Tool, Not Miracle: How SMEs Use AI as an Amplifier — and Where to Leave It Alone

There are two sentences about artificial intelligence that are both true — and that seem to contradict each other at first glance.

The first: AI is the most useful new technology since the smartphone. It writes, summarizes, translates, sorts, helps you code, answers questions about documents nobody wants to read anymore. Ignoring that means leaving real productivity on the table.

The second: The industry supplying this technology is burning money and electricity at a scale without historical precedent — and inside the companies adopting AI, nearly half of all projects fail before they ever reach production.

Both are true. And precisely from this tension follows the only stance that still holds in 2026: AI is a tool, not a miracle. An excellent tool — for the right jobs. An expensive disappointment — for everything else. This piece makes both sides concrete, with numbers instead of mood, and ends on a question every SME can answer this week: Where does AI amplify my business — and where should I leave it alone?

The Bill You're (Not Yet) Paying

Let's start on the supply side, because it explains why the euphoria is so loud: there is an unfathomable amount of money on the table.

Global investment in data centers has nearly doubled since 2022, reaching roughly half a trillion US dollars in 2024 (IEA, Energy and AI). The four large hyperscalers alone — Alphabet, Microsoft, Meta, Amazon — are planning around 700 to 725 billion US dollars in capital expenditure for 2026, an increase of more than 60 percent over an already record-breaking previous year (CNBC; Tom's Hardware, based on quarterly earnings). Microsoft alone cites around 190 billion for 2026 and states it will remain "capacity-constrained at least through 2026" — demand is bigger than what they can build.

And then there's OpenAI. By November 2025 the company had signed compute commitments of around 1.4 trillion US dollars — against annual recurring revenue (ARR) of roughly 20 billion (Yahoo Finance; TechRepublic). That's a 70-to-1 ratio between what it has promised and what it takes in.

That the investment hits the business for real shows up in free cash flow. Analysts model negative free cash flow for Amazon in 2026, a drop of around 90 percent for Alphabet, a similar decline for Meta and negative figures from 2027/28 on as well (CNBC, based on notes from Morgan Stanley, BofA, Pivotal and Barclays). A Barclays analyst puts it dryly: this is "likely what we eventually see for all companies in the AI infrastructure arms race."

Whether this is a bubble is open and honestly contested. A skeptic at SLC Management calls Meta's once "capital-light money machine" a "capital-intensive incinerator"; a bull at Jefferies dismisses the bear thesis flat out as "garbage" and points to real revenue growth. Both could turn out right. But for you as a user, the bubble question is secondary. More important is the lesson that holds either way: if the providers have to keep this pace to survive, then the euphoria is their business model — not yours.

The Electricity Meter Is Running Too

The second bill is physical. Data centers consumed roughly 415 terawatt-hours of electricity in 2024, about 1.5 percent of the world's electricity, and have been growing at around 12 percent per year for years — more than four times as fast as electricity consumption overall (IEA). In 2025 alone, data-center demand rose 17 percent, while global electricity use grew by just 3 percent.

The International Energy Agency's forecast: by 2030 consumption doubles to around 945 TWh — "slightly more than Japan's total electricity consumption today." AI-specialized data centers are set to triple. In the US, data centers account for nearly half of all electricity demand growth through 2030. This is not an abstract climate topic but a very concrete one: grids, connection timelines, and ultimately electricity prices — for everyone else in the region too.

Water comes on top. US data centers directly consumed roughly 21 billion liters of cooling water in 2014, and 66 billion by 2023 — a tripling in nine years, at 1 to 9 liters per kilowatt-hour of server energy (peer-reviewed study, AGU Advances 2026).

For honesty's sake, the counter belongs here: in terms of global CO₂ emissions, data centers remain below one percent according to the IEA. Whoever attacks AI primarily on climate grounds hits its weakest point. The real problem isn't the global CO₂ share but the local concentration: a single data center pushing a region to the edge of its power and water supply. And the point for your own business is a different one: these costs are real, they are rising, and sooner or later they show up in your cloud bill. "Just ask the AI" is never free. It's just someone else paying today.

Expectation Meets Reality

Let's switch sides — from the providers to those who actually adopt AI. This is where it gets interesting for SMEs, because this is where the most expensive misconception lives.

The expectation often goes: we sprinkle a bit of AI everywhere, and then everything gets faster, better, cheaper. Reality looks different. Gartner predicted that by the end of 2025, at least 30 percent of all generative-AI projects would be abandoned after the proof of concept — due to poor data quality, unclear business value, and escalating costs (the figure was reportedly revised upward in 2026). A broad survey by S&P Global / 451 Research of more than 1,000 companies (October 2025) paints the same picture even more sharply: the share of firms that scrap most of their AI initiatives rose from 17 to 42 percent within a single year. On average, 46 percent of AI pilots were scrapped before they ever reached production.

What matters here is what is not being said. These numbers are not proof that AI is useless. They are proof that the way many companies adopt AI is useless: as a fad, as an answer to a question nobody asked, as a feature that "has to go in too" because a competitor announced something with AI. Tools you buy because they shine, rather than because you have a screw they turn, end up in a drawer. With AI, the drawer is just more expensive.

The Hidden Bill: What an LLM Really Costs

The main reason for failure is rarely the demo — that almost always looks good. It's the costs after the demo. Gartner puts the implementation costs for generative AI, depending on approach, at 5 to 20 million US dollars (simple retrieval ~750,000) and stresses: these costs are "not as predictable as other technologies."

