I Closed the Laptop

The third email that week. Take your interview anytime. Our AI is available 24/7.

So I sat down at my desk in Tenerife and switched on the webcam. Before the avatar started, the screen showed me the rules: don’t switch tabs. Don’t alt-tab. Don’t look away. Don’t use your phone. Stay there, in front of the recording, and answer the questions.

About 15 minutes in, the avatar started lagging. I repeated my answer. It lagged again. I slowed down, repeated it more carefully, and at some point caught myself feeling something close to empathy. For a machine that has none.

That was the strange part. I was making myself more patient for software that could not be patient back. I was adjusting my pace, my tone, and my attention for something that could not return any of it.

That was the moment.

I closed the laptop. And I felt relief.

Then I added the company to a list. The list of companies I will never work with again, regardless of whether the relationship is an employee role, a contract, or a freelance engagement.

That distinction matters. Not everyone in a hiring process is in the same position. Some candidates already have a job and are testing the market. Some are unemployed and cannot afford to refuse a bad process. Some are contractors or freelancers, professionals choosing clients as much as clients choose them.

I am closer to the third group. I am not begging for permission to work. I am deciding who I am willing to work with.

I am writing this because that list is getting long, and because I am not the audience this matters most for. I have more than 16 years of experience. I am not a junior trying to prove my value to the market at any cost. I can close the laptop.

The vast majority of people sitting in front of these systems right now cannot. That power imbalance, the candidate exposed and the company performing nothing, is the actual violation. It is not the technology. It is what the technology lets a company avoid doing.

Who I am writing as

For the last 16 years, I have worn different hats across different businesses.

I have been working with AI chatbots since before ChatGPT became the thing everyone was suddenly talking about. Before the viral moment, I was prototyping chatbots for the travel industry and experimenting with automations built through nodes, rules, statistics, and probabilities. Back then, it felt less like magic and more like engineering small decision systems, step by step.

I was in Barcelona, surrounded by a community of builders, meetups, and people who were already paying attention. The people who lived through that earlier chatbot era understood something others missed: the potential was already there, even before the interface became simple enough for the rest of the world to see it.

That is why some of us have been placing bets on AI for the last five years. I was also close enough to the field to remember, in 2018, that Amazon scrapped its internal AI hiring tool after it taught itself to penalize CVs containing the word “women’s.” That story is not new. The mistake the industry made was to treat it as a fixable bug instead of as a permanent property of the approach.

Today, I am positioning myself as a go-to-market and business engineer for AI deployments. I coach executives and founders on where AI belongs, where it does not, and how their own identity has to shift when the work starts changing.

My previous PM role was impacted so much by AI that a 14-day sprint can now feel like a two-day hackathon, or even a few hours of focused work depending on the complexity. AI did not replace the job. It compressed parts of it. It changed the shape of the work. Once that happens, old job titles start to matter less than the judgment, context, and responsibility behind them.

So when I talk to executives, I am not trying to make AI look evil. I am trying to draw the line between useful deployment and lazy substitution. The avatar interview is not a borderline case. It is the cleanest example of putting AI in the wrong place that I can offer you.

This is not a complaint. This is a diagnosis.

3 abuses, compounded

The avatar is the visible one. Two others come before it, doing more damage with less attention.

Abuse 1 - the ATS. A poorly configured Applicant Tracking System is not a filter. It is a discrimination machine that nobody audits. A 2021 study by Harvard Business School and Accenture, Hidden Workers: Untapped Talent, found that 88% of employers acknowledge qualified, high-skilled candidates are routinely rejected because they do not exactly match the configured criteria. The same research estimated that 27 million Americans had been filtered out of jobs they were qualified to do, not because they could not do the work, but because the system did not recognize them. Today, over 99% of Fortune 500 companies use ATS. A 2024 Gallup survey found that 93% of Fortune 500 Chief Human Resource Officers have integrated AI into their hiring process, but only about a third of their own employees know it.

