In the AI Era, There Is a Growing Anti-AI Trend Nobody Is Talking About

Every software company on earth is racing to add AI. In one corner of the enterprise — compliance, risk, and procurement — a growing class of buyers is racing the other way. They are asking vendors to remove AI, prove it isn't there, or sign that it never will be. Here is the quiet movement reshaping who wins compliance contracts in 2026.

For two years, the only acceptable answer to "what's your AI strategy?" has been a list of features. Every category leader has rushed to put "AI-powered" on the deck. Every challenger has rushed to bolt a model into the workflow. The market has assumed, almost universally, that adding AI is the same thing as adding value.

It isn't — at least not in compliance. While Sumsub, ComplyAdvantage, Middesk, and most of the KYC/KYB category lean harder into "AI-powered" positioning, a growing number of compliance, risk, and procurement teams have started saying so out loud. Not in op-eds or on LinkedIn, but where it matters: in procurement forms, in vendor questionnaires, in the contractual fine print that decides who gets the renewal. They are asking software vendors a new question, and the answer they want is increasingly the opposite of what the rest of the market is racing to deliver.

This is the trend that no one in the AI hype cycle is talking about. It is quiet, it is well-funded, and it is reshaping how compliance contracts get won in 2026. Here is what it looks like in practice.

Two weeks ago, a procurement contact at a major tech buyer sent us an email. They were filling out their internal AI risk form before signing the renewal. The form looked something like this:

The buyer was relieved. That answer is now the easiest one to give in compliance procurement — because every other vendor in the stack is being interrogated on AI risk, and most of them cannot pass.

Welcome to the quiet correction of 2026. After two years of "AI-everywhere" marketing, enterprise buyers are pulling back. Not because AI is useless — but because compliance, risk, and procurement teams have learned the hard way that generated text is not evidence, and the bill for pretending otherwise is now visible on the balance sheet.

The $67 billion problem nobody underwrote

Here is the number the vendor pitch decks leave out.

Global enterprise losses from AI hallucinations reached $67.4 billion in 2024 — direct and indirect costs combined.

That figure, from AllAboutAI's enterprise study and corroborated by Korra's analysis, covers bad business decisions made on hallucinated content, regulatory fines, legal liability, customer trust erosion, and the salaried hours spent cleaning up after AI errors. It is not a worst-case projection. It is what already happened.

The per-employee number is more useful for budget planning. Forrester estimates each enterprise knowledge worker now costs companies roughly $14,200 per year in hallucination mitigation — fact-checking, re-verification, and rework. Microsoft's 2025 productivity data puts the time tax at 4.3 hours per employee per week. Deloitte's Global AI Survey found that 47% of enterprise AI users have already made at least one major business decision based on hallucinated content.

Stack those into a procurement TCO model and the picture is unflattering. A 500-person enterprise running AI tooling broadly is absorbing roughly $7.1M a year in verification overhead, before counting the cost of decisions made on bad information. McKinsey reports that 42% of companies abandoned most AI initiatives in 2025, up from 17% in 2024. MIT's NANDA study found 95% of pilots delivered no measurable P&L impact.

The market spent $644 billion on enterprise AI in 2025. Between 70% and 95% of those pilots, depending on how you count, failed to reach production. The hallucination problem is not an edge case. It is the median outcome.

Why hallucinations are fatal in KYB and supplier verification

Hallucinations are tolerable when the cost of being wrong is low. A summarised email with a slightly off paraphrase is annoying. A creative draft that invents a metaphor is fixable. Compliance is different. Every one of the following has a binary, verifiable answer that lives in a primary source — and getting it wrong is a regulator-facing failure:

You cannot answer these questions from a confidence score. You answer them from Companies House, the IRS, the Bundesanzeiger, the ONRC, OpenSanctions, and 400+ other primary registries. Each record needs the same three things: source attribution, timestamp, and an audit trail. That is what a regulator asks to see — not a "78% confident this is the director" output.

The newer the model, the worse it can get

This is the part that surprises most procurement teams. The intuition that "AI is improving so this will fix itself" is wrong on the tasks that compliance actually cares about.

OpenAI's own data on its o3 reasoning model shows it hallucinated 33% of the time on PersonQA — double the rate of its predecessor. The April 2026 ICLR paper "The Reasoning Trap" formalised the finding: training models for stronger reasoning increases tool-hallucination rates in lockstep with task gains. Smarter models, more confident hallucinations.

