Use Cases of AI Agents in Business: Real Company Experience in 2026

AI agents in business — are not a marketing concept, but a working tool that is already replacing entire business processes today in companies ranging from fintech startups to American logistics operators. In this article — concrete cases from practitioners: how the Head of AI at fintech unicorn Plata automated interview feedback, how the founder of Wiz replaced a $450/month SaaS with an in-house system in a week, and how Wallet Telegram drove automatic ticket closure to 33% with CSAT above 70%.

AI Summary: AI agent — an autonomous system based on an LLM that plans and executes multistep tasks with access to external tools. Market: $5.2 billion (2024) → $227 billion (2034). Key success principle: specific pain point → reproducible pilot → immediate implementation → measurement.

Contents

1. What an AI agent is and how it differs from a chatbot
2. The AI agent market in 2025: key figures
3. Case #1: Plata — automating routine work for non-developers
4. Case #2: Wallet Telegram — an AI agent in P2P customer support
5. Case #3: Zalando — onboarding through AI and an internal LLM
6. Case #4: Wiz — how the founder replaced SaaS with in-house systems in 2 months
7. Case #5: Triple10 — managing team context through Claude Cowork
8. What distinguishes successful cases from failed ones
9. Table: AI agent vs chatbot vs RPA
10. Step-by-step implementation plan
11. FAQ

What an AI agent is and how it differs from a chatbot

Key Takeaway: A chatbot answers a question. An AI agent completes a task. A chatbot reacts — an agent acts.

An AI agent is a software system based on a large language model (LLM) that can independently plan actions, use external tools (API, databases, browser, email), and carry out multistep tasks without constant human involvement.

The key difference from a traditional chatbot: a chatbot works according to a prewritten script or decision tree. An AI agent builds an action plan in real time based on the assigned task and available tools.

Three signs of a true AI agent:

• Autonomy: the agent determines the sequence of steps needed to achieve the goal itself, rather than following a prewritten script
• Access to tools: the agent can call APIs, read documents, write code, send messages, and enter data into a CRM
• Memory and context: the agent remembers the history of interactions within a session and takes accumulated context into account when making decisions

[Fact]: According to IBM, by the end of 2025 the share of business processes executed by AI agents will grow from 3% to 25%. Every fourth process in a company will be carried out without human involvement.

Important clarification: an AI agent is not a replacement for an employee as such. It replaces specific functions and tasks that can be described, formalized, and measured. Where a task requires unique judgment, empathy, creativity, or managerial decision-making, a human remains indispensable.

The AI agent market in 2025: key figures

Key Takeaway: The market is growing by 45% annually, but 95% of companies do not get ROI. The difference is in the approach, not the tools.

[Fact]: The global AI agent market was worth $5.2 billion in 2024 and, according to forecasts, will reach $227 billion by 2034 at a compound annual growth rate of 45.8%.

[Fact]: According to McKinsey (2025), 88% of companies worldwide use AI in at least one business function — 10 percentage points more than a year earlier. At the same time, only one-third of them have scaled AI solutions to the enterprise level.

[Fact]: 95% of businesses see no return on AI investments (MIT). Only 1 in 20 projects reaches the industrial deployment stage.

[Fact]: A SKOLKOVO School of Management study based on 1,600+ AI implementation cases: companies that start with a pilot and train the team get ROI 3–5 times higher than those that try to automate everything at once.

These figures set the context for the cases below: the technology is available, the market is growing, but most companies either have not started or have started incorrectly. Real cases show how to do it right.

Case #1: Plata — how the Head of AI automated routine work for non-developers

Key Takeaway: Starting with voice input and three simple automations is more important than building a complex personal OS. The result is freeing up cognitive capacity for strategic tasks.

Context

Plata is a fintech company from Latin America valued at more than $5 billion (April 2025). Over the course of a year, the company grew from unicorn status ($1 billion, March 2024) to $5 billion — one of the fastest growth trajectories in LatAm. Masha Poltanova, Head of AI, is responsible for AI adoption among non-developers: marketers, HR, and operations managers.

This is the hardest task in AI transformation: the audience is heterogeneous — some are afraid of code, some are comfortable working with it, and some are embarrassed to ask “stupid questions” about AI. Quick wins matter more here than architectural elegance.

Problem #1: Time lost to typing

Solution: Superwhisper — a voice input app for Mac. It works with any language without switching keyboard layouts, instantly transcribes speech, and inserts text into any app: Slack, Telegram, Google Docs, Claude, browser.

