Everything About AI
NextFin News - Companies are spending an average of $186 million a year on artificial intelligence, 88% of organizations now use it in at least one business function, and 94% of executives say they will keep investing even without near-term returns. Yet 56% of chief executives report that AI has delivered neither higher revenue nor lower costs over the past 12 months, and only about four in ten organizations say AI has contributed positively to earnings before interest and taxes. The gap between AI ambition and AI accounting is the defining corporate-finance story of 2026 — and the binding constraint is not model quality, budget, or regulation. It is trust.
When an AI system cannot be verified, audited, or traced to a source, companies respond the only rational way: they slow deployment, add human checks, and demand proof before scaling. That response shows up on the income statement as verification work that the technology was supposed to eliminate, and in the boardroom as deferred enterprise rollouts. The return-on-investment gap, in other words, is a trust gap with a dollar figure attached.
The Numbers: Near-Universal Adoption, Single-Digit Returns
The scale of the disconnect is visible across the major 2026 surveys. PwC's 29th Global CEO Survey, covering 4,454 chief executives across 95 countries and territories, found that 30% of CEOs saw tangible revenue results from AI in the last year and 26% reported lower costs — but 56% reported neither, while only 12%, the group the survey calls the "vanguard," achieved both.
Adoption tells a different story. McKinsey's latest global survey on the state of AI finds that 88% of organizations now use AI in at least one business function, up from 78% a year earlier — but only 37% report that AI has contributed positively to EBIT, essentially unchanged from 2025 despite the growth in scaled deployments. Average enterprise AI budgets have reached a weighted global average of $186 million over the next 12 months, according to KPMG's Global AI Pulse survey of 2,110 C-suite and senior business leaders across 20 countries. And MIT's Project NANDA, drawing on a review of more than 300 publicly disclosed AI initiatives, 52 executive interviews, and 153 senior-leader surveys, found that 95% of organizations are getting zero return from generative AI, with just 5% of integrated pilots extracting millions in value.
Read together, those figures describe a market that has moved past the adoption question entirely. The question boards are now asking is not "should we do AI?" but "can we prove it works?" — and the answer, for most companies, is still no.
The Trust Deficit Is a Balance-Sheet Item
Trust is usually discussed as a soft, cultural attribute. In enterprise AI it has a hard cost. The mechanism runs through three channels: verification overhead, liability exposure, and the discount rate executives apply to AI-generated outputs.
Verification overhead is the most direct. When employees cannot trust an AI output, they check it. Analyses of enterprise AI usage patterns estimate that skilled workers spend several hours each week manually fact-checking AI claims — a verification tax that converts directly into wages paid for work the technology was supposed to eliminate. The cost recurs with every use rather than amortizing into infrastructure, which is why it shows up in operating expenses quarter after quarter instead of disappearing after implementation.
Liability exposure is the second channel, and it is no longer theoretical. In February 2024, the British Columbia Civil Resolution Tribunal ruled in Moffatt v. Air Canada that the airline was liable for a bereavement-fare policy its chatbot had fabricated, ordering it to pay roughly 812 Canadian dollars in damages and fees after the carrier argued the chatbot "was a separate legal entity" responsible for its own statements. The tribunal rejected that argument directly. In the United States, a federal judge in Manhattan sanctioned two attorneys $5,000 in Mata v. Avianca after they filed a brief containing fictitious case citations produced by an AI tool, finding they had acted with subjective bad faith. Each incident does two things: it imposes a direct cost, and it creates a precedent that makes general counsel and risk committees slower to approve the next deployment.
The third channel is the discount rate. A CFO evaluating an AI project does not discount its expected cash flows at the company's standard hurdle rate; she discounts them further for the probability that the system produces an unreliable, unexplainable, or non-compliant output. That extra discount is the market's price for unverifiability. It is why pilots that look profitable in a demo stall before they reach production: the projected return is real, but the trust-adjusted return is not.