A 2023 NBER working paper went further. When researchers fed AI screening tools identical CVs that differed only in the candidate’s name, candidates with names perceived as Black received 14% fewer positive recommendations than those with names perceived as white. The system has learned bias. It applies it consistently. Nobody can ask it why. A 2025 paper in the Journal of Law and Society documented that recruiters can intentionally use AI hiring systems to filter out candidates from protected groups while keeping the motivation invisible. The discrimination did not disappear. It got better at hiding.

Abuse 2 - the job advertisement. Read enough of these, and the fiction becomes obvious. Five years of ChatGPT experience required. ChatGPT was released three years ago. The hiring manager probably wants something specific. HR, with or without AI assistance, writes something else. The 2021 Harvard study found that 49% of companies admitted they automatically reject candidates with employment gaps of six months or longer. Companies are excluding people for reasons that have nothing to do with the work. They are also writing job descriptions that almost nobody can match, then wondering why they cannot find the right people. In many cases, the people they need were already filtered out.

Abuse 3 - the avatar. This is the final layer. By the time a candidate gets here, the ATS has rejected many applicants for the wrong word, and the job description has scared off many of the rest. The remaining few sit in front of a synthetic face and perform the most human qualities they have: communication, presence, judgment, adaptability, patience. They perform all of this for an audience that cannot perceive any of it.

The interview is no longer a conversation. It is a one-way data extraction.

These three abuses are stacked on top of each other. The candidate writes a CV the ATS will not reject. The job description is fiction. The interview is performed for a face that does not exist. By the time a real human enters the loop, if one ever does, the candidate has already been processed, ranked, and reduced to a vector.

This is what people are calling “efficiency”.

What efficiency actually looks like in this market

The companies running this stack think they removed the bottleneck. They just moved it.

The bottleneck used to be the time it took to identify good candidates. It is now the rate at which good candidates refuse to play. I cannot give you a clean number on that, because companies designed these systems specifically not to see it.

What I can tell you is what I see in my own network. Senior contractors with real track records, exactly the people you would want, look at an avatar interview invitation and quietly redirect to the competition. Without explaining. The companies running these processes never know who they lost or why.

In Europe, depending on the role, a remote knowledge-worker position can get close to a thousand applicants in 24 hours. The market feels saturated, so companies feel powerful. They are not powerful. They are sorting by desperation.

The candidate who cannot afford to refuse the avatar interview is not the same person as the candidate who would have been an exceptional hire. Some of those overlap. Many do not.

And the saturation is not a temporary post-pandemic correction. In the first quarter of 2026, between 78,000 and 92,000 tech workers were laid off globally, depending on which tracker you trust. About 76% of those losses were in the United States. According to Nikkei Asia’s reporting, roughly 47.9% of Q1 2026 tech layoffs were attributed by the companies themselves to AI and automation. Salesforce cut 4,000 customer support roles in late 2025 with Marc Benioff stating publicly that he needs “less heads.” Atlassian eliminated 10% of its workforce on a single day in March 2026. Block ran the math and decided its AI systems could resolve 70 to 80% of customer inquiries without humans, and acted on it.

Some of this is real productivity displacement. Some is what Sam Altman, of all people, has called AI washing, using AI as the public reason for layoffs that would have happened anyway. Cognizant’s Chief AI Officer, Babak Hodjat, has said publicly that he doubts most of these cuts are tied to verified productivity gains. Either way, the practical effect on the candidate side is identical. The market is full. Companies feel they can afford to insult the people standing in front of them.

This is the macro picture against which the avatar interview is being deployed: a labor market under structural compression, with both real and performative AI displacement, and with the leverage flowing one direction.

The companies installing the synthetic face are taking advantage of the moment. The moment will not last forever.

What is actually being optimized

The pitch for AI in hiring is always the same: it removes bias, it scales, it ensures consistency.