That's the crux. Classic software has an unpleasant but honest property: given the same input, it always does the same thing. An LLM does not. From that come cost categories no demo ever shows:

• Per-call costs that grow with usage. Classic software gets cheaper per user the more it runs. An LLM is paid per call — success makes the bill bigger, not smaller.
• Non-determinism as a permanent burden. What isn't reproducible can't be tested easily. Every provider model update can shift behavior — you maintain something that moves under your feet.
• Hallucinations as a business risk. A wrong but confidently phrased answer in a quote, a maintenance log, a compliance statement is more expensive than no answer at all. For critical processes, "mostly right" is often worse than "always traceable."
• Dependence on a provider you don't control. Price, availability, and even the very existence of a model lie outside your hands — a topic we dissected in detail in our piece on digital sovereignty.

The English-speaking engineering community has long had a name for this: "Choose Boring Technology" (Dan McKinley). The idea isn't that boring tech is better because it's boring. It's better because it's predictable — and predictability is hard cash in operations. Every new, exciting component brings unknown failure modes. A company only has a limited budget for such unknowns. Burning it on an AI chatbot that hallucinates the accounting department's phone number is a poor investment of that budget.

The Order That Saves Money

Does that mean you should avoid AI? No. It means: put it in the right place in the chain — and not in the first place. Most tasks in a company don't need a language model. They need a clear rule, cleaner data, or simple automation. An LLM is the most powerful — and most expensive — rung. You reach for it when the cheaper rungs don't solve the problem, not before.

The beauty of this order: it's not anti-AI, it's AI-frugal. It wastes no language model on a task a one-line IF-THEN rule handles — and it saves the expensive rung for the cases where AI genuinely shines. And that's exactly where the opportunity lies that gets talked about too rarely.

The Real Opportunity: AI as an Amplifier, Not a Replacement

Up to here it has sounded like a brake. It isn't. The sober view of the costs is the precondition for deploying AI really big where it works. And it is: big.

The decisive reframe is the word amplifier. AI is most valuable when it makes something your company already does faster and broader — not when it's meant to replace something it never understood. An employee who writes quotes writes three times as many drafts with AI — and checks them herself. A dispatcher who triages damage reports has them pre-summarized and categorized — and decides himself. The AI extends the reach of a human who knows what's right. It doesn't replace them.

That's more than a nice formula — it's the line between the projects that fail and the ones that run. "Replacement" projects fail because they require AI to be 100 percent right, and it isn't. "Amplifier" projects work because a human stays in the loop to catch the 5 percent of errors. The difference isn't the technology. It's whether you treat AI as a miracle (that can do it all alone) or as a tool (that makes a good craftsperson faster).

What an SME Can Do This Week

Theory is cheap. Here are the four steps we recommend in practice — in this order.

1. Write down three tasks where time is lost today. Not "where could we use AI," but "where does it hurt." The right order is problem first, tool second — never the other way around.
2. Place each task on the ladder (Figure 3). Does a rule suffice? Is it just clean data that's missing? Only when it's genuinely about understanding language or images is an LLM up next. You'll be surprised how far down most problems sit.
3. Where AI fits, build it as an amplifier — with a human in the loop. Let the AI suggest, never decide finally, as long as errors cost money or trust. And build an abstraction layer so you can switch providers without rewriting half the system.
4. Calculate the running cost, not the demo. What does it cost per month at real usage, including maintenance and the hours someone spends reviewing the outputs? If the answer is unclear, the project isn't ready yet.

If you face the more fundamental question of whether to build or buy, you'll find the logic in our make-or-buy guide for software. And if you want to go deeper on the energy and dependency angle, see the analysis of digital sovereignty.

Frequently Asked Questions

Isn't this just AI skepticism in a nice suit?

No. It's the opposite of skepticism: it's the precondition for taking AI seriously. Whoever treats AI as a miracle is disappointed at the first failure and drops it entirely. Whoever treats it as a tool deploys it where it delivers on its promise — and wins real productivity there. The sober view leads to more usable AI, not less.

Doesn't "prefer boring" mean falling behind technologically?

No. "Boring" means predictable, not outdated. The most reliable systems in the world — payments, air traffic control, accounting — are deliberately boring. In a company's operational core, "works every Monday morning" is a competitive advantage, not a lag. The exciting belongs at the edge, where errors are cheap, not at the center, where they're expensive.

Where should an SME actually start with AI?

With a task that has lots of free text and low risk: pre-sorting incoming emails, summarizing long documents, drafting quotes. Always with a human who reviews. This delivers visible value quickly without a single error costing money — and builds the experience to decide bigger steps on solid ground later.

Should I worry about energy consumption?

As a citizen: yes, especially because of the local power and water load in data-center regions. As a company: indirectly. The energy hunger is one of the reasons AI prices will stay volatile. Whoever uses AI frugally and in the right place is more robust against price jumps than someone who reflexively built it in everywhere.

Conclusion: The Best Tool Is the One You Hold Right

The AI boom is real, and it's expensive — more expensive than the headlines suggest. Hundreds of billions in investment, electricity use on its way to the level of entire industrialized nations, and on the adopter side an abandonment rate that would have long buried any other technology trend. That's no reason to avoid AI. It's a reason to treat it like a grown-up.

Grown-up means: AI is a tool, not a miracle. It extends what good people and good processes already do — it doesn't replace them. You reach for it when the cheaper, more reliable means aren't enough, not out of reflex. And you keep building a company's operational core from software that works every Monday morning: boring, predictable, maintainable.

The hype sells the opposite — AI as an all-purpose miracle, everywhere, instantly. Whoever follows that is highly likely to land in the half of projects that never reach production. Whoever instead takes AI for what it is — an exceptionally good tool for a clearly defined class of tasks — gets the benefit without the bill. That isn't modesty. That's the more expensive lesson, learned in advance.

Want to find out where AI really amplifies your business — and where reliable, boring software is the better investment? Talk to us.