"Isolated, tactical AI projects often don't deliver measurable value. Tangible returns come from enterprise-scale deployment consistent with business strategy," PwC's CEO survey report concluded, identifying strong AI foundations — integration-ready technology, a defined roadmap, formalized responsible-AI and risk processes, and an adoption-supporting culture — as the dividing line between the vanguard and the rest.
The trust deficit, then, is not an obstacle beside the ROI problem. It is the ROI problem, expressed through verification costs, legal exposure, and a higher discount rate on every AI dollar.
Why Pilots Fail and the 12% Do Not
The most important correction to the popular narrative is that the 95% failure rate is not a model-quality problem. MIT's researchers were explicit: the divide "does not seem to be driven by model quality or regulation, but seems to be determined by approach." Generic tools like ChatGPT and Copilot enhance individual productivity; they stall inside companies because they do not retain feedback, adapt to context, or improve over time. The core barrier to scaling is learning — systems that cannot learn a specific workflow cannot move a profit-and-loss statement.
This is the line between the 95% that fail and the 5% that do not — and it is also the line between the 56% of CEOs seeing nothing and the 12% in the vanguard. PwC's data shows the vanguard applies AI to products, services, and customer experiences at more than twice the rate of other companies: 44% versus 17%. They are not running more pilots. They are embedding AI into revenue-generating workflows where the before-and-after baseline is measurable.
Budget allocation compounds the problem. MIT's research identified an investment bias: budgets favor visible, top-line functions over the high-ROI back office. The largest realized returns sit in back-office automation — eliminating business-process-outsourcing contracts, cutting external agency spend, and streamlining finance and operations workflows that are boring to demo but carry a hard dollar baseline. The money flows to what is easy to approve; the return sits where it is easy to measure.
The companies in the small group reporting real ROI share a narrow, high-volume, well-instrumented pattern, almost always in customer service. Klarna's AI assistant handled 80% of customer chats and contributed roughly $40 million in savings in 2024, per the company's disclosures, while cutting average resolution time from 12 minutes to under two minutes. Salesforce reports that its internal Agentforce deployment handles about 32,000 customer conversations per week at an 83% resolution rate, with human escalations cut in half. OpenTable resolves about 70% of inquiries autonomously. The pattern is unambiguous: ROI appears where the use case is narrow, the volume is high, the workflow is instrumented, and the counterfactual cost is known. Everything else is a pilot.
Abandonment is rising alongside spending. Research tracked by S&P Global Market Intelligence found that the share of U.S. companies abandoning most of their AI initiatives jumped to 42% in 2025, up from 17% the prior year, with the average company scrapping 46% of proofs of concept before they reach production. That is not a technology cycle turning. It is a measurement regime arriving.
Agentic AI Raises the Stakes — and the Trust Requirement
The trust gap would be manageable if AI remained an advisory tool. It is not. The industry is moving into agentic AI, where systems do not just suggest actions but take them — processing customer data, invoking internal APIs, and chaining decisions across workflows with minimal human oversight of individual steps. McKinsey's latest survey finds organizations are now deploying agentic systems capable of acting autonomously across workflows, and individual employees report real productivity gains even as organization-wide financial impact lags.
That shift changes the risk profile fundamentally. A chatbot that gives a wrong answer is a reputational problem. An agent that executes a wrong action is an operational and legal one. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. The same research forecasts that 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, and that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.
This is the second-order effect of the ROI gap that most analyses miss. The first-order effect is that companies do not see returns. The second-order effect is that, as AI systems gain the capacity to act autonomously faster than they prove reliable, the trust requirement rises faster than demonstrated reliability — and that widening gap threatens to cancel the next wave of investment before it reaches production. Gartner's separate guidance warns that applying uniform governance across all AI agents leads to failure, because agents operate at different autonomy levels; governance must match the level of independent action, with humans reviewing exceptions, audit logs, and aggregated outcomes rather than individual decisions.
