I recently read a thought that made me uncomfortable. The gist: companies will build new things from scratch. With a new team. 10x smaller. And the old one — will be let go.
Harsh. I looked at the numbers.
Numbers You Can No Longer Ignore
In March, OpenAI published a framework of five AI business transformation models. The first level is Workforce Empowerment. The third is Systems and Dependency Management. The fifth is Process Re-Engineering with Agents, where agents redesign entire processes. Between the first and the fifth — three years. Maybe less.
Capgemini calls 2026 the “year of truth.” Average ROI on AI projects is 1.7x. Payback period is 1–2 years. GenAI delivers 26–31% savings on operational costs. Companies see this and start figuring out who’s needed and who’s not.
HBR, meanwhile, writes about the “last mile problem.” Companies launched hundreds of AI pilots, rolled out Copilot and ChatGPT. Transformation is stalling. Not because the technology isn’t ready — because people haven’t adapted.
Two Types of Employees
Devendra Goyal introduced terms that resonated with me: AI-native vs AI-dependent.
AI-dependent uses ChatGPT as a better Google. Asked — got an answer. Copied. Pasted. Task closed.
AI-native rethinks the process entirely. Not “how to speed up what I’m doing,” but “why am I doing this manually at all.” The difference is fundamental: one optimizes steps, the other eliminates them.
IBM tripled hiring for entry-level positions. But redesigned them for AI. Not “junior developer writes code,” but “junior developer manages agents that write code.”
Elena Verna wrote in February: “The demand for AI-native employees is exploding. Right now.” The window is measured in months, not years.
My Path: From “Better Specs” to Automation
In 2023, I was writing specs and documentation with AI — threw text into ChatGPT, got an “improved” version back. Time saved? About 20 minutes. Then I built custom GPTs — slightly better, but essentially just prompt engineering. Nothing serious could come from that.
The turning point was January 2025. Found the right community, three intensives back to back, started learning automation properly. Built my first documentation pipeline — a serious one, with well-designed context, a connected knowledge base, and a thought-through structure. That’s when it clicked. The gap between “throwing text into a chat and getting an answer” and “building a process that works reliably” is enormous.
Then I discovered Claude Code. That same documentation pipeline went through an evolution: from ChatGPT prompts to a full-fledged system with a release management site, automatic context collection, and Claude Code integration. Everything changed.
Over time, I built numerous projects and products personalized specifically for me. Several of them replaced paid subscription services. Launched a content pipeline across three platforms.
I gained the ability to do things I simply never had time for before. The quality of output increased compared to what I used to produce manually. Mistakes happen. Edits after review — that’s normal and part of the process. Every correction is feedback that improves the automation/system/agents for the next iteration. The final result is my responsibility.
Use Cases by Role That Might Resonate
Product Manager:
- Meeting transcript processing: an agent extracts decisions, action items, open questions and organizes them into documents. Notes from a one-hour meeting — in minutes.
- Feature request pipeline: incoming customer request → code research → gap analysis → 2–3 solution options with trade-offs → stakeholder summary.
- Generating user stories from requirements with quality standard checks, versioning, and TBD tracking.
- Feedback synthesis from 10+ sources (CRM, support, App Store, Slack, Jira): deduplication, prioritization, a ready-made list for the roadmap.
- PRD in 45 minutes instead of 8 hours. A clickable prototype instead of a spec — in one Zoom call.
- Ad-hoc analytics without an analyst: an agent with database access answers “LTV by region,” “why did conversion drop” in 3 minutes.
- Voice-recorded Decision Records: you describe the decision, the agent asks clarifying questions and generates a DR with context, alternatives, and a review timer.
- Preparing documents for any team member: designer, architect, business, whoever.
Developer:
- A 52-page specification + ~280 feature files in Gherkin in 3 hours. An agent wrote the code from them overnight. Two evenings instead of two sprints.
- Full cycle: wishlist → specs → planning → implementation → verification. All through agents. Human — on review.
- Reverse-engineering from open source in unfamiliar languages: an agent extracts logic from Swift+TypeScript repositories and writes documentation with examples.
- Test coverage from 33% to 95% after a two-day investment in test hierarchy documentation.
- Legacy code audit before modifications: scanning, dependency index, documentation — only then new features.
- Neo4j code dependency graph: AI agent accuracy on tasks with hidden dependencies jumped from 75% to 99.4%.
Marketer:
- Finding and scoring 200+ influencers in hours instead of 60 over several days. Parsing, enrichment (geo, ER, fake followers), personalized briefs.
- Analysis of 1,700 ad creatives ($2.5M in spend) in 1 hour. Found one creative that wasted $15K.
- An agent manages ads via API: compares actuals against plan (CPA, ROAS), scales or pauses campaigns.
