Across industries, executives who rushed into AI projects are finding that the promised financial payoff is slower, messier and far more complex than the pitch decks suggested.
AI hype meets the finance department
The latest warning signal comes from a major international survey by PwC, spanning 4,454 business leaders in 95 countries. These are not cautious laggards; they are the ones who have already put real money into AI.
More than half of executives who invested in AI say it has neither lifted revenue nor cut costs in the most recent financial year.
According to the study, 56% of respondents report that AI has not yet translated into measurable financial gains. That means no noticeable bump in sales, no visible reduction in operating expenses.
The picture is not entirely bleak. Just under 30% say AI has helped increase revenue. A smaller group, 12%, report hitting the jackpot: higher revenue and lower costs at the same time. That narrow slice of winners fuels the ongoing race, even as many early adopters struggle to justify their budgets to shareholders.
The tension is stark. AI features heavily in strategic plans, investor presentations and media interviews. In profit-and-loss statements, the impact is often faint.
The mirage of “easy” AI savings
A handful of companies tried a brutal shortcut: fire people, plug in AI, book the savings. Several quickly regretted it. Systems that looked impressive in demos stumbled when exposed to messy, real-world workflows.
Some firms that laid off large numbers of staff and loudly trumpeted “AI-driven efficiency” ended up facing service failures, angry customers and expensive damage control. Rebuilding teams, re-training staff and fixing broken processes can wipe out any theoretical gains.
AI is proving far weaker at replacing humans outright than some executives and investors claimed — at least for now.
➡️ Diese wohnungstemperatur entscheidet ob sie diesen winter gesund bleiben oder den schimmel gleich mit einziehen lassen
➡️ Winter sales at Boulanger: our selection of must‑grab deals
➡️ Warum der wunsch nach sicherheit im job deine karriere ruiniert und weshalb bequemlichkeit keine entschuldigung mehr ist
➡️ China gives Airbus an early Christmas present as Tianjin plant hits symbolic production milestone
➡️ Wölfe zurück im dorf für die einen naturwunder für die anderen existenzbedrohung
➡️ Rentner soll landwirtschaftssteuer zahlen weil er seine wiese an einen imker verpachtet hat und jetzt leer ausgeht
➡️ Motten loswerden mit einem geruch, der mehr über deine nachbarn verrät, als dir lieb ist
➡️ Diese wohnungstemperatur trennt vernünftige winterplaner von heizfanatikern und schimmelzüchtern
This aligns with warnings from leading researchers such as Geoffrey Hinton, who argue that large-scale, unconstrained deployment of AI can be both risky and economically disruptive. In his view, the huge AI bets only become financially rational if machines eventually replace a majority of workers — a scenario that is far from today’s reality.
AI is not plug and play
One of the clearest lessons from the past two years is that AI rarely works as a simple “install and profit” tool. It is not a mouse you plug into a laptop and start using in seconds.
In many companies, AI lives in a parallel universe: flashy pilots, innovation labs and one-off experiments that never touch the core processes that actually generate revenue.
- Customer-service bots that handle only the easiest questions
- Proof-of-concept tools for marketing copy that no one fully trusts
- Data-science models that stay in slide decks, not in production systems
These projects consume time and cash but rarely show up in financial results. The fear of missing out — on talent, investors or media attention — is often a stronger driver than a clear, grounded business case.
When integration breaks down
According to PwC and other analysts, most firms are still struggling to embed AI in “mission-critical” workflows: pricing, supply chains, risk assessment, manufacturing, or large-scale customer operations.
As long as AI remains trapped in side projects, its return on investment will look disappointing on any CFO’s spreadsheet.
Beyond integration, the technology itself causes friction. A report from MIT last year found that around 95% of attempts to roll out generative AI inside companies failed to deliver a rapid boost to revenue. Tools that promise to generate emails, code or reports often hallucinate facts, misinterpret instructions or break down on tasks that seem simple to humans.
That forces companies to keep humans in the loop, which reduces labour savings and increases training and oversight costs. At the same time, data-security concerns hang over many deployments. Executives worry that confidential information fed into AI models could leak back out in future outputs, to competitors, customers or the public.
