What AI Agents Actually Cost You

What You’ll Gain

• The two pricing models for AI agents, in plain numbers. Pay for effort or pay for completion, and what each one means for your bill.
• A third option that costs one subscription. What a working AI sales employee looks like when you build it yourself, and what it is made of.
• Why 6% of companies get transformational results from AI and the other 84% who use it do not.
• How advertising is changing now that the ad is an agent you talk to, not a page you click to.
• Where the week’s AI money went, including a $1.3B round for software that watches trucks and drivers.

Biggest Takeaway to Implement: Pick one boring, repeated job in your business and write down how it gets done, step by step, in a file. That file is the start of an AI employee you own. Chris’s version: AI is not something you ask questions to. Its primary purpose is to do things.

0:00 The Anthropic Resignation 172 Million People Read

Chris opens on the story that set the tone for the week. On September 9, Jacob Coxon, a pretraining researcher at Anthropic who had worked at OpenAI before that, resigned publicly and said neither company was acting responsibly. This time an insider agreed. Anthropic’s alignment science lead, Evan Hubinger, replied that he personally puts the probability that AI could kill all humans above 10% within the next decade. The post passed 172 million views. Three days later Anthropic’s CEO published an essay asking the industry to pace itself. Olga then opens the show and tells a friend’s question from the weekend: why she does it. Her answer is that without the Thursday prep she would have the biggest fear of missing out of her life, because nobody with a job and a family can keep up with this on their own.

3:30 Why Every Big Lab Suddenly Wants to Slow Down

Dario Amodei’s essay, We Must Pace the Frontier, asks for pacing, not a pause. Its first step is a board of outside evaluators with employee-like access to the models. Sam Altman agreed and said OpenAI would do the same. Elon Musk wrote “Dario is right.” Mark Zuckerberg broke ranks and said each lab is responsible for pacing itself, and President Trump called the whole thing a hoax. Chris is blunt: he reads it as a regulatory play, rules only the largest labs can meet, and a chance to get cash flow in order after three years of an expensive race. Olga pushes back. Drug companies answer to the FDA, and AI answers to nobody. Chris’s reply is that if he were China he would be cheering for the slowdown. Olga’s conclusion: nothing is actually slowing down, so keep getting more AI fluent.

8:30 What Changes When Your CRM Lives Inside Claude

Anthropic shipped Salesforce inside Claude: 37 prebuilt sales skills, in open beta on all paid Claude plans. The named ones are account research, call prep, pipeline review and CRM updates. Your admin requests access through AgentExchange, you log in with your own Salesforce credentials, and the data comes in under permissions you already have. Anthropic says writes need your approval by default. The shift is where you work. You are in Claude and Salesforce feeds it. The same week, Salesforce’s CEO called fears of a “SaaS apocalypse” nonsense and said the future has no user interface.

11:07 The Subscription Fight Enterprise Doesn’t Care About

Chris explains the tension underneath. Consumers and small businesses have subscription fatigue. Enterprise mostly does not, and will take on another subscription if it reduces human error and liability. So the market is being pulled two ways at once. Salesforce is betting that enterprise subscriptions are a new frontier, not a dying one. Chris adds that any company that has been a tech giant for twenty years depends on recurring subscription revenue for its forecasts, and will not give it up easily.

12:47 The Claude Update That Makes Canva Optional

Claude merged Cowork into the main chat: one product, one prompt box, no toggle. It decides on its own whether a request is a quick answer or a multi-step job that runs in the background. Anthropic also launched Claude Docs and Claude Slides in beta on paid plans, and the deep dive later in the episode was presented from a deck built that way. Olga cancelled Canva because Claude now handles the design she needs. Chris admits the last Google Doc he opened was a contract and a spreadsheet of disc golf scores. The point for everyone else: have your AI write the outline, and it builds the deck.

16:02 Amazon’s Ads Just Moved Into ChatGPT

Two steps in one week. On September 10, Amazon started letting its advertisers run campaigns inside ChatGPT, as a pilot with select US advertisers and Delta Vacations first. That comes after Amazon spent months blocking outside AI bots from its store. On September 16, OpenAI launched Sponsored Agents: the conversation hands off to the brand’s own agent, which has the whole product catalog and can take your lead or finish the sale without you leaving the chat. Separately, and reported by Sensor Tower rather than OpenAI, ads shown per user per hour on ChatGPT’s US mobile app rose 163% in August against April. Chris calls sponsored agents enormous: a new lead generation channel that runs as a live conversation.