Worse, MIT's January 2025 research found that AI models are 34% more likely to use confident language — "definitely," "certainly," "without doubt" — when generating incorrect information than when telling the truth. The wronger the AI, the surer it sounds. That is a catastrophic property in a verification workflow where someone has to decide whether to onboard a counterparty.

The compliance failure modes

The financial damage is one thing; the operational damage is more specific. In KYB and supplier verification, an AI hallucination is not a bad summary or an awkward draft — it is a sanctioned counterparty that gets onboarded because the model invented a clean record. It is a shell entity that passes UBO screening because the model fabricated a beneficial owner. It is an invoice paid to a dissolved supplier. It is an AML report filed on content that does not match any registry record.

None of these errors are recoverable through better prompting. They are inherent to using a generative system in a verification workflow. The same query, asked of an AI versus asked of a registry, returns fundamentally different outputs:

What "verified" actually means in 2026

Vendors throw the word "verified" around. It is mostly a marketing term. To compliance buyers — and increasingly to the procurement contracts they sign — it now has a specific meaning. Across the compliance and procurement conversations we have had over the past two quarters, the same five properties keep showing up. The first three are now appearing as contractual requirements, especially for firms operating under EU AML6, the U.S. Corporate Transparency Act, and the EU AI Act's high-risk system rules taking full effect on 2 August 2026.

1. Source attribution. The exact registry, filing, or authority that produced each field. Not "our database." Companies House, the IRS, Bundesanzeiger, ONRC, OpenSanctions — name the source. Buyers want a registry name, a URL or filing ID, and a date — not a vendor logo.
2. Timestamps and recency guarantees. Two timestamps, not one: when the record was retrieved from the source, and when the source itself last updated. A registry record from 18 months ago is a liability. Static datasets fail this test, and so do AI tools that synthesise from cached web pages of unknown vintage.
3. Audit trail. A reproducible reference — filing ID, registry URL, document ID — that another reviewer can independently verify. The question being asked in vendor reviews is no longer "Does this tool work?" It is "Will this tool withstand scrutiny if challenged?" AI-generated outputs cannot reproduce themselves identically on a second run, which means they cannot be audited.
4. Versioning. Historical states of the record, so you can show what was true on the date of the onboarding decision — even if the record has changed since.
5. Standardisation. The same field, in the same format, across every jurisdiction. UBO data from Germany should be parseable identically to UBO data from Mexico.

None of these five properties are satisfied by a generated summary. All five are satisfied by registry-sourced data with a documented pipeline. This is the gap.

Why this trend is structural, not a fad

The pendulum is swinging. Three forces are pushing compliance buyers off the AI-everything pitch and back toward verified primary sources — and none of them are reversing in 2026.

Regulators are catching up. The EU AI Act's high-risk system obligations take full effect on 2 August 2026. Compliance and KYB tooling that makes autonomous decisions about people or businesses falls inside that scope. Vendors that can answer "no, this is registry data, not a model output" simply have less compliance burden — and their buyers know it.

Procurement is becoming AI-aware. The procurement email at the top of this article is now standard. AI-usage disclosure is a new section in vendor onboarding forms. Compliance buyers are scoring vendors on whether they use AI for the answer or for the workflow around the answer. Models are fine for routing tickets. They are not fine for deciding if a supplier exists.

The CFO has noticed. When 95% of AI pilots deliver no P&L impact (per the MIT NANDA study) and verification overhead runs $14,200 per employee per year, the AI line item gets a column in the next budget review. Aon's 2026 AI risk report explicitly recommends "human-in-the-loop review for high-stakes AI outputs" — which, for KYB, means the loop never closes if the underlying data was hallucinated to begin with. The cheaper architecture is to skip the model and pull the registry record directly.

The takeaway

AI is genuinely useful for the workflow around compliance — routing alerts, summarising case files, drafting internal notes. It is not useful for the answer. The answer to "is this company registered?" lives in a registry. The answer to "who is the UBO?" lives in a beneficial ownership filing. The answer to "are they sanctioned?" lives on a list with a date.

Enterprises burned $67.4 billion in 2024 learning that lesson. 2026 is the year compliance buyers stopped paying the tuition — and started rewarding the vendors who never asked them to.