their agility as an advantage rather than assuming AI is only for companies with large budgets.
So which metrics should you track?
I asked Sam directly for the three to five metrics a leader could put on a dashboard. Her answer was that it depends, and that she would distrust anyone who handed over a universal list.
She used a gym analogy. Three people training with different goals (building muscle, staying healthy, or losing weight) will track completely different things and eat completely different diets. The metrics follow from the goal, not the other way around. Firms need to decide what they are trying to become in the age of AI, and derive the measurements from that north star.
That answer will frustrate anyone hoping for a template, but it is the honest one. The conversation an executive team needs to have first is not about metrics at all.
Where to find Sam
Sam publishes free reports and white papers on the Insights tab at samzolfagharian.com, covering workforce preparation, hiring, employee sustainability concerns, and AI governance to protect IP and confidentiality.
AI in AEC 2027
Artificial Intelligence in Architecture, Engineering and Construction, AI in AEC, continues on 17 and 18 March 2027 in Helsinki and online. Presentation proposals are open until the 30th of September. AEC Business is the conference’s content media partner for the third consecutive year. Details at aiaec.net.
Listen to the interview here and subscribe on your favorite podcast platform, or watch it on YouTube.
Podcast: Play in new window | Download
Subscribe: RSS
