Let me say first that I am a big fan of coffee. Its ability to get my day started, or give me a lift when I begin to flag, is an essential part of my routine. Increasingly I look at AI in my daily life in much the same way. It sits in the background, and when I need a helpful piece of information or data, it’s there to give me the intelligence to stay on track.

But as with coffee, it’s an addictive habit, and too much reliance isn’t healthy. I think we’re now entering a second, more sober phase of how individuals and businesses will use AI, one defined less by how much of it we consume and more by how deliberately we’ve built the foundation underneath it.

There are a few ways to reflect on that. In recent weeks you’d have had to be in a cave with no reception to avoid the wave of articles flagging the danger of AI and the risk to humanity from uncontrolled development. The calls from the leaders of the major AI companies for governments to partner with them on catastrophic safety risk need to be heeded, but my ability as an individual to influence that side of things is well beyond my pay grade.

There is, however, a readily addressable aspect of how we embrace AI, both as individuals and as business leaders, that is becoming increasingly clear. We first need to curb the blind enthusiasm for the seemingly unlimited ability of AI to reinvent almost anything we, or it, can think of. In an extraordinarily short period, AI has infiltrated all our lives. Most of us have a subscription to a favoured agentic tool, perhaps even a personal AI assistant we’ve built ourselves. We’ve quietly adopted an AI habit that sits alongside Netflix or Spotify: a seemingly indispensable part of how we live.

A note of caution: we’ve adopted this technology at an enormously subsidised price, and I’d suggest an unsustainable one that individuals and businesses alike are about to have to reckon with. We’re increasingly aware of the enormous cost, in investment and in impact on the planet, of the infrastructure behind it: the water and energy consumption, the data centres, often reported in the billions upon billions of dollars. That insatiable need for capital has fuelled trillion-dollar valuations and repeated all-time highs on the US financial markets. But how does all that spend get paid for?

This is where the phase shift becomes impactful , not just for individuals but for how businesses will actually spend money on AI. The last 18 months have seen business leaders rushing to work out whether AI is a threat or an opportunity, hoping to build a strategy that keeps them in touch. The simple answer is it is probably both. But the 2025 GenAI Divide study from MIT, which reviewed over 300 enterprise GenAI initiatives, put a number on the gap between enthusiasm and result: of an estimated $30-40 billion spent on enterprise GenAI, roughly 95% of organisations saw zero measurable P&L impact. Even among companies that built their own solutions rather than buying one, the failure rate was 95%.

The report’s lead author was clear that the core issue isn’t model quality. It’s a knowledge gap. Generic in-house builds don’t adapt to real workflows the way expert vendor tools, which understand the complexity of process, system and integration, do. Strip away the specifics and the underlying cause is the same one that’s shaped every technology cycle before this: businesses spent on the model layer while skipping the harder, less exciting work of getting their own data structured, connected and mineable in the first place. AI applied on top of disorganised data and undefined process doesn’t produce a transformed business. It produces an expensive experiment with nothing to show for it.

For SaaS businesses built to deliver structured data architecture for clients, this AI exuberance has made winning new business hard, while prospects either sat on the fence waiting for an AI silver bullet, or ran an internal build with little to show for it. Thankfully, studies like MIT’s are starting to shift that. Executives are increasingly realising that if their internal business and process data isn’t structured and mineable, there is little AI can do to move the needle on efficiency or effectiveness, and that seeking out tools already designed to deliver that outcome is both faster and cheaper than building it from scratch. That’s particularly true for project-based businesses carrying a heavy margin sensitivity from accurate pricing, forecasting, delivery and demonstrable continuous improvement.

There’s a related benefit worth highlighting : properly designed tools usually come with permission-based architecture, guardrails around who can access, input or amend project data. That in itself drives more efficient, reliable use of AI in mining data and applying it in a genuine continuous-learning loop. The alternative is expensive in ways that are hard to budget for. Uber reportedly burned through its entire 2026 annual AI coding budget in four months. No guardrails, no cap. And the unit cost of enterprise AI token consumption is only going to rise, likely sharply, against ROI that remains largely unproven.

That wastefulness hasn’t gone unnoticed by the top-tier developers either. Anthropic’s Model Context Protocol has seen rapid, industry-wide adoption precisely because it addresses this: it lets mid-tier SaaS suppliers route data traffic to each other behind corporate firewalls, letting businesses assemble a bespoke digital architecture rather than pouring everything through one generic model. It’s a structural response to the same problem, and, not coincidentally, it also keeps the primary interface, and the revenue, anchored to whichever foundational tool a business has chosen: Claude, ChatGPT or Gemini.

So here is the new phase, stated plainly. The first phase of enterprise AI was defined by how much you could bolt onto your business, model access as the answer to everything. The one we’re moving into now is defined by how well-built the foundation underneath that model actually is. The MIT numbers aren’t a verdict on AI itself; they’re a verdict on skipping the architecture. The practical takeaway for any business leader is to stop asking “how do we get more AI into the business” and start asking “is our data structured well enough for AI to do anything useful with it.” That means auditing what’s actually in your tech stack, right-sizing it to what your business genuinely needs, and right-speccing the tools you choose so the data underneath them is clean, connected and governed before AI is layered on top. Buy or build, the order of operations is the same: architecture first, intelligence second.

Structured inquiry and a tailored, well-specced digital foundation, not unlimited AI consumption, will be what separates businesses that get a return from this technology from the 95% who didn’t.

That, and the sensible, regulated consumption of coffee.