# Replacing Human Departments With AI Swarms *(wait, what are they exactly?)* > A talk by Aashish Pahwa. July 2026. A talk about replacing human departments with AI swarms. It opens with how far AI has come (AI actors selling at over $15M ARR, music and text passing the Turing test, agents that use tools like humans) and the macro shift: the agentic internet, bot to bot marketing and a bot-to-bot AIDA funnel (be understood, be cited through GEO, be validated by humans, be transactable, with Zomato's MCP as the example), and the autonomy spectrum from program to network. Then the individual level: OpenAI's five-level AGI scale, what an agent is (memory, tools, goals, actions, environment), the step from automation to workflows to agents, Aashish's own AI PAs running LinkedIn, email, brand deals, content, micro SaaS and social media, an email agent that makes about $2k to $3k a month at a cost of $50, a personal VA system like the one Fyxer sells for $270 a year, and the options for personal agents (connectors, Openclaw and Hermes, Open Human or Me.bot) with the rule that you cannot trust a memory you cannot read. Single agents are bad at multitasking, so you build a team: an AI swarm of specialists that coordinate on one goal, shown with a research, newsletter and video script team. It ends on orchestration, one agent managing other agents without human intervention, and examples of back office automation, an orchestrator replacing a team, and inbound lead management over calls. > Slides and speaker notes; the live version with the AI Q&A is at https://aashishpahwa.com/talks/ai-swarms ## Slide 1 of 45: Replacing Human Departments With AI Swarms - (wait, what are they exactly?) Aashish says: AI swarms. It sounds like a sci-fi film. It is really just a team of agents, each with one job. By the end you should be able to draw one on a napkin. ## Slide 2 of 45: Which of these is real & Which is AI? - Which of these is real - & - Which is AI? Aashish says: Two videos, same bedroom, same dance. One is a real person. One is generated. Pick one. Most rooms split close to half and half, which is the point. ## Slide 3 of 45: These businesses took this opportunity in 2026 - These businesses took this opportunity in 2026 - (They're selling AI Actors with over $15M ARR) - Create winning ads with AI - AI Actors - Create winning studio quality video ads with AI Creators Aashish says: And people are already selling that. AI actors for ads, over $15M ARR. Nobody on these sites is a person. The brands buying them do not seem to mind. ## Slide 4 of 45: Music passed the Turing Test in October 2025 - Music passed the Turing Test in October 2025 - Let's test it out - Ethan Mollick: It looks like AI music is following the same path as AI text: people are only 50/50 in identifying older Suno vs. human songs (but 60/40 when two songs are the same genre) - It takes much less time to create a song in Suno v5 than to listen to a song. Aashish says: Music went the same way. People could only tell Suno from a human about half the time. Scan the code and try it yourself. I got it wrong more than once. ## Slide 5 of 45: Textual-based Turing Test was passed in 2024 - Textual-based Turing Test was passed in 2024 - It's 2026 now and these agents can use tools like humans now - GPT-4 judged human 54% of the time - GPT-4.5 with persona judged human 73% of the time - LLaMa-3.1-405B with persona judged human 56% overall Aashish says: Text went first. GPT-4.5 with a persona was picked as the human 73 percent of the time. More often than the real humans. That was talking. What changed since is that they can now use tools the way we do. ## Slide 6 of 45: So, why are we not talking about the future? - The web, as we know it, is dying. - Agentic internet has already taken over. Even Google and Meta are pushing for it. - SAAS is dying. (SAAS as a service is the next big thing) - LLM Models > humans at non novel tasks. - Agents are already buying - Indian consumers embrace AI agent-led shopping, says Accenture Aashish says: So why do we still talk about AI like it is next year's problem. The web is changing, SaaS is changing, and agents are already buying things. Even here, Accenture says Indian shoppers are ready for it. ## Slide 7 of 45: About Aashish - An AI consultant, generalist and coach. Taught AI to over 100k people. - Founder of Koso.ai - an AI custom development company - Founded, bought, sold, and currently working on 20+ digital assets - An AI creator (100k+) and a community builder (50k+) - An AI enthusiast - A part-time Domainer - OKAashish, Aashishpahwa Aashish says: Quickly, me. I run Koso, we build custom AI systems. I teach AI and I make videos about it. Most of what follows I run on my own work first, so the examples are