Trust Is the Most Underpriced Variable in Enterprise AI

In October 2025, IBM's Institute for Business Value published a survey of 915 executives with two numbers that belong side by side. In 2024, companies spent 7% of their AI budgets on AI ethics. The quarter of companies spending 10% or more reported, on average over two years, 30% higher operating profit from AI than the quarter spending 5% or less.

If even part of that gap comes from the spending, it has a name. The trust premium is the value an organization earns because people trust its AI enough to use it, buy from it and fund it. It has three parts: the losses trust avoids, the returns it brings and the value markets assign to trust they can see. I call them the floor, the engine and the moat.

My claim is that most enterprises underprice it. They book trust as a compliance cost, sized to pass the next audit, when the evidence points to an asset that pays out on lines nobody traces back to it. If you are the CFO or COO setting next year's AI budget, that mismatch is where the money is.

Why the budget gets the price wrong

Trust is underpriced for an accounting reason before it is a moral one. The spending sits on one line: a governance team, a review board, model documentation, access controls. The return lands on other lines, months later, owned by other people. Adoption rises in a business unit. A breach costs less than it would have. A due-diligence call goes smoothly. None of those numbers carries a label saying trust paid for it, so the budget meeting sees only the cost and argues with it.

IBM's respondents describe the same mechanism from the user's side. More than half said ethics-oriented concerns are key barriers to AI, and the report states the consequence plainly: AI that is not trusted does not get used. An unused system has a cost and no return, however accurate it is. Budgeting trust as overhead treats the precondition for adoption as an optional extra.

The mispricing shows up in the companies' own forecasts. For 2024 they had projected AI ethics at 4.3% of AI spend, and they spent 7%. IBM reports that organizations keep revising this line upward, a pattern it says it does not usually see in other technology investments, and executives plan to double its share of AI spend between 2023 and 2026. The survey cannot say whether that overrun was cost arriving early or value found late. Either way, the planners set the line before they understood it.

Three bodies of evidence

What broken trust costs

The floor is the easiest part to measure, because its numbers are the closest thing to invoices. IBM's 2025 Cost of a Data Breach Report found that breaches at organizations with high use of unsanctioned "shadow" AI cost USD 4.74 million on average, USD 670,000 more than at organizations with little or none. Thirteen percent of organizations reported a breach of their AI models or applications, and 97% of those had no AI access controls in place. Since 2 August 2025, the EU AI Act has let regulators fine a prohibited AI practice up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher. That top tier covers practices the Act bans outright; most other breaches of its obligations carry a ceiling of EUR 15 million or 3%. Litigation is following the same curve: AI-related securities class actions in the US rose from 7 filings in 2023 to 15 in 2024, Cornerstone Research counted. Australia's Robodebt scheme, which relied on data matching, automated decision-making and income averaging to raise welfare debts, is a government case. It still prices a risk any enterprise takes when it automates decisions about people: every wrong decision made at scale has to be found, reversed and repaid. The Royal Commission recorded refunds of A$746 million to some 381,000 people and A$1.751 billion in debts written off, and the refunds returned money the scheme had taken from the people it decided about.

What earned trust returns

The engine is harder to measure and the numbers are softer, so read them that way. The 30% operating-profit gap in IBM's survey is a correlation between spending quartiles, and IBM itself says ethics spending explains only "a relatively modest portion" of AI's benefit. In the same survey, organizations reported 12-month improvements they put down to their AI ethics work: 22% in customer satisfaction and retention, 20% in preventing AI ethics incidents and 19% in AI adoption. PwC's 2025 survey of 310 US business leaders points the same way: 58% said responsible AI improves return on investment and efficiency, and 55% said it improves customer experience and innovation. These are executives describing their own results. On its own that is weak evidence. It earns its place because it agrees with the harder numbers on either side of it.

What demonstrated trust is worth

The moat shows up where outsiders can see it. MIT's Center for Information Systems Research found that companies whose boards were both digitally and AI savvy had an average return on equity 10.9 percentage points above their industry average, while companies with non-savvy boards averaged 3.8 points below. Only 26% of boards met that bar in 2024. Board savvy is not trust itself, and the link is another correlation. It belongs here because a savvy board is one of the few signs of AI governance an outsider can check. Beyond the boardroom, the 2025 Edelman Trust Barometer found that 49% of people across 28 markets say they trust AI, falling to 32% in the United States and rising to 72% in China. A company that can show its AI deserves trust is selling into a market where roughly half of people do not yet say they trust AI at all.

What none of them prove

None of this proves that spending on trust causes profit. Every return figure above is a correlation. The likeliest rival explanation is simple: well-run companies do most things well, governance included, and would have earned their profits anyway. Profitable companies can also afford a bigger ethics budget, which runs the arrow backwards. The surveys add a bias of their own, because executives who funded a program have reasons to say it worked. And the samples hold only the companies still around to answer. Firms that invested in trust and failed anyway are missing from the data. The windows are short, too. Most of these figures cover a year or two, so they can't yet show whether the returns hold over the longer periods in which trust is supposed to build.

Edelman's figures carry a different caution. A survey answer about trust is not a purchase, and none of these sources measures what people will pay for AI they trust.

So the strongest objection runs like this: if the upside is unproven, fund the minimum the regulator demands and spend the rest on capability. My answer is that the floor rests on firmer ground than the upside. An Article 99 fine or the Robodebt refunds follow directly from the failures they answer. The breach figure is softer, a comparison between groups of breached organizations, but it counts losses rather than opinions. A budget that covers only the floor still has to price that floor, which means adding up the exposure system by system. Above the floor, the evidence is directional. Sources using different methods and pointing the same way is what most capital decisions run on. Nobody asks for a randomized trial before funding a brand. The practical answer is to treat the upside as an option rather than a forecast: price the floor in full, then fund the work above it in stages, each one tied to a measured change in adoption or, for systems that decide about people rather than serve them, in complaints and appeals.

A test for the budget meeting

Three questions will tell you whether your organization prices trust or ignores it.

1. What share of this year's AI spend goes to trust work? Count governance, documentation, monitoring, access control and training. IBM's respondents reported 7% in 2024, and the quarter with the higher profits spent 10% or more. IBM asked about AI ethics spending, so expect your count to differ from theirs. The number matters less than whether anyone can produce it, and if nobody can, that's the first finding.
2. Who owns the floor? Name the person who can state, for your three highest-risk AI systems, the worst-case fine, the likely cost of a breach and the cost of unwinding a bad decision at scale. If that person doesn't exist, the floor is unpriced. In IBM's survey, 53% of executives said their governing bodies for AI ethics are not effective, so a committee on the chart is not an answer.
3. Which part is weakest? A solid floor with no adoption is insurance with no return. Strong adoption on an unguarded floor is a return you may have to give back.

To answer the third question properly, score your organization on the Trust Premium Score: fifteen dimensions across the three parts, scored on what runs today, with a 90-day plan for the weakest. Then read the two together. A low share with a weak floor means the floor is unfunded. A high share with a low score means the money is going to the wrong parts. Fund the weakest part first, in a stage tied to a measured change.