19:42 Why Your Website Needs a Second Front Door

If agents are shopping for people, your product information has to answer a question asked in a sentence, not match a keyword. Olga’s version: one site for humans and one for agents. Chris goes a level deeper. When an agent arrives at a site, it looks for something to connect to, an API or an MCP connection. What it gets back on connecting is a welcome handshake, and today that is usually generic information about the site. His prediction is that this handshake becomes the new first contact with a customer, which makes it advertising. PageMotor already uses it to tell agents what it can do.

21:58 OpenAI and Anthropic Just Picked Different Roads

Olga points out that OpenAI is putting real resources into ads, while Anthropic has said no to advertising and keeps simplifying for builders and enterprise. Chris agrees they are no longer two cars in the same race. He sees Anthropic as the toolkit for professional AI users, and OpenAI as focused on a question Anthropic seems not to care about: how businesses stay visible in internet commerce, now that websites, organic search and landing pages matter far less.

24:10 The AI Training Number That Doesn’t Add Up

The iCIMS September workforce report: 47% of US job seekers picked up AI skills in the last six months, up from 41% a year earlier. Self-teaching rose from 22% to 30%. Employer-provided training stayed flat at about one in six. Chris does not trust the small jump after a year of nonstop AI news. His bigger point is incentives: employers can see why AI-fluent staff would help them, but most employees cannot see what is in it for them. Olga adds the missing piece, which is a picture of the finished product. People know what a fluent speaker or a good driver looks like, and nobody has shown them what AI-native looks like. Chris’s prediction: employers will soon treat AI upskilling as an emergency.

28:41 Siri Can Read Your Screen. Would You Let It?

Apple shipped the assistant it promised three years ago. In iOS 27, Siri reads what is on your screen, pulls context from Messages, Mail and Photos, and acts inside apps. It has its own app, and chats sync across devices. It launched as an English-only beta with a waitlist, and more languages arrive in October. Apple’s models were trained in collaboration with Google and distilled from Gemini, but Gemini does not run on your phone: Apple’s own models run on the device and on Apple’s private cloud. When the Simplified AI newsletter asked readers whether they would let a screen-reading Siri drive their Mac, 100% said keep it locked out. Chris uses it to describe where this is going. You will have one brain, your instructions and procedures in a single place, and every AI you use on every device will point at it.

32:32 How One Dog Owner Found a Surgeon in 30 Minutes

Mopsy, a ten-year-old dog, needed urgent dental surgery that was already damaging her eye. Emergency hospitals do not do dental surgery, and the specialists were booked. Her owner made a dozen calls, got nowhere and gave up. Then she asked ChatGPT. It found every practice within an hour’s drive, wrote an individual urgent email to each one from her own Gmail, contacted more than 20, and built a tracker of who had been asked. Half an hour later a vet replied: bring her in tomorrow. Olga’s point is the shape of it, many outreaches at once from a non-technical person. Chris’s point is the vet side: for local businesses, how fast you respond is about to decide who gets the work.

36:07 Two Conferences, One Week, Opposite Bets

The deep dive. Salesforce held its conference and HubSpot held UNBOUND in the same week, and most businesses use one or the other. Olga’s reason to watch: if you are building a product, this is where you see the market being reshaped and which openings are closing, without spending months building something nobody wants. Both companies agreed that the screen is the problem. Chris’s Photoshop comparison: a complicated interface used to need an expert, and now anybody can describe what they want and the AI plays the expert. Salesforce brought Anthropic and Nvidia on stage and planted itself firmly in enterprise. It launched a way to reach Salesforce data and workflows from any AI interface, and its own CRM model. HubSpot stayed with smaller businesses.

41:39 90% Use AI. Why Only 6% See Results

HubSpot surveyed 6,000 companies. 90% are using AI. 6% call the results transformational, and that 6% is four times more likely to hit revenue targets. The detail Chris stops on: the 6% run four or five use cases out of fifty tracked. The others run more. The winners pick specific workflows and go hard on them, and the rest spread AI thin across the whole organization. Olga ties it back to the missing vision: what an AI-native seller or company actually looks like. HubSpot co-founder Dharmesh Shah spoke about growing in the age of the builder. Twenty years ago he was the builder and his co-founder was the user, and now anyone can build. Olga’s own lesson from doing both: you can build fast, and you still have to be able to sell.