Why leaders are still doubling down
Despite the underwhelming financial metrics, most executives are not walking away from AI. If anything, they plan to invest more aggressively over the next two years. PwC points to 2026 as a potential tipping point, when today’s experiments either harden into real productivity gains or turn into an expensive write-off.
Part of this persistence is strategic. Boards fear that if they slow down, rivals will not. AI is becoming a signalling tool: a way to show investors, partners and potential hires that a company is modern, ambitious and future-facing.
Leaders are betting that the current gap between AI promise and AI profit is temporary, not structural.
The key question is whether they adjust their strategy: moving from scattered pilots to focused, measurable, high-impact projects that tie directly to core business metrics.
What a realistic AI ROI strategy looks like
Firms that are starting to see returns tend to follow a more disciplined approach than the average AI enthusiast. Instead of chasing flashy use cases, they focus on places where AI can quietly reduce friction or speed up repetitive work.
| AI focus area | Typical goal | Common pitfalls |
|---|---|---|
| Customer support | Shorter response times, 24/7 coverage | Bots giving wrong or confusing answers, escalating complaints |
| Back-office processes | Lower manual workload, fewer errors | Poor integration with legacy IT, staff bypassing tools |
| Data analytics | Better forecasts, smarter decisions | Low data quality, models that no one trusts |
| Content and coding | Faster production, new ideas | Hallucinated facts, insecure or buggy code |
Return on investment tends to materialise when companies:
- pick a narrow, well-defined problem
- measure baseline performance before AI
- deploy in production, not just in labs
- track financial impact for at least 12–24 months
This is slow, unglamorous work, and it clashes with the breathless pace of AI marketing. Yet without it, even the most powerful models turn into sunk costs.
Key risks executives are underestimating
The ROI anxiety around AI is not only about money; it is also about risk management. Some of the most serious dangers sit just beneath the surface of “experimental” deployments.
Hallucinations and hidden labour
Generative AI systems are notorious for producing confident but false answers. In a consumer chatbot, that might be amusing. In a bank, hospital or law firm, it can create legal and financial liability.
To cope with this, many organisations quietly add extra layers of human review. Staff check AI-generated content, rewrite outputs or run parallel systems “just in case”. That hidden labour eats into the savings that were originally used to justify the project.
Data leakage and regulatory pressure
Another unresolved issue is data governance. When employees paste confidential information into AI tools, companies have limited visibility into where those data fragments travel next. That raises questions about trade secrets, client confidentiality and compliance with privacy law.
Without strict guardrails on data use, AI projects can generate more legal risk than financial return.
Regulators in Europe, the US and Asia are already signalling heavier scrutiny of AI use in finance, healthcare and employment. Compliance costs, audits and reporting requirements will likely rise, adding another line item to the AI bill.
Terms that help decode the debate
Two expressions crop up repeatedly in board-level AI discussions:
- Generative AI: Systems that create new content — text, images, code, audio — based on patterns learned from training data. Tools like chatbots, image generators and AI coding assistants fall into this category.
- Hallucination: When an AI tool fabricates information that looks convincing but is factually wrong. The system is not “lying” in a human sense; it is assembling patterns without understanding truth.
Understanding these terms is not just jargon. It shapes expectations. Generative AI can speed up creative or analytical tasks, but it will occasionally be wrong in surprising ways. Treating it as an infallible oracle is a quick route to disappointment — and poor ROI.
What the next few years could look like
Imagine two companies starting from the same position in 2024, both under shareholder pressure to “do something with AI”. The first rolls out chatbots, content generators and internal tools with minimal planning, assuming quick savings. The second chooses three core processes, sets clear targets and invests in training and redesigning workflows.
By 2026, the first company may have cut headcount but finds customer satisfaction slipping, error rates rising and staff quietly working around unstable tools. Any early savings are offset by churn, reputational damage and rework. The second moves more slowly but starts to see steady, traceable efficiency gains — not dramatic, but defensible.
This kind of scenario, more than glossy keynote slides, is what will determine whether AI ultimately looks like a historic productivity engine or just another expensive management fad. For now, the return on investment question is turning from awkward side note into a central concern in every serious AI conversation.