From where I sit, after years of working with automation and AI deployments, bias removal is the first lie. Not the kind of AI consulting where people copy workflows from YouTube and call it expertise. I mean the boring part: testing systems, seeing where they break, understanding what happens when automation meets real operations, real incentives, and real people.

A model trained on historical hiring data does not remove bias. It launders it into a place nobody can audit. The candidate cannot ask the AI avatar why she was ranked 8 out of 10. The avatar does not know either. HR does not know. The auditor does not know. The Sheard study from 2025 made this point cleanly: AI hiring systems make existing discrimination less visible while preserving its effects.

Scaling is real, and it is the wrong thing to optimize. Screening 10,000 candidates instead of 50 processes more volume. It does not make better contact with any of them. The qualities that hiring actually requires, judgment, empathy, trust, and the ability to work with the people already on the team, show up through interaction, not throughput.

The consistency argument is the most insulting. Asking every candidate the same question in the same way sounds fair until you remember that the exceptional candidate is usually the one who answers a question you did not ask. An AI avatar cannot follow that thread. So the system ends up consistent at finding consistent people, which is exactly what most companies do not actually need.

What is being optimized in an avatar interview is not the candidate experience or even the hiring decision. It is the company’s convenience. Setting up a real interview takes a calendar, a person, attention, and time. The avatar removes all of that. The company saves a few hours per role and tells itself it has “innovated”. What it has actually done is offload the part of hiring that requires it to show up.

This is what convenience-based deployment looks like everywhere AI is being misused right now. A company decides it does not have time, does not have the headcount, does not have the patience to do something properly, so it automates the appearance of doing it.

The result is not efficiency. It is paying for a Ferrari to drive 10 meters to the corner shop.

And the same companies will then publish a careers page about their values, their inclusive hiring practices, and their commitment to diversity. The marketing, PR, communication, and sustainability budget for the values page is real. The compliance budget for the avatar tool is real. The investment in actually meeting the candidates? That is what got cut.

Welcome to greenwashing applied to human resources.

The deeper pattern

The avatar interview is not the disease. It is the first visible symptom of a wider system: a labor market learning to extract human judgment, behavior, and attention without giving much humanity back.

Three patterns matter here.

Cognition outsourcing. A June 2025 MIT Media Lab study by Nataliya Kosmyna and colleagues, Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (arXiv:2506.08872), found that participants who relied on LLMs to help write essays showed weaker neural activation, lower memory of what they had produced, and what the researchers called cognitive debt. The convenience came at a measurable cost.

We are not just outsourcing tasks. We are outsourcing the thinking that used to come with them. Job seekers feel this directly when they are pushed into AI-rewritten CVs and AI-coached interview answers, performing for AI graders. The whole system becomes a closed loop in which nobody is actually thinking about the question of whether this person can do the job.

Work becoming training data. A growing share of employee work can now be recorded or reconstructed from screen activity, mouse clicks, voice, decisions, habits, tools, prompts, and documents. Tools like ScreenPipe, a local, free, open-source screen and audio memory project, make the personal version of this obvious: you can search your own work, document your habits, and give agents context about what you actually do.

That can be useful. It can help you save time, document your own process, and spend less time in front of a screen. But inside companies, the same logic has a sharper edge. First, the work is captured. Then repeated patterns become workflows. Then workflows become skills, automations, SOPs, or agents. The goal is not to capture the whole person. A screen cannot capture taste, judgment, trust, political intelligence, or responsibility. But it can capture enough visible work to make replication economically tempting.

That bigger story deserves its own article: Your Work Is Becoming Training Data. For this piece, the point is simpler. The same logic that replaces the interviewer with an avatar is moving inside work itself. Observe the human. Extract the pattern. Automate the interaction. Then call it efficiency.

The fear of being obsolete. White-collar workers are now putting in 12 to 14-hour days because AI is around, not in spite of it. The mechanism is what some have started to call FOBO, the fear of being obsolete. The cost of producing artifacts, reports, decks, summaries, content, spreadsheets, and research, is collapsing toward zero.