mine. ## Slide 8 of 45: There's a new B2B now - There's a new B2B now - (Bot 2 Bot marketing) - 1 Awareness: be understood. Visibility optimization helps LLMs know what you sell and mention you. - 2 Interest: be cited. GEO helps LLMs cite you when customers ask for the best products. - 3 Desire: be validated. Human proof matters: communities, forums, and influencers. - 4 Action: be transactable. Optimize web assets so bots can take action. Coming next. Aashish says: There is a new B2B. Bot to bot. Your buyer's agent talks to your agent. The funnel is still AIDA. Every stage just has a different reader now. ## Slide 9 of 45: Make agents find you first. - Make agents find you first. - (Next year, it won't be just ChatGPT, Gemini, or Claude crawling the internet) - Is Your Site Agent-Ready? Aashish says: Awareness first. An agent has to understand what you sell before it can recommend you. And it will not be three crawlers for long. Every company will have its own agents reading your site. ## Slide 10 of 45: Focus on citations - Focus on citations - GEO is like SEO but without focus on just links. - SEO, GEO, Essential for both Aashish says: Interest means getting cited. GEO overlaps with SEO a lot. The difference is you are not chasing a link. You are chasing a mention inside the answer. ## Slide 11 of 45: Human Validation - Human Validation - ORM and social validation matter more in the AI world than before. - Reddit's AI search influence goes beyond training data Aashish says: Desire is the ironic one. The bots trust humans. Reviews, Reddit threads, creators. Your reputation online matters more now, not less, because the model reads all of it. ## Slide 12 of 45: Zomato built the bot-to-bot AIDA - Zomato built the bot-to-bot AIDA - Known → cited → trusted → transacted - Awareness: city, cuisine, dish, menu, and 'near me' pages make Zomato legible to search-powered LLMs - Interest: best-of collections, ratings, photos, and reviews create answer-ready evidence - Desire: reviews, memes, creators, and forums add human proof at internet scale - Action: official MCP: discover → menu → cart → order → track → QR pay - The full agentic commerce loop. Aashish says: Zomato has done all four. Pages LLMs can read, reviews they can quote, memes people share. And the last step is an official MCP. An agent can find a restaurant, fill a cart and pay without opening the app. ## Slide 13 of 45: So yes, agentic commerce is (almost) here. - So yes, agentic commerce is (almost) here. Aashish says: So yes, agentic commerce is almost here. The almost is shrinking every quarter. ## Slide 14 of 45: The Autonomy Spectrum - The Autonomy Spectrum - From fixed rules to networked agency - 1 Program: human sets the rules. AI follows. - 2 Assist: AI advises. Human decides. - 3 Assemble: AI builds the plan. Human approves. - 4 Authorize: human sets guardrails. AI executes. - 5 Autonomize: AI acts. Human is notified. - 6 Network: agents coordinate. Humans oversee. - More autonomy → less intervention → stronger guardrails Aashish says: Here is the ladder. Most companies are at step two. AI advises, a human decides. Swarms live at step six. Agents coordinate with each other and humans oversee. Notice the guardrails get stronger as you climb, not weaker. ## Slide 15 of 45: But that's the macro part. - But that's the macro part. - Let's get to the individual level. Aashish says: That is the big picture. Now the part you can do something about on Monday. Your own work. ## Slide 16 of 45: OpenAI's Five-Level AGI Scale - OpenAI's Five-Level AGI Scale - 1 Current AI, like ChatGPT, that talks with humans. - 2 AI that can solve basic problems like a PhD w/o tools. - 3 AI agents capable of taking actions on a user's behalf. - 4 AI that can create new innovations. - 5 AI that can perform the work of entire organizations of people. Aashish says: OpenAI's own scale. Five levels. Look at the top one. The work of entire organizations of people. That is the same thing as this talk's title, and it is their stated goal. ## Slide 17 of 45: We are here - June, 2024: level 1 - May 2025, O4-mini: level 2 - We are here: level 3 - The scary stuff starts from here (conscious models): level 4 Aashish says: Level one in June 2024. Level two by May 2025. We are at three now, agents acting for us. Level four is where it gets strange. It took under two years to climb two steps. ## Slide 18 of 45: What is Agentic AI? - What is Agentic AI? - Memory, Tools, Goals → Agent → Actions → Environment → Observations Aashish says: So what is an agent. A model with memory, tools and a goal. It acts on the world, looks at what happened, and goes again. That loop is the whole idea. ## Slide 19 of 45: Like this financial management agent- - Like this financial management agent- - Dexter: