98% of Large US Companies Have AI Rules, but 47% Bypassed Them in a Rush
Businesses are firing specialists to make way for AI agents, but they have not settled who answers for the agents' mistakes. According to an EY survey, 98% of large US companies have written rules for working with AI, yet 47% failed to apply them at least once during a rushed rollout. It is like a safety manual that gets shelved whenever deadlines are on fire. So before any layoffs, name a responsible person for each agent and install a kill switch.
AI-processed from Habr AI; edited by Hamidun News
Companies are laying off marketers, SEO specialists and programmers on the assumption that AI agents will replace them, but they are not answering a simple question: who is accountable if an agent does something wrong? The EY survey numbers show what this looks like in practice. Written rules for working with AI exist at 98% of large American companies, which means almost all of them.
Yet 47% admitted that during a rushed AI rollout they did not apply their own rules at least once. Think of a safety manual that gets stuffed in a drawer when deadlines are burning. The author of this analysis on Habr, a consultant who works with AI for analytics and website code, raises this question on every project.
It is a personal column, not a study, but it contains plenty of practical ideas.
Rules on Paper and Reality
The piece was prompted by a panel on AI deployment security at Kazan Digital Week. Specialists from banks and cybersecurity firms spent an hour and a half discussing who bears responsibility for an agent's actions. Commenters laughed about it, but the author did not find it funny.
The panel mentioned the EY survey, and the author tracked down the original source and checked the figures. The gap between 98% and 47% shows that having rules guarantees nothing; under deadline pressure, they get bypassed.
From the presentations, the author drew several practical ideas:
- For agents, the reverse principle applies: anything not explicitly permitted is forbidden;
- The model an agent runs on can be swapped for a "poisoned" one, and agents will start causing harm, so external monitoring is needed;
- A kill switch is needed, meaning a button that quickly takes the agent out of operation;
- Each agent must have a responsible human.
A vivid example from the panel: an employee asks to give an agent his electronic signature so the agent can use the CryptoPro plugin on its own. The security officer replies that it is allowed, but for one action only: signing the employee's resignation letter. It was a joke, but the principle is clear: agents should get the minimum rights, scoped to a specific task.
Who Will Be Giving AI Tasks in Five Years
The author's central thesis is that someone has to train agents, assign them tasks and check their results. If you fire the people who can do this, in five years there will be no such people left. The hope that someone else will solve the problem for you, or that models will learn to do everything on their own, the author calls a utopia.
From this he forecasts that the labor of narrow specialists and security experts will become more expensive. He expects the internet to split into "AI-made" and "human" spaces, with paid platforms offering practitioners' knowledge that is closed to neural networks. An indirect sign: social networks have already started labeling generated content, and Meta has been putting an "AI info" tag on such posts since 2024. This is the author's forecast, not an established fact.
He describes his own experience without enthusiasm. He uses AI to speed up processes, but he double-checks everything that goes onto a website, and he occasionally verifies data collected through APIs by hand. He criticizes the idea of a "content factory" producing dozens of texts a week: who will proofread them, and who will make sure the pages get indexed without cannibalizing each other? In SEO, he says, AI delivers "the average temperature of the hospital," and real results begin where you dig deeper.
Two Stories from Practice
The first is about an intern. He arrived with a setup from vibe coders: in SEO, the main thing is to write articles, and an AI assistant will advise on the rest. For six months he ran his project that way, and the site collapsed: articles without assembled semantic cores, poor internal linking, weak product cards, and pages competing with each other. Now the author gives him a foundation in SEO processes, and he learns on his own project. The takeaway: specialists cannot be found "from nowhere"; they have to be grown, and those who start now will come out ahead.
The second is about a voice AI lead qualifier. The owner hired two contractors who promised everything would run like clockwork. In the end, there were lost leads during testing, pre-trial claims and contract termination. The contractors themselves admitted that AI is not suitable for this niche.
What to Do in Practice
The author is not against adoption: without AI, a business will lose to competitors. The question is how to do it. Before replacing a person with an agent, he suggests answering three questions:
- What happens in three months when something goes wrong?
- How much will you lose while marketing is stalled and you look for a new team or agency?
- How much time will it take to clean up what the agents broke?
The limitations are obvious. This is the opinion of one practitioner, the EY survey figures are given without methodology details, and the stories from his projects should not be treated as statistics. But the checklist is workable: assign a responsible person to every agent, give them minimal rights, set up external monitoring and a kill switch, and check key results by hand. And do not fire the people who know how to assign tasks and verify AI's work, because bringing them back later will cost more.
- Nov 21, 2026Open-Weight Inkling 2 Mini Matches Qwen3.8-27B on Independent Benchmark, Runs on 16 GB of Memory
- Nov 21, 2026Qwen3-0.6B Fine-Tuned on a Mac mini: Errors Dropped to 3 out of 30 from 9
- Nov 21, 2026Thinking Machines Releases Open-Weight Inkling 2 Mini with 26.4B Parameters for 32 GB Macs
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
Key Questions Answered
What are the key details on Rules on Paper and Reality?
The piece was prompted by a panel on AI deployment security at Kazan Digital Week. Specialists from banks and cybersecurity firms spent an hour and a half discussing who bears responsibility for an agent's actions. Commenters laughed about it, but the author did not find it funny. The panel mentioned the EY survey, and the author tracked down the original source and checked the figures. The gap between 98% and 47% shows that having rules guarantees nothing; under deadline pressure, they get bypassed.
Who Will Be Giving AI Tasks in Five Years?
The author's central thesis is that someone has to train agents, assign them tasks and check their results. If you fire the people who can do this, in five years there will be no such people left. The hope that someone else will solve the problem for you, or that models will learn to do everything on their own, the author calls a utopia. From this he forecasts that the labor of narrow specialists and security experts will become more expensive.
What are the key details on Two Stories from Practice?
The first is about an intern. He arrived with a setup from vibe coders: in SEO, the main thing is to write articles, and an AI assistant will advise on the rest. For six months he ran his project that way, and the site collapsed: articles without assembled semantic cores, poor internal linking, weak product cards, and pages competing with each other. Now the author gives him a foundation in SEO processes, and he learns on his own project.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.