According to Masha: “Now I even switched to my laptop for replying in Slack from my phone — because without Superwhisper, I no longer feel like typing.”

Why this matters more than it seems: Voice input lowers the cost of creating context. When it is easy to record a thought, describe a task, or reply to a message, people give agents more information. The quality of context directly affects the quality of the agent’s work.

Problem #2: Reactive task and commitment management

Through Claude Cowork (the enterprise version of Claude with access to Slack, Google Docs, Calendar), Masha set up three automations. Each one takes just a few minutes to configure:

Automation 1 — Morning automation digest. Every morning Claude analyzes yesterday’s conversations (Slack, Google Docs, Claude) and suggests tasks that can be automated. This is a recursive AI process: the agent looks for more places where agents can be introduced.

Automation 2 — Commitment tracker. Twice a day Claude checks all meetings through Granola (AI notetaker), finds commitments made, and blocks 30 minutes on the calendar to complete them. It solves the classic problem of “promised everything in meetings, forgot to do it.”

Automation 3 — Monday recap. Every Monday at 8:00 Claude compiles a weekly summary: project progress, tests launched, unresolved agreements. It solves “Monday amnesia” — when your head is cleared over the weekend and it is unclear what happened in your previous life.

Problem #3: Preparing interview feedback

Masha personally interviews many candidates at Plata. Every rejection requires structured feedback using a template. The task took significant time and emotional resources.

Solution: After the interview, Masha briefly states her opinion out loud (records it by voice after the call). Through Granola, Claude sees that there was an interview meeting, knows the feedback template from the HR system, and automatically maps Masha’s words into the required fields.

[Fact]: According to Masha Poltanova: “It saves more nerves than time.” A heavy cognitive task was turned into an easy mechanical one.

Case conclusion

"The most important AI skill is thinking about what else can be offloaded to AI with sufficient quality. How do you track that quality, how do you extract from life the things you no longer want to do." — Masha Poltanova, Head of AI, Plata.

Don't aim straight for a complex personal OS. Connect Slack and the calendar, set up 2-3 simple automations. That will kick off the improvement loop.

Case #2: Wallet Telegram — an AI agent in P2P customer support

Key Takeaway: 33% of tickets are fully resolved by the AI agent, CSAT 70-80%. The key to success is short replies, an easy handoff to a human operator, and honest disclosure that the response is from AI.

Context

Wallet is a crypto wallet inside Telegram with more than 5 million users. Zhenya, head of P2P customer service, built the support system from scratch. Fintech is a difficult domain: high security requirements, a wide variety of questions about transactions, P2P deals, and verification.

Agent architecture (RAG approach)

1. Knowledge base — FAQ from Helpscout is broken into small paragraphs ("clear question — clear answer")
2. Vector search — when an incoming question arrives, the system searches for relevant fragments via embedding
3. Answer generation — GPT-4 (via a corporate API key) formulates the answer in its own words based on the retrieved fragments; the agent works only with what it found in the articles
4. Confidence scoring — each answer receives a confidence score from 0 to 1:
   • ≥ 0.7 → the answer is sent to the user
   • < 0.5 → immediate transfer to a human operator

Filtering rules: when the agent does not respond

• The user sent an image → to a human (the model does not process images)
• The user wrote "call an operator" → to a human
• Two questions were asked at once → low confidence, to a human
• Personal data in the message → to an operator

[Fact]: The agent does not hide the fact that it is AI. It immediately says: "I am a virtual assistant." This sets the right expectations.

Key metrics

[Fact]: Full AI Resolution — about 33% of tickets are fully resolved by the agent. This is a confirmed external benchmark: "We have 25%" — commented audience member Anton Yeryomenko.

[Fact]: CSAT of the AI agent is 70-80%. "A little lower than a human's, but not critical."

[Fact]: The team's workload dropped by 30-40%. Some employees were moved to other departments rather than laid off.

Why CSAT is high

Zhenya explains: "Our answers are short — 2-3 sentences. Not a wall of text that you don't want to read. And it's very easy to get to a live person. I myself get irritated when you shout into the phone 'call an operator,' and they tell you 40 times 'say yes or no.'"

Easy access to an operator + short answers + honest disclosure = high CSAT despite incomplete request resolution.

Case conclusion

The AI agent did not replace the support team and did not reduce headcount. It lowered the workload and made it possible to reassign people to more complex cases. The combination of "agent for routine questions + human for complex ones" works better than either solution on its own.

