In partnership with
Most firms rushing into AI agents have a quieter problem underneath. Sophie Planken-Bichler, industry analyst at Verdantix, estimates that 90% of firms currently using agents lack the data and processes to get the full potential out of them.
🎧 Short on time? Hit play: Watch the full episode on YouTube
Sophie sits in a useful spot. She talks to the vendors building AI for AEC and real estate, and to the firms trying to use it on real projects. What she keeps seeing is a widening gap between the demos and what most organisations can actually absorb.
"I'd say 90% of the firms who are using the agents in practice today... don't have the data and the processes in place to actually get the full potential out of AI agent."
The knock-on effect is predictable. Outputs look unreliable, people can't validate them, and trust drops, especially on safety and compliance work on site. One bad first experience and a risk-averse firm quietly parks the whole AI programme.
How this lands on a real project
If you're about to buy or build an agent, start smaller than the pitch deck. Sophie's three-step, 12-month roadmap is blunt:
Define one narrow problem with a measurable cost or delay (a safety observation review, a schedule risk). Set a baseline before you add AI.
Upskill and set guardrails. Train people on what to do and what not to do, including data privacy. Find out what tools they already use.
Own the data for that use case. Know where it sits, who controls it, and who updates the model. Governance is where most of these efforts quietly die.
"Firstly, you just really need to clearly define the problem that you're trying to solve. So what is the measurable cost or delay that you're currently experiencing?"
She gets into the full roadmap from 30:35 in the episode.
AI won't tidy your mess. It will broadcast it
I pushed on whether AI can help spot holes in bad data. Sophie's answer matched the episode title.
"Adding on the AI overlay can sometimes even expose the weaknesses in people's data and amplify those and extrapolate the failures in the system... there's almost a perception that AI can fix the problem, but actually it amplifies it even further."
That's the bit project teams feel first. Layer a clever agent on incomplete site diaries, inconsistent CDE fields or WhatsApp threads with no owner, and you don't get magic. You get louder mistakes.
Try a shadow AI amnesty
One practical idea we landed on: ask teams what they're already using, with an amnesty. No punishment, just honesty. Sophie was clear.
"I definitely recommend that, 100%... there's a lot of almost embarrassment in some cases of how much people are using AI for certain cases, and it still is demonized a lot of the time."
You can't set guardrails for tools you pretend aren't there. That section starts around 32:20.
We also get into AI literacy reaching site level (not just the back office), why organisational change and job fears slow adoption, and Sophie's 2030 view that AI becomes table stakes you stop talking about. Verdantix publishes a lot of this thinking on its Insights pages, and Vantage has free research if you want the analyst cut.
Why it matters: Buying an agent without data foundations, a narrow use case and basic literacy is how you burn trust. The firms getting value aren't the ones with the flashiest demo. They're the ones who fixed ownership of the data first.
Links and Stuff
Sophie Planken-Bichler on LinkedIn: linkedin.com/in/sophie-planken-bichler-618880196
Verdantix: verdantix.com
Verdantix Insights: verdantix.com/insights
Verdantix Vantage (free research): verdantix.com/vantage
Listen on
▶️ YouTube: full episode (all videos: youtube.com/@Project_Flux)
🎧 Spotify
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
Not subscribed yet? Subscribe to the Project Flux newsletter for the week's AI and tech news for project delivery, every week.
Project Flux is independently owned, funded, and operated by its founder. Any external references or features are for community benefit only and do not imply endorsement, control, or ownership.