your AI assistant for deep financial research. Aashish says: Like Dexter. You ask it a finance question and it goes off, pulls the data, checks itself and comes back with research. Same loop. Goal, tools, act, look, repeat. ## Slide 20 of 45: My AI PAs Handles - - I have a mix of Openclaw, Hermes, Claude Code, and custom agents.. - My LinkedIn accounts – posting, replying, commenting, outreach, etc. - My emails – outreach, negotiations, and even lead nurturing. - My brand collaborations – outreach, inbound, negotiations, contracts, etc. - My content department – research and scripting for videos; writing, posting, and updating content on websites, etc. - Micro SAAS – Ideating, building, and managing microSAAS projects. - Social media – Carousels, Videos (Even editing) Aashish says: This is what my agents do today. LinkedIn, email, brand deals, content, micro SaaS, even video edits. It is not one tool. It is Openclaw, Hermes, Claude Code and some I built. Each one has a job. ## Slide 21 of 45: But let's understand technical stuff first. - But let's understand technical stuff first. Aashish says: Before swarms, two minutes of plumbing. I promise it is short. ## Slide 22 of 45: Automation → Workflows → Agents - Automation → Workflows → Agents - The difference is how much judgment the system can use. - 1 Traditional automation: follows fixed rules. Best for repetitive, predictable tasks. Example: copy form data. - 2 AI workflows: adds judgment inside a process. Best for classification and flexible decisions. Example: triage support tickets. - 3 AI agents: pursues a goal across many steps. Best for open-ended, multi-step work. Example: research and prepare a brief. - Preset instructions → adaptive decisions → goal-directed action Aashish says: Three things people mix up. Automation follows rules. A workflow puts a bit of judgment in the middle. An agent gets a goal and works out the steps itself. The difference is how much judgment you hand over. ## Slide 23 of 45: A simple AI Agentic Workflow looks like this - A simple AI Agentic Workflow looks like this Aashish says: This is a simple one. A message comes in, an AI agent with a model, memory and a couple of tools handles it. Boxes and arrows. Nothing magic. ## Slide 24 of 45: A complex one can replace your teams - A complex one can replace your teams - This agent makes me ~$2k – 3k every month at the cost of $50. Aashish says: This is my inbox. Every form submission here gets read, sorted and answered by an agent. It makes me about $2k to $3k a month and costs $50 to run. That used to be somebody's job. ## Slide 25 of 45: But honestly, it's like solving a puzzle. - But honestly, it's like solving a puzzle. Aashish says: Honestly, building these is a puzzle. You take a task a person does and break it into pieces a machine can hold. The skill is not coding. It is seeing the pieces. ## Slide 26 of 45: Puzzles become Systems. - Puzzles become Systems. - And systems, when attached with appropriate tools, can replicate human working. - When I receive a reply to my newsletter. - Reply like Aashish - A long-term memory of How Aashish Talks - Here are all the Gmail commands - Make them join the WA community Aashish says: One puzzle, solved. Someone replies to my newsletter. The agent answers the way I would. It has a memory of how I talk, the Gmail tools, and one goal: get them into the WhatsApp community. ## Slide 27 of 45: Like My Personal VA System - Like My Personal VA System Aashish says: Put enough of those together and you get a personal assistant. This is mine, sorting my mail as it arrives. ## Slide 28 of 45: Fyxer is selling exactly this for $270 PA - Fyxer is selling exactly this for $270 PA - At a valuation of $60 million. - Your assistant for writing replies Aashish says: And Fyxer sells exactly this. $270 a year, at a valuation of $60 million. So the puzzle you solve for yourself might be someone else's company. ## Slide 29 of 45: Also, I know you might already be doing it - Using Claude or ChatGPT connectors (easy but prone to fail. Plus it's just time based) - Technical AI Personal Agents like Openclaw & Hermes (scary black code screen) - Non technical AI Personal Agents like Open Human or Me.bot (I prefer them) Aashish says: Some of you do a version of this already. Who uses Claude or ChatGPT connectors? Three ways in. Connectors are easy but break. Openclaw and Hermes are powerful but it is a black code screen. I prefer the non-technical ones. ## Slide 30 of 45: You can't trust a memory you can't read - You can't trust a memory you can't read Aashish says: This is Open Human. It checks my calendar and tells me honestly when it cannot reach something. The rule I use: if I cannot read what the agent remembers about me, I do not trust it. ## Slide 