The bottleneck has moved from execution to optionality: too many things you could be doing, too many tools to choose between, too many parallel threads. The illusion is more output, more options, more leverage. The reality, for the worker, is exhaustion. Companies are extracting the productivity gain. The worker is absorbing the cognitive cost.

This is also why token budgets will start to matter in hiring. In some white-collar roles, staying ahead with AI will depend on whether the company gives people enough access to experiment, burn tokens, test tools, build artifacts, and learn. A company that expects AI-native output but gives no real AI budget is asking people to compete in a race while hiding the fuel.

So the avatar interview lives inside a larger system: cognition outsourced, work captured, workers exhausted, and companies rewarded for removing the human parts of the process first.

Naming the AI avatar as the symptom is correct. Pretending the AI avatar is the disease is not.

What the law is about to do about it

Start with Europe. The European Union is not going to let this run indefinitely.

Article 26 of the EU AI Act is unambiguous. Once the deployer obligations come into force on August 2, 2026, a company using a high-risk AI system in employment decisions has documented responsibilities: a named human with the authority to intervene in any decision the system makes, a Fundamental Rights Impact Assessment when the system affects access to employment, and audit-grade records of how the system is actually being used in practice. The penalties for non-compliance scale higher than GDPR. We are talking up to €35 million or 7% of global annual turnover for the most serious violations.

Sectors with regulatory weight, healthcare, finance, government contracting, will be enforced more strictly. Companies operating in any of those sectors and still using avatar interviews are not running an HR experiment. They are holding a compliance liability that gets hotter every month.

The United States is moving differently. It is not aligned around one federal AI law. It is a patchwork. Colorado is the clearest example: its AI Act is aimed at high-risk systems used in consequential decisions, including employment, and is scheduled to take effect in 2026. But the federal government is already pushing back against state-level AI rules, arguing that fragmented regulation can slow American competitiveness in the race against China. That means companies will do what companies often do: move toward the least restrictive jurisdiction, at least until the law becomes federal, Congress acts, or a future administration changes direction.

China is not waiting either. It does not yet have one EU-style AI Act, but it already regulates algorithmic recommendation systems, deep synthesis, generative AI services, and human-like interactive AI services. So the global split is not “regulated Europe versus unregulated everyone else.” It is more uncomfortable than that. Europe is regulating through rights and compliance. The United States is trying to preserve speed and competitiveness while fighting over state versus federal control. China is regulating AI through platform control, content control, and state supervision.

Different philosophies, same direction: AI used on people will not stay outside the law forever.

The argument that “we used a vendor tool” is not going to shield companies for much longer.

This is not a future problem. The window for a company to fix its hiring stack quietly, before anyone files anything, is approximately the next 12 months.

What working with AI in hiring actually looks like

There is a version of this that works. AI in the background. Humans in the conversation.

This is the part people get wrong when they try to make AI look evil. AI is not the problem when it helps where help is actually needed. Use it to clean the grammar of a job description. Use it to compare CVs against clear criteria. Use it to summarize experience before the call. Use it to prepare better questions. Use it to take notes during the interview, draft a humane rejection letter, handle scheduling, or surface references.

All of that is valid.

The line is the interview itself. That moment still has to be two people, both present, both exposed enough to make a real judgment. AI can prepare the room. AI can document what happened in the room. AI should not replace the human being in the room, or on Zoom.

That is the pattern.

In my own practice, I draw a clear line for businesses between where AI belongs and where it does not. Back-office automation, scheduling, summarization, document drafting, knowledge retrieval: yes. Decisions about who is qualified to work with the team: no.

The line is not technical. It is about who carries the responsibility when the decision is wrong. AI cannot carry it. The manager who deployed it can, and should.

I am not nostalgic about inefficient hiring. Plenty of human-only processes are slow, biased, and unfair. The fix is a better human process with AI handling friction in the right places. Not a synthetic face performing humanity while a vendor collects the candidate’s data.