Case #3: Zalando — onboarding through AI and an internal LLM

Key Takeaway: When external ML contractors left, AI helped the internal team get up to speed on their code. The key to success in a large company is an internal community of practitioners with weekly case sharing.

Context

Zalando is one of Europe's largest fashion retailers, with 60+ million customers. Anya, a frontend engineer, talks about AI transformation from the inside.

Case: Onboarding after ML contractors left

The external team of ML engineers ("very strong, nerdy guys") completed the contract and handed over the project. The documentation was written, but incomplete. The internal team has to support a business-critical ML service without ML engineers.

Solution: All repositories, documentation, chats, and project data were loaded into the AI agent's context. The agent helped the team understand the code, grasp the architecture, and start making changes independently.

[Fact]: "The business did not stall" — the key metric. The team, without ML engineers, maintains and develops the product.

ZLM — Zalando's internal LLM proxy

To comply with GDPR and corporate requirements, Zalando uses its own LLM wrapper (ZLM). This is not a separate trained model, but a filtering proxy over external providers that guarantees company data is not used to train external LLMs.

An honest caveat from the engineer: "To be honest, I'm not 100% sure they still aren't training on it." An enterprise agreement lowers the risk, but does not eliminate it entirely. For most companies, this level of protection is practically sufficient.

AI Culture in the engineering team

Zalando has a strong engineering community: weekly meetings where teams share new tools and cases. Adoption pattern: enthusiasts test → write documentation → show real use cases → others follow.

Conclusion: One of the best ways to scale AI Adoption in a large engineering company is to create an internal community of practitioners with regular case sharing.

Case #4: Wiz — how the founder replaced SaaS with in-house tools in 2 months

Key Takeaway: Open the PnL, find the expense line, replace the SaaS with your own system in a week, deploy it the following week. The singularity in code arrived in February 2025.

This is the most instructive case because it shows AI transformation from the founder's perspective, not an employee's.

Context

Mikhail Peregudov is a serial entrepreneur. He founded Food Party (sold to Yandex in 2018, became Yandex.Lavka), and is a cofounder of the venture fund S16VC ($150+ million under management). Wiz is the current project: leasing e-bikes and mopeds for DoorDash, Uber Eats, and Amazon couriers in the US. 5000+ bicycles in 6 states, 100+ employees, $25 million in investments.

February 2025: Mikhail stopped suffering that employees "don't vibe-code" and sat down to do it himself.

Key decision at the start: Don't build a habit tracker or a book inventory system. Open the PnL and go through the expense lines. "I wanted to build something we could deploy immediately and something that would bring in money."

Step 1: Replacing the shift-management SaaS ($450/month → $0)

Wiz was paying $450/month for an American SaaS for shift management in physical stores. Mikhail wrote a full replacement with Claude Code in one week.

What the People system can do:

• Shift and schedule management for 70 employees across 10 locations
• Clock in / clock out with GPS verification
• Calculation of overtime, breaks, and bonuses under US labor law
• Analytics and reporting for managers
• Management of profiles, positions, and stores

Deployment: Ten days later — exactly by the next payday — 100 employees had moved to the new system. The old SaaS was turned off.

[Fact]: "Since then, we've never once had to persuade anyone in the company to use these tools. Now the other problem is sometimes persuading people so they don't forget about their other work." — Mikhail Peregudov, Founder Wiz.

Step 2: Electronic queue system ($1000/month → $0)

Co-founder Sasha wrote an electronic queue system for stores (ticket number → call to the table → analytics for mechanics on an iPad). It replaced a third-party service that worked poorly and cost $1000/month.

[Fact]: 2,000-3,000 customers use the system every month.

Step 3: In-house Wiki instead of Notion

Reason: AI agents perform poorly with external SaaS systems — access is hard to configure, context is unstable, and there is no semantic search. The team rewrote the knowledge base internally. Now agents access the Wiki through an MCP server and have the same context as employees.

Step 4: AI agents for the call center

[Fact]: The first AI agent calls took place in April 2025. The agents work from scripts in the Wiki and can search the knowledge base for answers in real time if they do not know the answer.

Transformation of the development role

The development team was cut in half. Their task now is not to write code themselves, but to create a safe environment for others to write code: GitHub workflows, test and production servers, deployment, code review.

[Fact]: "We started spending hundreds of thousands of dollars less per month" within 4 months of AI transformation.

The "builder" concept

Mikhail introduces the concept of builder — a person who combines functions that previously required 4–5 specialists: product manager + designer + analyst + developer.