The companies that close the ROI gap will be the ones that treat trust infrastructure — provenance, audit trails, runtime controls, verifiable outputs — as capital expenditure rather than compliance overhead. Trust is not a soft virtue here. It is the infrastructure that makes autonomous action financeable.
The Counter-Thesis: Patience Is Rational, and History Rhymes
The strongest case against reading the ROI gap as a structural trust crisis is the adoption-curve argument, and it has real force — and real backing. BCG's 2026 AI Radar, surveying 2,360 executives including 640 CEOs, found that four out of five CEOs are more optimistic about AI ROI than they were a year ago, and that AI investment is projected to double to about 1.7% of organizational revenue in 2026 from roughly 0.8% in 2025. From this view, the 56% figure is not a warning sign but a normal phase of a technology S-curve. Cloud computing, enterprise software, and mobile transformation all followed the same pattern: heavy upfront investment, a long period of diffuse and hard-to-measure spending, and returns that concentrated late and unevenly. CEOs keeping spend rising are behaving like rational long-term allocators, not victims of hype. The trust language, on this reading, is a post-hoc rationalization for a simple timing mismatch: the money went out before the returns came in.
There is truth in that. Diffuse pilots always precede concentrated value, and premature measurement can kill projects that would have paid off. But the analogy breaks on two points. First, cloud and mobile did not require a verification tax on every unit of output; a server either ran the workload or it did not. AI produces probabilistic outputs that must be checked, and that checking cost recurs with every use — it is embedded in the operating model, not amortized into the infrastructure. Second, the agentic turn is new: previous technologies did not gain the capacity to act autonomously before they proved reliable. The sequence has inverted, and with it the risk.
Patience is rational only if the trust infrastructure is being built during the waiting period. If it is not, the patience is not an investment horizon — it is a delay.
What Comes Next: Beneficiaries, the Exposed, and the Signal That Would Prove This Wrong
The mechanism points to a clear asymmetry. Over the short term, the beneficiaries are the vendors and service providers selling trust infrastructure: audit trails, provenance and watermarking, runtime governance, model monitoring, and verification tooling. The exposed are companies whose AI programs remain pilot-heavy, concentrated in low-measurability use cases, and lacking formal responsible-AI and risk processes — precisely the profile of the 56% seeing no return.
Split by time horizon, the picture differs. In the short term, sentiment will remain tied to spending announcements and pilot launches, and the gap will widen as more companies report zero measurable impact. In the medium term, fundamentals will separate the embedders from the experimenters: companies that move AI into instrumented, revenue- or cost-baseline workflows will start showing returns, and the rest will face board pressure and budget cuts. In the long term, the structural question is whether verifiability becomes a built-in property of AI systems or a bolted-on layer. If it is built in, the trust gap closes as the technology matures; if it is bolted on, the gap persists as a permanent cost of doing business with AI.
Three scenarios frame the path. The base case: ROI concentration continues, with the vanguard share rising slowly from 12% toward 20% to 25% over the next two years as trust tooling matures and weak pilots are culled. The upside case: a verifiability breakthrough — reliable provenance and near-zero hallucination rates in enterprise-grade systems — compresses the trust discount and unlocks a wave of agentic deployment. The downside case: a high-profile autonomous-agent failure triggers regulatory intervention that raises compliance costs and slows deployment across the sector.
The falsifying signal is concrete: if, over the next four quarters, the share of CEOs reporting both revenue growth and cost reduction from AI rises above 25% without a corresponding increase in measured spending on governance, audit, and verification tooling, then the trust-gap thesis is wrong — the ROI gap was a timing and integration problem, not a trust problem. That share is tracked in the annual CEO survey cycle.
The AI ROI gap will not close because models get smarter. It will close when AI output becomes as auditable as a financial statement — because until then, every AI dollar carries a trust discount, and no CFO will scale what she cannot verify.
Explore more exclusive insights at nextfin.ai.
