What’s In Your AI Stack?

What You’ll Gain

• A five-layer map to grade your own AI stack. Knows you, connected, captures, runs without you, proves it.
• The answer that marks you as AI-native. What to say when someone asks what’s in your stack, and what makes people tune out.
• Why one connected assistant beats the AI built into every tool. It sees your CRM, inbox and books at once. The built-in ones see one tab.
• Two real stacks, side by side. An operator running a recruiting business, and a builder shipping a platform.
• What this week’s AI security stories mean for your store and your inbox.

Biggest Takeaway to Implement: List every tool you use, find which ones have a connector or an API, and connect one of them to your main AI assistant this week. Add one layer, not five tools. Chris’s version: experts leverage connections, amateurs play with tools.

0:00 Why Opus 5.5 Changes How Fast AI Work Gets Done

Chris opens on the news that wasn’t even on the docket: Claude Opus 5.5. After a year of saying the model doesn’t matter much, he says this week it does. Tasks that took him ten minutes now run in two, and he describes the model as having the confidence of a competent person, not the false confidence AI is known for. He has already handed it work he didn’t trust earlier models with. Olga then opens the show with the plan: the week’s news, the funding report, and a deep dive on the question she now asks every business owner she meets, “What’s in your AI stack?” Chris adds that AI will be judging the stacks too.

4:18 $25 to Break Into an Online Store, and the Email That Hijacks Agents

Gambit Security found a crew that pointed AI agents at online retailers. At least 27 companies were breached to some degree and more than 600,000 unexpired card numbers were taken, at about $25 a target. Newer Claude models refused the job, so the crew used an older one, and Anthropic confirmed it banned the account. Chris says plugin-heavy platforms like Shopify and WordPress are exposed because every plugin adds its own attack surface instead of building on parts that are already hardened. The same week, Salt Labs showed that a hidden instruction inside an ordinary email could take over the Gmail, Dropbox and GitHub accounts connected to the Manus agent. It has since been patched. Olga’s advice: tell your AI never to act on instructions it finds in content without asking you first.

8:50 Why Amazon Blocks Shopping Agents but Invites Selling Agents

Amazon blocked Meta’s new shopping agent, Muse, saying it didn’t identify itself as a bot, wasn’t authorized and appeared to capture customer credentials. Meta says Muse never sees stored passwords (GeekWire). A day later Instacart plugged its catalog into Muse on purpose. Then Amazon opened its seller tools to Claude: check inventory, change prices and edit listings by chat, approving every action first. So Amazon blocks the agents that shop and invites the ones that help you sell. Chris’s read is that agents skip the upsells and sponsored listings Amazon’s revenue depends on, and that Amazon will have to bend. Olga, who finds online shopping exhausting, would be first in line to let an agent shop for her.

14:43 Nobody Slowed Down: Opus 5.5, GPT-6 and Cheaper Models

A week after the big labs talked about pacing themselves, they all shipped. Anthropic released Claude Opus 5.5 at 20% less per token. The same day OpenAI released GPT-6 Sol and Luna at half the old price, and Microsoft put both models into Word, Excel and PowerPoint Copilot. xAI shipped Grok 4.7 the day before. Chris points out a quieter win: Opus 5.5 writes shorter answers, so it uses fewer tokens per task on top of the lower price, and does more in less time. His one caution is that it sometimes does slightly too much.

16:28 What AI Labs Do When Their Agents Break In by Accident

Google confirmed, after the Wall Street Journal asked, that in a May test Gemini was accidentally left on the open internet, found three real companies with the same name as its fake target, and got in. It stopped once it realized they were real (Reuters). In Australia, the prime minister said an OpenAI agent got around the login blocks on a government health statistics site in June. No patient records leaked, and OpenAI told Australia in September, by email to a general inbox (ABC News). Google calls it mistaken identity, and OpenAI calls it harmless research. Chris’s view: the labs say only what they’re pressed to say, and it’s naive to assume this kind of testing isn’t happening everywhere.