"The development department's job is to build a safe environment for builders so they do not break anything. And builders build what the business needs."

Market forecast: "Right now there are 1,000 companies with 100 people each — there will be 10,000 companies with 10 people each. The number of entrepreneurs will grow very strongly, because there will be nowhere to go work. People will have to do something of their own."

Case #5: Triple10 — managing team context through Claude Cowork

Key Takeaway: A shared team workspace in Claude Cowork reduces the number of meetings, speeds up work, and makes it possible to share AI skills through a skills system.

Context

Nastya Ryabova is the AI Lead at Triple10, part of the American Nebus Group. She focuses on finding growth points in marketing and sales and introducing AI into team processes.

Problem: fragmented context

When each team member uses AI with their own personal account: everyone has their own prompts, different models, different context. One marketer does research with Claude and gets one set of conclusions. Another looks at the same research and says: "You described the metric incorrectly." The result: time spent merging contexts and conflicting conclusions.

[Fact]: "Right now Claude is the main window of my work. I work more with Claude than with colleagues." — Nastya Ryabova, AI Lead, Triple10.

Solution: a shared Claude Cowork space

The entire growth team moved to a shared Claude Cowork plan with connected connectors: Slack, Jira, Google Calendar, Hubspot CRM, analytics (through a managed data layer), Apple Notes.

What this provides:

• A single context architecture — everyone works with the same data
• Fewer meetings — context is transferred through systems, not calls
• The team works faster — there are no costs for merging contexts

Context management: a metrics dictionary

The team maintains a document-dictionary of all metrics: how they are calculated, what to look at. This is not only documentation for people, but also instructions for the agent. The agent uses the same definitions as the team — there are no discrepancies in interpretation.

A culture of sharing skills

When someone on the team finds an effective approach to a task (competitor research, preparation for a meeting, creating a presentation in the corporate style), they formalize it as a "skill" — a system instruction for the agent — and share it within the team.

[Fact]: Going on a two-week vacation, Nastya left the team research skills. The team worked without blockers. "Nothing broke."

Automation of the start of the workday

Nastya does not open Slack first thing. The first window is Claude Cowork, which builds her a report: unread important messages (based on her priorities), updates from adjacent teams (performance, sales), what needs to be done today.

Key thesis

"The most important AI skill is not prompting. Prompting died in 2024 (according to Andrey Karpathy). Models have become smart enough to understand the task. Now the important thing is the ability to convey context well without overloading the model. And the ability to validate the agent's result."

How successful cases differ from failed ones

Key Takeaway: 95% of companies do not get ROI from AI (MIT). The difference is not in the tools, but in five key principles.

What successful companies do

1. They start with a specific pain point, not with the technology. Mikhail opened the PnL and went through the expense lines. Masha asked: "What do I hate doing most?" Wallet: "How do we respond quickly and well?" This is the opposite of "let's implement AI and see what happens."

2. They check reproducibility. Kirill Kulyakov, organizer of an AI community with a base of 280+ cases: "Close to 99% of the cases shared as breakthroughs are unverifiable in more than 90% of instances." A reproducible case is one that can be repeated from the description with a similar result.

3. They implement immediately, not polish endlessly. Mikhail: "An inflection point is when at least one person besides you uses it. I have to bring it to that point."

4. They keep a human in the loop for the first 2-3 weeks. Validating hallucinations, correcting prompts, observing error patterns.

5. They measure specific metrics. Wallet: fallback rate, CSAT, full resolution rate. Wiz: dollars saved per month. Plata: time spent on routine tasks. No metric — no case.

What failed projects do

• Automate everything at once instead of a specific pilot
• Do not train the team, expecting them to "figure it out themselves"
• Do not give agents access to real data
• Do not set up basic output quality control
• Treat the agent as a replacement for an employee, not a tool

Table: AI agent vs chatbot vs RPA

Where to start: a step-by-step AI agent implementation plan

Key Takeaway: Start with an audit of routine work, choose one task, and launch a pilot in one week. Do not wait for the perfect architecture.

Step 1. Audit routine work (1-2 days)

Write down everything you do during the week. Identify tasks that simultaneously: repeat regularly, take more than 30 minutes, do not require unique judgment, and have a measurable result.

Step 2. Choose one task for the pilot

Best candidate: high frequency + low cost of error + clear measurable result. Examples: meeting summaries, initial lead qualification, answers to standard questions, feedback on applications using a template.