31 of 45: But they are single agents. - But they are single agents. - And a single agent suck at multitasking. - (Even with a long term memory) Aashish says: Here is the catch. All of these are one agent. And one agent is bad at doing many things at once. Give it five jobs and it gets confused, even with a good memory. Same as a person, really. ## Slide 32 of 45: So we make a team, and we call it an… AI Swarm - So we make a team, and we call it an… - AI Swarm - Many specialists. Shared signals. One adaptive goal. - 1 Specialize - 2 Coordinate - 3 Emerge - The group solves more than any one agent. Aashish says: So you do what companies do. You hire a team. That is a swarm. Each agent specializes, they share signals, and together they solve more than any one of them could. ## Slide 33 of 45: Let's simplify it. - Let's simplify it. - (The visual way though) - First team research for AI news from Telegram, Reddit, HackerNews etc and adds it to google sheet. - Agent verifies it and writes newsletters using the same - Agent verifies it and writes video scripts using the same Aashish says: Here is one I run. One agent reads Telegram, Reddit and Hacker News for AI news and puts it in a sheet. Two more pick it up. One checks it and writes the newsletter. The other writes video scripts. ## Slide 34 of 45: Let's simplify it. (The visual way though) - One agent researches daily AI news -> - One agent writes the newsletters - One agent writes the video scripts Aashish says: Same thing with real screens. The research digest at the top, the newsletter and the scripts underneath. That used to be a content team of three. ## Slide 35 of 45: Earlier I used AI Workflows for the Same - Earlier I used AI Workflows for the Same Aashish says: Last year I built it as a workflow. Look at it. Every box is a thing that can break. It worked, but I was the one fixing it every week. ## Slide 36 of 45: But that was 2025. - But that was 2025. - (human in the loop) - Today, you can spawn multiple agents in Claude itself. Aashish says: That was 2025. Today Claude can spin up several agents itself, with you approving along the way. No canvas, no wiring. You describe the team. ## Slide 37 of 45: But what these chatbots can't do well is Orchestrating. - But what these chatbots can't do well is - Orchestrating. Aashish says: But there is one thing chat apps still do badly. Orchestrating. They can start a team. They are not good at running one. ## Slide 38 of 45: Orchestration is when an Agent can manage other Agents - Orchestration is when an Agent can manage other Agents - Without human intervention - CEO: Claude - CMO: OpenClaw - CTO: Cursor - COO: Claude - CodexCoder, Engineer: Codex - ClaudeCoder, Engineer: Claude Aashish says: Orchestration is one agent managing the others, without a human in between. An org chart where the CEO is Claude, the CTO is Cursor, and the engineers are Codex and Claude. That is a department. ## Slide 39 of 45: A (really) Simplified Agentic Orchestration Demo - A (really) Simplified Agentic Orchestration Demo Aashish says: A really simplified one. An orchestrator in the middle writes articles by handing work to sub-agents and checking what comes back. The orchestrator decides who does what next. Not me. ## Slide 40 of 45: Now that we're talking about agents - Now that we're talking about agents - How are you using agents for internal productivity? Aashish says: Your turn. Hands up if you already use agents inside your company. Not for customers, for your own team. Here are a few we have built. ## Slide 41 of 45: Like Automating Back Office - Like Automating Back Office Aashish says: Back office first. It is the boring work, which makes it the best place to start. Counts, invoices, status. Agents are good at boring. ## Slide 42 of 45: Or an orchestrator to replace a team - Or an orchestrator to replace a team Aashish says: Or an orchestrator that runs a whole team of writing agents. You give it the brief, it gives you the finished work. ## Slide 43 of 45: Or inbound lead management using calls - Or inbound lead management using calls - Never miss a call. Ever again. Aashish says: Or voice. An AI receptionist that picks up every inbound call, answers questions and books the lead. Never miss a call again. That one is a department too. ## Slide 44 of 45: Let's Talk More - Let's Talk More - (I run koso.ai) Aashish says: If you want one of these for your own company, this is where to find me. I run Koso, and this is what we build. ## Slide 45 of 45: You can find me here - You can find me here - OKAashish - Aashishpahwa - @OKAASHISH Aashish says: And I am okaashish on most places. I post what I build, usually the week I build it. --- Invite Aashish to speak: aashish@koso.ai. All talks: https://aashishpahwa.com/talks