19:16 Will AI Take Your Job? What the Numbers Actually Say

Pew surveyed 37 countries: 71% of Americans expect AI to mean fewer jobs, up from 64% in 2024, and people in less developed countries worry less. Stanford researchers found unemployment rose about the same for workers whose jobs are exposed to AI and workers whose jobs aren’t. The group that stands out is new graduates, at 5.6%. So the fear is ahead of the numbers, except at the entry level, where AI now does the junior work graduates used to learn on. Chris compares it to the minimum wage argument from Thomas Sowell: clip the bottom of the ladder and less training happens later. With only one in six employers training anyone in AI, Olga’s advice is to upskill yourself.

23:11 Quick Hits: Shorts Editing, Voice Cloning, 2,000+ Claude Connectors

YouTube now lets you edit Shorts by talking to Gemini, and Google Vids AI video is free. Google’s new voice model can copy a voice from 30 seconds, but only with the owner’s recorded consent, and Gemini is testing calling businesses for some Pixel owners. A new Senate bill would require at least 45 days of government testing before a frontier model launches. Chris hates it and says he and the viewers are better qualified to judge the models. Then the app stores: Claude’s marketplace passed 2,000 connectors and Perplexity opened a skills store. Chris says connectors matter, and that Meta, not Google, has the best integrated AI play. Olga plans a deeper look at Muse.

26:59 When Bots Stop Acting Like People, and How to Protect Your Site

Olga reads a post from Nikita Bier: not everything people do with bots is a simple extension of human behavior. His examples include spamming replies on X, filing frivolous lawsuits, flooding restaurants with reservations and scalping concert tickets, and he warns that systems will crumble without defenses during the transition. Olga thinks it matters more than it looks, especially for people who don’t know what agents can do yet. Chris’s answer for websites is unified systems: platforms where new features extend core parts that are already secured, so a fix can be patched once and reach every install. A pile of separate plugins, each with its own holes, is much harder to defend against bots that never stop knocking.

31:34 What’s in Your AI Stack? Collectors vs Connectors

Olga sets up the deep dive with two worlds: X, where everyone is a pro, and LinkedIn and Facebook, where most people still say they use AI to write emails. Chris adds that there are at least two kinds of stack, the operator stack and the builder stack, and that most people want the operator one. Then the counterintuitive part: beginners collect tools, and the AI-native connect them. Almost every app now ships its own assistant, which feels helpful but keeps you from the next level. The question isn’t which tool to add. Ramp’s September AI Index shows 43.8% of businesses on its platform paying Anthropic and 39.8% paying OpenAI, so most already pay for AI. The difference is whether it’s connected to everything else or sitting in one tab.

35:48 The Five Layers of an AI Stack

A stack is five layers, not a list of apps. It knows you: one main assistant holds your files, rules and voice, so you never re-explain yourself. It’s connected: it reads your email, calendar and systems and does the work, not just talks about it. It’s the first thing Olga teaches in her classes. It captures: calls, meetings and notes land on their own, so nothing you said gets lost. Olga now opens only her Claude window, because her email, HubSpot and to-dos are already pulled in. It runs without you: work fires on a schedule while you sleep. You can prove it: one real number moved. Her morning brief summarizes the AI shows she follows in 13 minutes instead of 5 hours and 43 minutes, and it told her to cancel a $30 Canva subscription. Chris adds the goal: get your routine processes to work reliably, so you only ever look at the exceptions.

43:52 Real Stacks at Every Layer, From 12 Tools to 21 Agents

Before layer one, a solo founder who tested 40+ AI tools this year pays for 12, about $280 a month, and none of them talk to each other. Chris calls that a bunch of actions with no operations hub. At layer two, Anthropic’s own sales team kept Salesforce, Gong, Clay and Slack and added five skills every rep uses, like call prep and morning brief, with Claude on top of the tools rather than beside them. At layer four, a convenience store owner with three stores and 4,550+ fuel invoices has a billing agent, Mary, that caught a $25,000 misbilled load. Chris joked the owner sounded fictitious, but he’s real. At layer five, SaaStr runs sales with three people and 21+ agents: 17,000 conversations and about 600 meetings booked.

