AI Is Expensive Because We’ve Taught Everyone to Generate More Noise

There’s a slightly awkward truth sitting underneath the AI boom.

Everyone is using it. Everyone is excited by it. Everyone is producing more with it.

But a lot of it is still loss-making.

Not because AI is useless. It clearly isn’t. Used properly, it can be extraordinary. But at the moment, a lot of AI is being used like a very expensive photocopier attached to a firehose.

People are generating more emails, more documents, more meeting summaries, more reports, more content, more decks, more transcripts, more “thought starters”, more automated responses, more customer interactions, more everything.

Which sounds productive, until you realise that productivity is not the same as volume.

We’ve confused “more output” with “more value”.

And AI has made that confusion dangerously cheap at the point of use, but very expensive underneath.

The economics are brutal

At the consumer end, AI feels almost free. You type a question, get an answer, say “thanks”, ask it to rewrite something, ask it to make it warmer, punchier, shorter, more professional, more human, less corporate, more strategic, less salesy, and so on.

But behind that tiny interaction is infrastructure. Data centres. Chips. Electricity. Cooling. Networking. Storage. Engineers. Model training. Model inference. Security. Governance. Compliance.

That is why Goldman Sachs asked the uncomfortable question in 2024: “Gen AI: too much spend, too little benefit?” Their report estimated that tech giants and others were set to spend around $1 trillion on AI capital expenditure in the coming years, with relatively little to show for it at that point. (Goldman Sachs)

And it is not just theoretical. Reuters recently reported that OpenAI remains unprofitable, with a projected $25 billion burn in 2026, despite extraordinary revenue growth expectations. (Reuters)

That should make everyone pause.

Because if one of the most successful AI companies in the world is still battling those economics, what chance does the average enterprise have if it is simply throwing AI at badly organised knowledge, unclear workflows and unmanaged content?

The world is drowning in information

The amount of information being created is absurd.

IDC previously projected the global datasphere would grow from 45 zettabytes in 2019 to 175 zettabytes by 2025. (Seagate) Other estimates using Statista data put global data creation, capture, copying and consumption at around 149 zettabytes in 2024, rising to 181 zettabytes in 2025. (Rivery)

That is not just data. That is noise, duplication, contradiction, half-finished thinking, old versions, unnecessary copies, forgotten documents, dead policies, transcripts nobody reads, dashboards nobody trusts and content generated because someone thought they should generate something.

AI has not reduced that. In many places, it has accelerated it.

• Before AI, writing a mediocre report took effort. Now it takes a prompt.

• Before AI, producing ten versions of the same article was annoying. Now it is a Tuesday morning.

• Before AI, a meeting generated some notes and a few actions. Now it generates a transcript, a summary, an action list, a sentiment analysis, a follow-up email, a CRM update and three slightly different versions of what may or may not have been agreed.

Again, some of that is useful. But a lot of it is just digital cholesterol. It clogs the arteries of the organisation.

The real problem is not AI cost. It is wasteful AI use.

People often talk about AI being expensive because of the computing power. That is true, but slightly incomplete.

The deeper problem is that organisations are paying AI to process rubbish.

They are asking expensive models to read messy content, summarise duplicated documents, answer from contradictory sources, interpret old policies, search badly maintained knowledge bases and generate new material on top of already bloated information estates.

That is like hiring a Michelin-starred chef and giving them a fridge full of expired leftovers.

You might still get something edible. But it is not a good operating model.

And this is where the gap between AI adoption and AI value starts to show.

McKinsey found that 78% of organisations were using AI in at least one business function in its 2025 survey, up from 55% a year earlier. (McKinsey & Company) But adoption is not the same as financial impact.

MIT’s 2025 State of AI in Business report was even more blunt. It found that despite $30–40 billion in enterprise investment into generative AI, 95% of organisations were getting zero measurable return from their GenAI initiatives. (MLQ)

That is the bit nobody wants to put on the conference slide.

The problem is not that companies are not experimenting.

They are.

The problem is that most experiments are not becoming measurable value.

We’ve created an AI content machine, not an intelligence machine

So here’s my "chat in the pub" bullet point version :-)

• AI is brilliant at producing more stuff.

• But most businesses don't need more stuff.

• They need more clarity.

• They need better answers. Better decisions. Better service. Better compliance. Better knowledge. Better execution.

• Instead, too many organisations are using AI to increase the speed of content production without improving the quality of the underlying knowledge.

• So the system starts to feed itself.

• A poor knowledge base produces a poor AI answer.

• Someone spots the answer is poor, so they ask AI to rewrite it.

• That rewrite gets saved somewhere.

