History has taught us that all novel, exponential-growth-enabling concepts must go through a few phases. Mostly known as the Gartner Hype cycle, it boils down to early recognition of a technology’s extraordinary transformative potential (railways, electricity, the internet…), committing significant investment in pursuit of first-mover advantage, but basically taking a leap of faith. What follows is disillusionment when overinflated expectations aren’t met, and only then do the true mass-market productivity gains emerge.
Through my work with agentic systems, including hands-on experience with tools such as Claude Code, I’ve seen how easy it is to reach for a more powerful model than a task at hand requires.
For a while, the economics made it easy to postpone these questions. The priority was adoption, and cheap access encouraged organizations to experiment broadly, consume more tokens, and assume that value would follow. Now the bill is becoming easier to see. So, how did we get here, and what do enterprises need to do to secure value now that AI is no longer “interest-free”?
How have we gotten here
Only a few years have passed from the first commercially available models, which in hindsight functioned more as a novelty and an early glimpse into the future, to the plethora of capabilities we have today.
To accelerate adoption, providers kept the barrier to entry low, effectively “lowering interest rates to 0” to bring more users on board, but this could only grow if the frontier model was truly a premium.
Yet in technology, premiums rarely go unchallenged for long. DeepSeek was the first major disruptor to that logic, demonstrating that open-source models could be viable and that capable AI could run on lighter infrastructure. In doing so, it called the economics of the prevailing approach into question.
The consumption race
The race moved quickly, with quality of output often trumping privacy concerns, but with cost of value being on top of everyone’s mind. Meanwhile, open-source ecosystems have matured rapidly, with a flood of open-source models coming out of China but more recently out of the US market as well. Naming all of them would likely age badly, as the front-runners swap positions on benchmark scores weekly.
And they all have enterprises in their sights. The same enterprises that have embraced AI adoption as a competitive necessity, at the same time facing slower growth and increasing pressure to find new ways to accelerate it. That combination pushed executives to start incentivizing token consumption at scale. And when consumption alone becomes the KPI, chances are it’s not directly correlated with value.
So, where’s the value?
In just the last 12 months, we have seen enterprises embrace frontier models, adopt AI coding agents such as Cursor, Codex, and Claude Code, and move toward agentic workflows with tools such as OpenClaw. Token consumption has scaled rapidly, while questions around cost, governance, and long-term sustainability have largely remained in the background. The underlying conviction is that it will all pay off, with the potential rewards far outweighing the risks.
Meanwhile, the economics are becoming harder to ignore: running and advancing these models requires ever more compute, infrastructure, and power. Somebody ultimately has to absorb those costs. As organizations become more reliant on a handful of AI providers, those providers gain more flexibility in how they package and price their services, creating a powerful incentive to turn growing adoption into growing revenue.
Lastly, the one thing that often gets priced out of value equations is the true cost of governance. Building a piece of software is one dimension of the cost. Managing and operating it is another. And when you have the ability to swap out off-the-shelf solutions for tailor-made ones that fit your workflows perfectly, organizations get excited and creative, and you quickly end up with dozens of new applications you need to keep running.
What to do?
We saw something similar during the early cloud boom. Organizations rushed to move workloads into the cloud, convinced that speed of adoption would translate directly into business value. Sometimes it did. Sometimes it didn’t. What followed was a much more practical conversation about costs, governance, security, and whether the outcomes justified the investment. We’re starting to see the same shift happen with AI.
The solution then, and the solution now, boils down to:
- Assess — Build your own task taxonomy: what work is actually happening, how much is reliability-gated versus routine, where the price of AI automation doesn’t outweigh the returns. How and where to route these (sovereign, frontier or open source) requires domain knowledge only your organization possesses.
- Architect — Some things really do need frontier intelligence. On the other hand, some data and business intelligence need a moat. Decisions around routing, gateway tooling, and sovereign or on-premises deployments should be driven by data sensitivity, regulatory exposure, and vendor-concentration risk.
- Engineer — The context and data layer that determines how much work is safely routable to a smaller model. Better retrieval and structured context raise the ceiling on what the lower-cost tier can handle. Not everything needs Mythos-level intelligence, quite the contrary.
- Govern — As things stand, this is not a build-once-and-forget technology, at least not yet. The landscape reprices in weeks, and getting the call right needs someone who can assess, build, and re-test in short, disciplined cycles. This is the natural entry point for an HTEC Forward-Deployed AI team: small, embedded, fast-moving, present for as long as active iteration is genuinely needed.
What comes next – savvy investment?
As someone who’s spent my entire career at the intersection of market-facing & business value on one hand, and technical challenges on the other, it’s not the first (nor the last) divergence between value expectations and reality I saw. Value is sometimes hard to quantify; we easily devolve into vanity metrics, and simple numbers on a chart just to validate our thoughts.
So just as the FED keeps a close eye on economic data and pivots rate, accordingly, so too must enterprises adopt a process where the right value indicators are tracked and have a consensus on when to invest liberally or when spending gets expensive. It’s a continuous evaluation and pivoting process, and what made sense today, perhaps won’t be the best option in a quarter.
The “zero-interest-rate” phase of AI was never going to last forever. That doesn’t make AI less valuable. It just means organizations need to become more deliberate about where and how they invest.