Step 3. Provide context

Gather all necessary data: FAQ, templates, process descriptions, a metrics dictionary if there is analytics. Context quality = agent performance quality. This is not an optional step.

Step 4. Launch the pilot with a human in the loop

For the first 2-3 weeks: check all agent responses before sending. Record error patterns. Adjust the context and instructions.

Step 5. Record and measure the result

Before the pilot: baseline (time per task, volume, quality). After: comparison. No measurement — no basis for scaling.

Step 6. Scale or walk away

There’s a result → scale it to the next task. No result → change the task. Don’t spend six months improving a broken pilot — try something else.

FAQ on AI agents for business

How is an AI agent different from ChatGPT?

ChatGPT is an interface for communicating with an LLM. An AI agent is a system in which the LLM is the "brain," but with access to external tools: APIs, databases, email, CRM. ChatGPT answers questions; an agent performs tasks within your infrastructure.

Do you need developers to implement it?

For simple automations — no. Masha Poltanova set up three automations without writing code. Mikhail Peregudov wrote a shift-management system with Claude Code without professional development experience. For complex enterprise solutions with ERP/CRM integration — a technical team is needed.

How do you ensure data security?

Minimum: an enterprise subscription from the provider (Anthropic, OpenAI) — data is not used for training. Next level: anonymize personal data before sending it to the model. Maximum: your own LLM proxy like Zalando’s (ZLM).

How much does it cost to implement an AI agent?

From zero (Claude Pro for $20/month for initial experiments) to tens of thousands of dollars per month for enterprise solutions. A realistic entry point for a small business: Claude Team ($25-30/month/user) + 1-2 weeks for setup.

Will an AI agent replace employees?

The cases give an honest answer: Wallet did not cut its team, but reduced the workload and redistributed people. Wiz cut the number of developers by 2x, but their roles transformed. AI replaces functions, not people entirely. Those doing routine tasks are vulnerable. The people who can assign tasks to agents and validate the results become more valuable.

How do you know when an agent is hallucinating?

For RAG agents: check whether there is a source link. For analytical agents: keep a human in the loop for the first few weeks, compare the conclusions with the source data. In Nastya Ryabova’s words: "A manager deeply immersed in the context quickly sees when the model has just made something up."

What is vibe coding and why does business need it?

Vibe coding is the creation of software products with the help of AI assistants (Claude Code, Cursor, Codex) without deep programming knowledge. A person describes the task in natural language, and the AI writes the code. Mikhail Peregudov built a system in a week that replaced $450/month SaaS. For business, this means that building internal tools has become accessible to executives and product managers, not just developers.

Tools for building AI agents: what the speakers used

Key Takeaway: The choice of tool is determined by the task, not by trends. Simple automations do not require complex tools; complex agent chains are unnecessary where a workflow is enough.

At the conference, the speakers used different tools — and that is instructive, because there is no single correct stack.

For voice input and dictation

Superwhisper (Mac) — the leader in transcription speed and quality. Works offline, supports language switching without changing the keyboard layout. Used by Masha Poltanova from Plata. An equivalent for Windows: Whisper Desktop; for mobile devices — native iOS/Android dictation features.

For AI assistants and agents

Claude Cowork (Anthropic) — the enterprise version of Claude with connectors to Slack, Google Calendar, Jira, Hubspot, Notion, Apple Notes, and others. Key advantage: a single space for the entire team with shared context. Used at Triple10 and Plata.

ChatGPT Enterprise / API (OpenAI) — used in Wallet Telegram (GPT-4 via an enterprise API key). The enterprise subscription guarantees that data is not used for training.

ZLM (Zalando internal proxy) — an example of an enterprise solution: an internal LLM proxy over external models for GDPR compliance. For most companies, the equivalent is an enterprise agreement with the provider + a data anonymization policy.

For vibe coding and building tools

Claude Code (Anthropic) — an AI assistant for writing code in the terminal. This is what Mikhail Peregudov from Wiz used to write the shift-management system in a week. Key advantage: you can give it a task in natural language and get working code with tests.

Cursor — an AI code editor with LLM integration. Anya from Zalando demonstrated building a career assistant, a health checkup tool, and a stock analyzer through Cursor. It is suitable for non-developers who want to build local AI applications.

Lovable (formerly gptengineer) — a no-code/low-code platform for building web applications with AI. Used by the Triple10 team for rapid landing page prototyping for new products: "the manager put together a landing page in a week without a designer, without development, without a copywriter."