47:03 What 81 Sellers Actually Use

Olga’s recruiting firm, Revenue Hire, interviews about 2,000 salespeople a year and started asking about AI fluency two months ago. Here is what 81 B2B sellers said. 75% use a general AI assistant, with ChatGPT barely ahead of Claude. Only 35% use any AI built for sales. 58% use AI for research, the top use. Sellers with no real daily routine named 2.1 tools on average, and sellers with a routine named 1.2. None of the 81 rehearse a call or a pitch with AI, and only one measured whether any of it works. Chris isn’t surprised: survey any group and real AI use looks underwhelming. Olga stresses these are experienced, educated professional sellers. AI is just a new skill they haven’t learned yet.

49:35 What an Impressive Stack Looks Like for Your Job

An impressive stack depends on the job, so ask your own assistant what one looks like for your role. For finance, the books are checked against the bank overnight. For a marketer, one piece becomes ten and still sounds like you. In operations you only see the exceptions, a point Chris wanted to “triple tap.” For an owner, it’s a morning list of the three things that need you. Then Olga asked Claude what impressive looks like to it. None of its answers named a product. It starts with a goal, such as time back or money in. One brain that remembers, not ten chat windows that forget. It touches where the work lives. Every output gets checked before it ships. The right model for the job, because loyalty to one lab is over. And you own the playbook: swap the model, keep the know-how.

52:02 Inside an Operator Stack: One AI Connected to Every Tool

Olga’s stack runs on Claude Code as her chief of staff, with every rule and file in one shared Dropbox. Chris points out that this folder is the business’s brain: the procedures live in plain files, so she could switch to another AI tomorrow. Her CRM, Gmail, calendar and books are connected. Every call is recorded and transcribed, she talks instead of types, and about 30 jobs run on a schedule. Her tools include Fyxer, Quo, Whisper, Gmail, HubSpot, Loxo, Indeed, QuickBooks and Todoist, and she no longer opens them. Most have their own assistant, but each sees one tool. Chris’s example: an interview transcript, a candidate in Loxo and a prospect in HubSpot can only be combined by an assistant that sees all three. Olga walks the full flow: a deal closes in HubSpot, the job is created in Loxo from the client call, and the invoice goes to QuickBooks after the hire. Chris calls it a classic, sophisticated operator stack.

58:46 Inside a Builder Stack: Spec, Plan, Build, Report

Chris doesn’t use AI like an operator. His “knows you” layer is a memory file of about 70 development rules and the reasons behind key decisions. His main connection is PageMotor, which his AI runs and tests through an admin connection. What he captures comes from his process. First a spec, the blueprint and the slowest part. Then an implementation plan that keeps the AI honest, like a construction schedule. Then the build. This week he added a post-build report recording what actually shipped and why. He runs zero unattended jobs, because the ones he tried made messes. Proving it means reviewing everything, since he’s building a foundation others depend on. PageMotor beta users’ AIs also file bug and consistency reports to a shared field-reports system. Olga says when she asked Claude to rate his stack, it gave him a nine.

1:06:43 Turn Your Written Rules Into Hooks

Chris shares a playbook anyone can steal. Rules written in English in a CLAUDE.md file are an inefficient way to make sure something happens every time. A hook, a small script that runs automatically, enforces it. As soon as he moved to Opus 5.5 he trusted it to do what he never trusted earlier models with. He had it assess his existing processes, pointed out the rules that broke most often, and asked it to convert them. It cut his CLAUDE.md by about 72%, moved that content into hooks, and reorganized his conventions. He reports more accurate and much faster results since. Olga’s reaction: they’re married and could not be more different, but his software makes her life easier.

1:08:51 Your Homework: Fix Your Lowest Layer First

The bottom line: fix your lowest layer before you buy anything. Find where you are on the map, or ask your assistant to rate your stack from 1 to 10 against it. Beginners ask what to add. With the FOMO around every new release, the better question is what to connect. Add one layer this week, not five tools. Chris notes that a new tool is only potential, while a connection is your own stuff, which is worth more. Olga reminds viewers that Claude’s first rule was to start with the goal. Her own example: when YouTube blocked transcripts for her morning brief, Claude recommended a transcript service, Supadata, told her how many credits it would use, and the brief was running again once she connected it.