- AI video creative in 5 minutes and $10 — indistinguishable from real footage, good metrics.
- 65% of sales through an AI ambassador platform: registration, onboarding, scoring. CAC 2–2.5x cheaper than digital.
- A marketer without a developer vibe-coded an interactive game-landing page in a week. The special project collected leads into CRM.
Product Designer (UI/UX):
- Google Stitch: text, sketch, or screenshot → clickable prototype with React/HTML export. Minutes instead of 3–4 hours on wireframes. Idea validation before Figma.
- Figma Make: describe a screen in text — AI generates a prototype integrated into your existing design system (styles, components).
- AI UX audit: an agent opens a browser, clicks through the interface, and performs an audit against NN Group and iOS HIG standards. No manual screenshots.
- Figma → working React in 15 minutes: via MCP, select components in Figma → get a React App with preserved design system.
- Design system in 45 minutes of “taste” + a couple hours of description: AI generates mockups, the designer curates and systematizes. Instead of months.
- AI-generated UI from a brief as development references. Saves iterations between designer and developer.
QA:
- Full E2E QA workflow: user story → test plan → 60 automated tests → self-healing → commit. 1–2 hours instead of 5–7 days.
- Self-healing tests: UI changed — the test fixed itself, no human involvement.
- Creating test cases from Jira, Confluence, Figma, and Slack in a minute. Minus 90% of manual time.
- QA workload reduced by 50–70% according to teams that adopted AI testing.
DevOps / SRE:
- Overnight alert triage: an agent handles low-urgency alerts, delivers a morning summary — what to look at manually.
- MTTR down 40–75% through AI observability. Specific case: from 72 hours to 18 hours.
- Noise reduction: 50–70% of alerts are false positives. AI groups related events, delivers one incident instead of 50 notifications.
- Root cause analysis from hours to minutes: an agent correlates logs, traces, metrics, and deployment history.
Analyst:
- Data normalization from dozens of Excel/SQL tables (40MB): from 5 days to a few hours.
- Text-to-SQL: a non-technical specialist queries a database in natural language.
- 40,000 hours of user conversations analyzed for product analytics.
- Run-rate dashboards with forecasting: actuals + plan + end-of-month projection based on 7-day trend.
Engineering Manager:
- Team knowledge base in 2 hours: standards, processes, incidents in markdown. An agent queries it instead of meetings.
- AI-drafted performance review from accumulated data: 1:1s, Slack, PR comments → structured review in minutes. Instead of 60–90 minutes of painful recollection (“recurrent amnesia” — you remember the last 2 weeks, forget Q1).
- Automatic postmortem in 15 minutes instead of 60–90: incident call transcription, Slack timeline capture, draft generation. Up to 80% of the work automated.
- Sprint planning: AI analyzes velocity, holidays, WIP limits, and suggests a realistic plan. Saves 2.5 hours of manual updates every week.
- Tech debt prioritization through systematic analysis: AI determines what to fix first — “what will yield the greatest return.”
- Monitoring AI impact on the codebase: how much code is AI-generated, where AI code gets reverted more often, who on the team uses AI effectively.
- Redesigning interviews for AI-nativeness: adding a stage with an intentionally hidden problem — will the candidate stop and reframe the task.
Sales:
- AI scoring of candidates after a call across 50 parameters. Decision to let go/retrain: from 3 weeks to 1–3 days.
- Telegram bot for role-playing sales training with client avatars and tricky reactions.
- A German bank found 1M+ potential customers through millions of parallel AI agents.
HR:
- AI scoring of candidates from interview transcripts: green/red signals against a skills matrix.
- Employee cards with auto-updates from 1:1 transcripts, Slack, voice notes.
- Performance review draft from accumulated data for the period.
Customer Support:
- Automatic ticket routing: an agent checks documentation, similar tickets, identifies the team.
- AI automation of first-line support: basic processing, escalation. Human — for complex cases.
Entrepreneur / Solo Founder:
- Board of Advisors from AI sub-agents: 4 agents with different expertise provide independent perspectives.
- Extracting promises from all communications: who promised what to whom, with direct quotes. History spanning a year.
- Automatic blog from AI work logs — 4,000 visitors in the first month, ~90 posts in 1.5 months.
What I’ve Learned
73% of AI pilots don’t scale (HBR data). The reason — organizations wait for “systematic adoption” instead of letting people experiment. Companies are slow. An individual is not.
They’ll build the new thing alongside the old. With a small team. But while companies are ramping up, everyone has a head start.
The recipe is simple: one dumb, repetitive process. Automate it. Show the result. Then the next one. Each step is small, but they compound.
The window is measured in months, not years. One process. One result.