What candidates can do right now

This is the part that is rarely written down. If you are on the candidate side of this, you have more leverage than the system pretends.

Build your own list. Make a private record of every company that sends you an avatar interview, an AI-coached video assessment, or a four-stage automated screening process before any human contact. Note the date, the platform, the role. Three months from now, that list is your filter. You will see the patterns: which companies are doing this, which industries are doing it most, which roles are being processed without human oversight. The list will tell you who is worth engaging with.

Use your data rights. Under GDPR, if you are an EU resident and a company has processed your application through an automated system, you have the right to request your data and, in many cases, to request deletion under Article 17. The company generally has 30 days to respond. If it does not, you can file a complaint with your national data protection authority.

This is where AI should also work for the candidate, not only for the company. Businesses use automation to lower their costs, increase their leverage, and make legal processes feel like normal operations. Individuals could learn to use the same tools to lower the cost of defending their rights.

Nothing personal. For a company, answering a formal request or going through a legal process is business as usual. It should become business as usual for citizens, too.

Like the healthcare system, the justice system has always been easier to access for people and companies that can afford lawyers, fines, delays, and paperwork. AI changes part of that equation. It can help draft a GDPR access request, find the privacy email hidden in the company’s legal pages, prepare a deletion request, track the 30-day deadline, and assemble a complaint if the company ignores it.

Most companies make themselves very visible when they want applicants, customers, or attention. But when you need to contact them directly, suddenly there is a chatbot, a no-reply email, or no obvious person. The contact point is often buried in the privacy policy, legal notice, terms page, or GDPR section. Somewhere, they normally have to tell you who controls your data and how to contact them.

A small open-source tool could make this easier. It could take a company website, find the legal or privacy contact, generate the data-access or deletion request, log the deadline, and prepare the next step if there is no answer. Today, that can be done with browser agents, scraping, or agentic web tools. Tomorrow, companies may need pages that are easier for AI agents to read directly, whether through structured privacy pages, machine-readable notices, or emerging files like llms.txt.

If companies are going to use AI agents as proxies for business, candidates, and citizens will need AI agents as proxies for their rights.

Ask the question on the call. If you do reach a human, ask them where in the process AI was used and how. If they can’t tell you, that is the answer.
If they tell you and you don’t like it, that is also the answer.

Refuse the avatar. When you have the option to refuse, refuse. The companies need to feel that this is costing them. The market will not correct itself if every candidate plays along.

A vague job advertisement promising a “competitive salary” without specifics is recruiting mercenaries, not missionaries. The mercenaries will leave the moment a better number comes through. The missionaries, the people who would have built something with the company, never applied because the description did not address them. The company will then complain about retention.
They have built the problem they are now describing.

The harder question

What is happening in white-collar hiring right now is not just inefficiency. It is a power dynamic. One side has been told, repeatedly, that they are interchangeable, replaceable, ranked by an algorithm, and not worth a real conversation. The other side has been allowed to outsource the discomfort of meeting another human and making a judgment about them.

The managers responsible for deploying these systems are not victims of technology. They made a procurement decision. They wrote the cheque. They approved the vendor. They have names and titles. The greenwashing budget, the careers page about values, the diversity statements, and the ESG reporting are run by the same people who signed off on the avatar contract.

The AI is a car. It does what the driver tells it. The drivers in this case are the executives and managers who decided that humans on the other side of the hiring process are interchangeable enough that a synthetic face will do the job. They are accountable. The AI is not.

I am not approaching this as a left-wing position or a union position. This is a justice question that is now technically tractable. The EU AI Act, the GDPR, state-level US legislation, and visible class-action exposure are converging on the same target. There is a window, roughly the next 12 to 18 months, for companies to fix this on their own terms before regulators do it for them on much harder terms.

If you are a manager who has set up an avatar interview process, and the system is still running, I have one question that does not require an answer from me.

