70.1% of respondents report that an enterprise AI application produced a materially wrong answer or action at least monthly in the previous 12 months
Paligo research finds seven in ten respondents report materially wrong AI answers or actions at least monthly.
STOCKHOLM, Nov. 12, 2026 (GLOBE NEWSWIRE) — AI errors are adding to business costs and delaying projects, according to new research from Paligo, a provider of governed context infrastructure for Enterprise AI. The study found that 14% of respondents report direct financial loss or revenue impact from incorrect or unreliable AI outputs.
More than one in five (22%) report internal process or project delays, while 18% report additional technology or computing costs. The findings cover respondents’ experiences during the previous 12 months.
At least 78 of the 411 respondents (19%) reported a consequence for safety, compliance or reputation. Furthermore, across the full sample, 49% reported a safety incident or increased safety risk (17%), a compliance, legal or regulatory issue (17%) or an impact on their organisation’s reputation (15%).
The international study surveyed 411 business leaders across the UK, US, Germany and Benelux involved in decisions about enterprise AI or the information it uses. Its findings appear in Paligo’s new report, What does it cost an organization when AI gets the answer wrong?
Rahul Yadav, CEO at Paligo, said: “Most AI errors aren’t model problems. They’re content problems. When an AI pulls an outdated manual or the wrong product version, it gives a wrong answer in a confident voice.
“Every wrong answer has a bill attached, and most organisations aren’t reading it. Half of those who hit errors spend five or more employee hours putting each one right, yet only around a quarter measure the time spent checking answers. That’s a potential context tax, paid quietly every day.”
Seven in ten respondents (70%) say an enterprise AI application in their organisation produced a materially wrong answer or action at least monthly. Only 15.6% say their organisation measures the cost of reaching a reliable or verified answer. One in five (20.4%) measure the cost per AI request and 16.1% track token consumption.
Some AI services charge for input and output tokens, the units that measure how much information a model reads and generates. Extra requests or unnecessary material can add to the bill. Nearly three-quarters of respondents (74.2%) say their systems frequently or occasionally retrieve the same or similar information. Processing that material again can use extra tokens, though a repeat check may be necessary to verify an answer.
Problems with the information AI uses can complicate efforts to get reliable results. Some 27% of respondents say investigations have identified outdated retrieved information, while 25% identify relevant information from multiple sources being combined incorrectly. Asked about applications in production or near production, only 22% say their organisation can always identify which version of a source was used.
Rahul Yadav continued: “Finding relevant information is only the beginning. A document can be accurate for one product and wrong for another. Before leaders trust an AI answer, they need to know which version it came from, whether it was approved and why it applied. If you can’t answer that, you can’t defend the answer.”
Verification and correction also consume business resources. Among the 334 respondents who encountered errors, 51% say a typical issue takes at least five employee hours in total to investigate, verify and correct. Some 83% of all respondents report frequent or occasional checks against original sources, but only 27% say their organisation measures the time spent on verification.
Content problems are also delaying AI deployment. More than three in five respondents (61%) report frequent or occasional delays to deploying or expanding AI because supporting information was insufficiently current, approved, structured or governed.
“The fastest way to more reliable AI is to fix the content at the source,” Rahul Yadav added.
“Make it current, approved and traceable, and have the AI flag a question or decline when no approved answer exists. Start with one product or one high-risk answer, and measure what changes.”
Paligo recommends recording token use and charges across every request needed to reach a usable answer, alongside the time staff spend checking and correcting it. Where missing context or conflicting versions prompt repeat requests, teams can test whether improving the source material or how it is retrieved lowers costs while keeping answers reliable.
Read the full report: What does it cost an organization when AI gets the answer wrong?
Notes to editors
For more information on Paligo, contact Stephen Sanders, Kiran Patel or Adam Gooseman at PR Agency One. T: +44 161 871 9140
E: paligo@pragencyone.co.uk
Methodology
The research was commissioned by Paligo and surveyed 411 people in October 2026 across the UK, US, Germany and Benelux involved in decisions about enterprise AI or the information it uses. Respondents covered experimentation, limited production and production across multiple functions.
The 70% figure is the share of respondents who reported materially wrong outputs at least monthly during the previous 12 months. The findings on correction time cover 334 respondents who encountered errors, including those who did so less than monthly. Employee hours measure the combined time staff spent on an issue.
Respondents could select several causes and consequences. The minimum of 78 people reporting a safety, compliance or reputational impact counts the largest of those three groups in each market (26 in the UK, 21 in the US, 20 in Germany and 11 in Benelux). Token economics figures are given to one decimal place. Other survey percentages are rounded to the nearest whole number.
The report describes a potential “context tax” in additional verification and retrieval work. The survey did not quantify token waste, potential savings, the overall financial cost or AI’s net effect on productivity.
About Paligo
Paligo provides governed context infrastructure for enterprise AI. Organisations in manufacturing, medtech and other regulated industries use Paligo to keep their documentation and operating knowledge structured, current and approved, so the people and AI systems that depend on it work from information they can trust. Founded in 2015, Paligo serves enterprises where an inaccurate answer can affect safety, compliance or production.
For more information, visit paligo.net.
A photo accompanying this announcement is available at https://www.globenewswire.com/NewsRoom/AttachmentNg/a4e2d5ff-d6ba-4833-a4be-27c3a531a7e1
