The Thermostat Effect: Why AI Feels Distant but Works Deep
The dial’s been turned—but the room is only just starting to warm up. On Schmuki, ROI and Azhar’s $100T puzzle.
When I talk to people about AI, I often get the same question: "When is it actually going to change how we work? I don’t notice it." It's a fair question, especially when you consider how much hype surrounds the topic.
My go-to metaphor? A thermostat. You turn the dial, but the temperature doesn’t change right away. It takes time. The room has to adjust, the system has to respond, and only then does the difference become noticeable. AI is working the same way: the dial has been turned, and the heat is coming.
At Schmuki, the digital and AI agency I co-founded with my partner, we explore these changes in real-world settings. I often draw on our work as a lens to think and write about how AI tools actually perform outside the lab.
One part of what we build is highly visible: public-facing chatbots that clarify, guide and respond. The other part goes deeper: internal agents and knowledge systems that support teams, speed up tasks, and quietly reshape workflows. These tools are already having an impact, but the ROI isn’t always immediate or easily measured.
This disconnect between capability and perception is what Azeem Azhar calls the capability-absorption gap. His recent article, The $100 Trillion Productivity Puzzle, resonated deeply with me. It explains why the numbers aren’t showing the AI revolution. Yet.
Models are improving rapidly, prices are falling, and tools are ready. But businesses are slow to absorb them. The gap is not in the tech; it’s in the uptake.
Three Frictions Blocking Change
Azhar outlines three structural reasons why organisations lag:
- Learning Time: Like the typewriter, which took 25 years to show productivity gains, AI requires cultural and managerial learning before it can deliver results.
- Organisational Complexity: AI doesn’t belong to one department. It cuts across customer service, HR, product design, data analysis. Who leads? Who funds? Who measures?
- Non-Determinism: AI tools can look right and be wrong. They create outputs that pass surface-level checks but require time-consuming verification. Trust is slow to build.
These are exactly the tensions we run into with clients. Our chatbots often perform well in user tests, but the real challenge is embedding them into existing structures. Internal tools show promise in test environments, but live implementation depends on how open people are to process change.
Azeem Azhar
And Yet, the Economics Are Changing
One of the most compelling points in Azhar's piece is about unit costs. The price of AI inference has been halving roughly every six months. This matters. The cheaper the tool, the more places you can apply it.
At Schmuki, we’ve seen clients save dozens of hours a week by letting AI handle intake forms, knowledge retrieval, and even onboarding content. These aren’t headline-grabbing use cases, but they compound over time.
Still, ROI questions persist. One crude but energising exercise is to compare the cost of an AI agent to the cost of a human. It’s not a perfect match, apples and pears, but it sets the mind racing. If an internal agent can do a task 80% as well as a junior hire, and do it 24/7, what does that mean over six months? Over a year? Now multiply that by ten roles or tasks.
Rob Hoeijmakers
Redesign, Not Just Replace
Here’s where things get practical. The ROI isn’t just about using AI to go faster. It’s about redesigning the work itself. Azhar uses the example of electricity: its economic impact only surged when businesses reorganised factories to take full advantage. The same is true here. Don’t just ask where AI can plug in. Ask what could be rebuilt if AI was your default assumption.
That’s the approach we take at Schmuki. Whether it’s automating content workflows, structuring healthcare communication, or integrating voice interfaces in education, our focus is always dual: use AI to improve clarity and to support human agency. The tools are here. What matters now is design.
While AI agents often pay for themselves in a matter of hours, most clients aren’t buying cost savings.
They're buying:
- Control over their own data and processes
- Trust in a secure, privacy-respecting setup
- Clarity through real-world experiments, not empty promises
ROI comes later—once the foundations are solid.
If you're selling AI, start with agency, not efficiency.
The Heat Is Coming
You may not feel it yet. But the system is already warming up. At some point, the change will be unmistakable, just like that moment when you realise the room is suddenly warm, and you’ve already taken off your jumper.
By then, the real work will have already been done: learning, adapting, redesigning. That’s why it’s worth pushing now.
Rob Hoeijmakers
Digital Strategist at Schmuki
Benjamin Todd