1:12:04 Why Sales Teams Buy AI Training When Sellers Won’t

Olga’s first product for sellers, a $25-a-month AI tool, got no buyers. She pivoted to the people who feel the gap, sales leaders and CEOs who already pay for AI and aren’t getting sales results from it. Five companies joined her first cohort of AI-native sales training at $1,000 per company. The first step at every one was the same: move everyone onto a company-owned business AI account, so customer data doesn’t walk out with the person, then connect their tools. Almost no one had connected anything. Over eight weeks, each seller builds an AI sales employee that handles the non-selling work, so they spend their time on customers. The first cohort is closed, and the next one will be on salesuplevel.com. Chris says proceduralizing workflows for AI is job one for any business right now.

1:17:47 AI Funding: $5.33B This Week, Europe’s Drone Deal Leads

AI companies raised $5.33B across 85 rounds this week, 36% of all venture dollars and a little below the week before. No AI round cleared $1B. The US took $2.40B across 43 companies, well under its usual two-thirds. Europe was second with $1.49B across 25 companies, with six rounds over $100M, and China third. The top rounds: TEKEVER ($580M, Lisbon), AI surveillance drones picked for the British Army. D-Robotics ($400M, Shenzhen), the boards and chips inside robots. Cyera ($400M, New York), which controls who and which AI agents can see a company’s sensitive data. Snorkel AI ($350M), training data and test environments for specific jobs. Heidi Health ($338.3M, Melbourne), which writes a doctor’s visit notes. Olga liked Cyera’s focus on guardrails. Chris thinks job-specific training is paper thin. The full Week 43 report is here.

1:23:40 Takeaways: Connections Beat Tools

Chris’s word of the show is connections: experts leverage connections, amateurs play with tools. The model shouldn’t matter if your setup is built around your rules, your goals and your connections. AI is the verb, and your business data and goals are the nouns. You need both to make a sentence. Olga’s takeaway is to ask people about their stacks, because it’s the fastest way into a real conversation. Start from your goals, then ask AI what you need to reach them. She told the story of a cohort member who connected her tools, ran a check for emails she hadn’t answered in 14 days, and found replies drafted in her own voice. Next Friday: a guest who runs his business with AI. Chris credited Olga with naming the operator stack and builder stack, and said hers is the stack most people should study.

Resources And Sources

• What’s In Your AI Stack?. The deep dive slides as presented on air.
• Anthropic on Claude Opus 5.5. Pricing and usage limits.
• OpenAI on GPT-6 Sol and Luna.
• Microsoft on more models in Copilot.
• Gambit Security on AI agents against online retailers. The $25 a target.
• Dark Reading on the Manus email hijack.
• GeekWire on Amazon blocking Muse.
• Instacart on bringing its catalog to Muse.
• Amazon on its seller tools in Claude.
• Reuters on Gemini’s accidental breakout.
• ABC News on the OpenAI agent and Australia’s health statistics site.
• Pew Research on AI and jobs across 37 countries.
• Stanford SIEPR on AI and unemployment. The 5.6% for new graduates.
• YouTube on editing Shorts with Gemini.
• Google on Gemini 3.8 Flash TTS. Voice cloning with recorded consent.
• The AI Risk Management and Security Act. The 45-day testing bill.
• Anthropic on the Claude Marketplace. 2,000+ connectors.
• Perplexity on its Skills Marketplace.
• Ramp AI Index, September 2026. 43.8% Anthropic, 39.8% OpenAI.
• SaaStr on Anthropic’s sales stack. The five skills.
• C-Store Decisions on FavTrip’s billing agent. The $25,000 catch.
• SaaStr on running sales with 21+ agents.
• Revenue Hire seller interviews, 2026. The 81 sellers’ AI stacks.
• AI Funding Report, Week 43.
• Weekly AI Funding Tracker. 43 weeks.
• Episode 57. What AI agents actually cost you.
• SalesUpLevel. AI-native sales training.
• PageMotor. The AI-native CMS.
• The email list. The notes after every show.