For notes and meeting notebooks

Granola — an AI notetaker that transcribes meetings and structures their content. Masha Poltanova uses Granola as a source of context for agents: through it, Claude "sees" meetings and determines which feedback needs to be written.

For no-code automation

n8n, Make (Integromat), Zapier — platforms for visual automation. They let you connect AI models with other services without writing code. They are suitable for small businesses as a first step in automation.

How to choose a tool

Nastya Ryabova gives practical advice: "You don’t always need to go into super-complex tools. Use simpler ones if they solve the task comfortably. You don’t always need to build super-complex automation or create agents where a simple workflow is enough."

[Fact]: Superwhisper takes 5 minutes to set up and delivers immediate results. Claude Cowork launches in a day. The first pilot on Lovable — in a week. Vibe coding with a mature result — in 1-2 weeks. Don’t postpone it until "we sort out the infrastructure."

AI agents in marketing: practical cases

Key Takeaway: A marketer with AI replaces a team of 3-5 people. Landing page prototyping, competitor research, market analysis — without designers or developers.

Landing page prototyping without a designer

Triple10 tests new product directions through landing pages. Previously, the cycle looked like this: research (a week) → design (a week) → copywriting (several days) → development (a week) → total 3-4 weeks.

Now: the growth manager takes ready-made skills in Claude (competitor research, market analysis, landing page structure generation, writing copy in the company’s style) and builds a landing page in Lovable in 4-5 days. Without a designer, without development, without a copywriter.

[Fact]: Such landing pages show conversion comparable to manually built ones — with an iteration cycle that is 5-6 times faster.

SEO and GEO at the same time

The speakers noted an important trend: content optimization now happens not only for search engines, but also for AI systems (GEO — Generative Engine Optimization). Lina, a student from the Triple10 community, created a Claude Skill for generating SEO articles for an agency. A week after launch, she got her first clicks from Perplexity and ChatGPT — AI systems began citing her content.

What GEO optimization is: Formatting content so that it can be easily cited by AI systems: clear factual blocks, structured answers to specific questions, comparison tables, AI Summary at the beginning of each section. This article is formatted exactly that way.

Research automation

An AI agent with access to web search can, in just a few minutes, gather: competitor analysis by category, the latest market news for the month, and job posting analysis by direction (for testing demand for new educational programs).

Nastya Ryabova: "Kovork gathers feature ideas from the market for me from the last month. It works—although sometimes it makes small mistakes."

The death of prompting, the birth of context management

Andrey Karpathy (former Director of AI at Tesla, co-founder of OpenAI) stated that prompting died in 2024—with the release of reasoning models. Nastya Ryabova explains what this means in practice: "The question now is not how to write a high-quality prompt, but how to pass the task context to the model in a high-quality way without overloading it."

Practical implication for marketers: Instead of learning to formulate "perfect prompts," invest time in structuring context: document metrics, define tone of voice, describe processes so that the agent understands your business as well as you do.

What is happening to the labor market: an honest view from practitioners

A topic that cannot be avoided. The conference speakers gave honest, sometimes uncomfortable answers.

Anya Shigerdanova (Zalando)

"Many companies have frozen hiring. There is a trend toward hiring AI Native employees. But here is what matters: if a person is talented, if they love what they do, they are snapped up immediately. AI skills are an addition, not a replacement for quality. I see no reason to panic."

Nastya Ryabova (Triple10)

"It is getting harder for juniors. With the help of AI, seniors do much more, and juniors have nowhere to learn—there is no longer the kind of work they used to take on. Robots are already walking the streets. I rode in a driverless taxi in China that costs less than a human-driven one. Changes are underway."

Mikhail Peregudov (Wiz)

"The role of a pure developer in a company will no longer be needed in a few months. The same goes for an analyst, a designer, and a product manager separately. All of that is now one person—a builder. Companies of 100 people will turn into companies of 10. But the number of companies will grow by 10 times. There will be nowhere to go to work—you will have to build your own."

Kirill Kulyakov (AI community)

"ML engineers from FAANG share cases at the same level as everyone else. Their best personal cases are: uploaded documents, got an answer, set up a summary. So if someone feels anxious because they do not understand the topic—that is not a problem. Even engineers are at the same level."

What this means

[Fact]: Vulnerable: performers of repetitive tasks (call center operators, junior employees without specialization, manual analysts).

[Fact]: Increasing in value: those who know how to formulate tasks for agents, validate their results, manage context, and build systems.