When was the last time you sat through one yourself?

Not as the reviewer. As the candidate. With your career on the line. Performing for a face that will not remember you. With instructions not to look away.

If you cannot bring yourself to take the test you are giving other people, you already know.

Close the laptop.

Otman, AKA El Capitano, is a GTM (Go To Market) Business Engineer and Automation Deployment Strategist, coaching and mentoring operators and executives on EU AI Act readiness, AI deployment, automation risk, ethics, identity transition, and the question of where AI belongs in their operations and where it does not.
The window to fix a hiring stack quietly closes in 2026.

Sources referenced

• Kosmyna, N., Hauptmann, E., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. arXiv:2506.08872.
• Fuller, J. & Raman, M. (2021). Hidden Workers: Untapped Talent. Harvard Business School & Accenture.
• Sheard, N. (2025). Algorithm-facilitated discrimination: a socio-legal study of the use by employers of artificial intelligence hiring systems. Journal of Law and Society.
• Hoffman, M., Kahn, L. B., & Li, D. (2023). Discretion in Hiring. NBER Working Paper.
• Nikkei Asia (April 2026). Q1 2026 tech layoff data.
• Layoffs.fyi (2026). Tech layoff tracker, Q1 2026.
  92,272 tech employees laid off ∙ 98 tech companies w/ layoffs
• karpathy.ai/jobs/ Andrej Karpathy US Job Market Visualizer
• anthropic.com/research/81k-economics Maxim Massenkoff and Saffron Huang (April 2026). What 81,000 people told us about the economics of AI. Anthropic. Economic Research
• Gallup (2024). Fortune 500 CHRO AI integration survey.
• WEF (2025). Future of Jobs Report 2025.
• Dastin, J. (2018). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters.
• EU AI Act, Article 26 (deployer obligations, enforcement Aug 2, 2026).
• GDPR, Article 17 (right to erasure).

Tools / projects referenced

• ScreenPipe. Open-source local screen and audio memory for searching and understanding your own work. GitHub: https://github.com/screenpipe/screenpipe. Website: https://screenpi.pe
• TinyFish. Web-agent infrastructure for AI agents that can navigate, authenticate, extract data, and automate web workflows. Website: https://www.tinyfish.ai
• Public GitHub examples: https://github.com/tinyfish-io.
• Microsoft 365 Copilot / Microsoft Graph / Viva Insights / Viva Glint. Microsoft’s enterprise workplace stack is not a simple “digital twin” product, but it is important because it connects Copilot usage, work signals, employee experience data, and organizational analytics. This is the kind of enterprise infrastructure that can make work patterns visible at scale.
• Claude Cowork. Anthropic’s agentic AI for knowledge work. It is designed to handle tasks autonomously across a user’s computer, local files, and applications. Website: https://www.anthropic.com/product/claude-cowork.
• Claude Code. Anthropic’s AI coding assistant for understanding codebases, editing files, automating development tasks, and working across multiple tools. Docs: https://code.claude.com/docs/en/overview.
• Claude Computer Use. Anthropic’s computer-use tooling for letting Claude interact with a sandboxed computing environment through tool calls. Docs: https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool.
• OpenAI Codex / Codex Computer Use. OpenAI’s coding and computer-use environment. Computer Use in the Codex app lets Codex use desktop apps on macOS, with availability restrictions in the EEA, UK, and Switzerland at launch. Docs: https://developers.openai.com/codex/app/computer-use.
• OpenClaw. Open-source personal AI assistant / agent that can run on your own devices and interact through familiar channels. GitHub: https://github.com/openclaw/openclaw. Website: https://openclaw.ai
• Browser-use style agents. Browser and computer-use agents are part of the wider shift from “AI that answers” to “AI that clicks, navigates, fills forms, and executes workflows.”
• llms.txt. Emerging proposal for making websites easier for AI systems and agents to read, understand, and navigate.

First published on Substack.