• Someone else later asks AI a similar question.

• Now the AI has even more conflicting material to choose from.

• Then someone generates a summary of the conflicting material.

• Then someone generates a deck from the summary.

• Then someone uploads the deck back into the knowledge estate.

• Congratulations. You have invented a very expensive confusion engine.

The energy bill is not imaginary either

This is not just a software margin problem. It is also a physical infrastructure problem.

The International Energy Agency estimates that data centres consumed around 415 TWh of electricity in 2024, about 1.5% of global electricity consumption, and that data centre electricity use had grown around 12% per year over the previous five years. (IEA)

The IEA also states the obvious but important truth: “there is no AI without energy.” Training and deploying AI happens in large, power-hungry data centres. (IEA)

So every pointless prompt, every unnecessary summary, every duplicated AI-generated document and every chatbot answer built on bad knowledge has a cost.

Not always a visible cost. But a cost nonetheless.

So how do we turn it around?

The answer is not “stop using AI”. That would be ridiculous.

The answer is to stop using AI as a content multiplier and start using it as a value multiplier.

That means changing the question from:

“How can we use AI to create more?”

to:

“How can we use AI to remove waste, improve trust and reduce cognitive load?”

That is a very different conversation.

It means fewer random pilots and more disciplined use cases.

It means not asking AI to generate yet another document unless that document actually improves a process, resolves a customer issue, reduces risk or helps someone make a better decision.

It means measuring AI by business outcome, not novelty.

• Did it reduce average handling time?

• Did it improve first contact resolution?

• Did it reduce escalations?

• Did it improve compliance?

• Did it remove duplicated knowledge?

• Did it help people find the right answer faster?

• Did it reduce the amount of content we have to maintain?

• Did it make the business calmer?

Because that last one matters more than people think.

A calm knowledge environment is a competitive advantage.

The fix starts before the AI layer

This is where knowledge management becomes painfully relevant.

Most AI economics are broken because the input layer is broken.

• The knowledge is duplicated.

• The ownership is unclear.

• The policies are outdated.

• The articles contradict one another.

• The documents are written for humans but not structured for machines.

• The audit trail is weak.

• The same question has five answers depending on where you look.

Then the organisation puts AI on top and wonders why the results are patchy.

AI does not magically create truth.

It retrieves, predicts, summarises and generates from what it has access to.

So if your knowledge estate is a mess, AI simply makes the mess more interactive.

That is why the future of AI value is not just about better models.

It is about better knowledge foundations.

What good looks like

Turning AI from loss-making noise into value requires a few basic shifts.

• First, clean the knowledge before scaling the AI. Identify outdated, duplicated, contradictory and low-quality content before allowing AI to rely on it.

• Second, connect AI to governed sources, not everything. More access does not mean better answers. Sometimes it just means more ways to be wrong.

• Third, reduce content, do not just generate it. AI should help retire old material, consolidate duplicates, flag contradictions and improve structure.

• Fourth, make answers explainable. If AI gives a recommendation, the organisation should know where it came from, whether the source is approved, who owns it and when it was last updated.

• Fifth, measure the right things. Usage is not value. Tokens are not value. Number of prompts is not value. Business impact is value.

Where index fits

This is exactly the problem index is built to solve.

index does not start with the assumption that the answer is “more AI”.

It starts with the assumption that trusted AI needs trusted knowledge.

index Scan looks across enterprise knowledge systems and identifies the hidden problems that make AI expensive, unreliable or difficult to trust: duplication, contradiction, outdated content, broken links, weak ownership, poor structure, low machine readability and governance gaps.

index Solve then helps organisations fix those issues through controlled remediation workflows, approvals, audit trails and human oversight.

In plain English, index helps organisations stop feeding AI with junk.

Because that is where the economics change.

• If AI is being used to process bloated, messy, ungoverned knowledge, it becomes expensive theatre.

• If AI is connected to clean, structured, governed, trusted knowledge, it becomes operational leverage.

The uncomfortable conclusion

AI is not loss-making because it lacks potential.

It is loss-making because too many organisations are using it to create more information instead of better intelligence.

• More summaries.

• More drafts.

• More reports.

• More noise.

But the winners will not be the companies that generate the most.

They will be the companies that know what to trust, what to remove, what to improve and what to automate.

The real AI advantage will not come from producing more content faster.

It will come from needing less content because the knowledge is better.

And that is the shift.

• From more output to better answers.

• From information overload to trusted knowledge.

• From expensive experimentation to measurable value.

Because in the end, AI does not become profitable by generating more mess.

It becomes profitable when it helps us stop managing the mess in the first place.

by Paul Tucker - contact@index-ai.net