45:45 Why Salesforce Gave Its Agents Human Names

Salesforce launched seven agents with human names and job titles. Piper handles inbound leads and Hunter goes prospecting. The naming is deliberate: they want you to think of Hunter as part of your team, someone you hire. Chris thinks it is smart and exactly what large companies want. HubSpot went the other way and named its agents by job function: prospecting, customer, content, campaign and data. You tell the Breeze assistant the outcome you want and it dispatches the right agent. Olga’s summary: Salesforce sells you employees, and HubSpot sells you the outcome and hides the robots. If you sell anything with AI in it, both are playbooks worth studying.

48:21 Three Ways to Pay for the Same AI Work

Olga lays out the three lanes. Salesforce bills effort: Core at $195 per user per month billed annually, then 10 cents per agent action, with 500,000 included credits per company per year that do not grow with headcount. Your team may also need its own AI subscriptions. HubSpot bills completion: $1 per lead the agent recommends, 50 cents per customer chat that resolves, nothing when a chat escalates. The third lane is building your own: one AI subscription, a CRM (hers is HubSpot’s free tier), no per-action charges, no credits, and skills she wrote herself. She raises the risk in the rented lane. A salesperson can arrive polished and fully prepared by an agent and still not be able to sell. Both paid models are early experiments, and nobody yet knows what usage will look like.

52:06 Renting an AI Employee vs Owning One

Chris names what the rented lane depends on: the scary unknown. It looks technical, so it seems easier to pay for a platform whose price keeps rising, and you never build an asset of your own. You also cannot train a rented agent in the vocabulary of your business. The reality, he says, is that an AI employee is some markdown files and a little data, and software will soon set it up for you. If you run your own, you build it in the image of your business, down to the follow-up sequence that sets you apart from the competition, and refine it over time. His comparison is web design: people see a design they like and want one made for them, not a copy. AI employees will go the same way.

56:05 What a Real AI Sales Employee Looks Like

Olga shows hers live. It is Midas, a Claude project that runs sales for Revenue Hire, her recruiting company. It covers lead generation, meeting prep, call debriefs, follow-up and the pipeline. It is connected to her meeting recorder, Gmail, calendar, CRM and Slack, and scheduled tasks run her lead generation and daily to-dos. Before a prospect call the day before, the brief was already researched from her calendar, with the likely pains laid out, and it took her two minutes to read. She no longer updates her CRM. Deals move as emails come in, and her CRM is more current than it has ever been. Chris asks whether her other agents talk to Midas. Not inside Claude: each project is a separate employee, and projects cannot yet share a brain. Because everything lives in her own files, she could switch CRMs tomorrow and it would keep working.

1:02:36 The Simple Files Behind an AI Employee

Chris walks the diagram behind the demo. Each responsibility, such as prospecting or meeting prep, is a group of related procedures. Each procedure is one markdown file with step-by-step instructions. All of them together are the job. Companies have always had a vague idea of what each role does without writing every process down, and AI employees need exactly that written down. Humans do not work well from a thick manual, and AI needs one. Olga closes with the question of the day: which lane are you in? Buy an agent your team never retrains, which is the right answer for some companies, or build your own and control it. Chris adds that either lane, done well, should beat the equivalent human cost, and suggests thinking of AI employees as department heads that track everything and bring you in only when a decision needs you.

1:09:19 Where $6.1 Billion Went This Week

AI raised $6.11B across 103 companies from September 10 to 16, 48.5% of all venture dollars. The United States took 67% of AI dollars across 57 companies, the highest US company count in the 42 weeks tracked. Europe funded 19 companies, led by hardware. The top rounds: Motive, $1.3B, San Francisco, dash cams and safety software for trucking and delivery fleets, where AI watches the road and the driver. Positron, $500M, Reno, chips that run AI models after they are trained, which is the bill that arrives every day. Anew Labs, $290M, Shanghai, ByteDance’s drug discovery unit, spun out and taking outside money for the first time (reported by Reuters, not confirmed by the company). Exein, $270M, Rome, security built into the device itself, Europe’s biggest AI round of the week. Ridgeline, $250M, Incline Village, back-office software for investment firms, funded by founder Dave Duffield, who also founded PeopleSoft and Workday. Olga on the trucking round: twenty years ago a $1.3B round for fleet dash cams would have meant an industry being overhauled, and now it barely registers. The full Week 42 report is here.

1:14:58 What Happens When Your Business Becomes Software

Olga comes back to Chris’s question from the demo and names the gap: people are building an AI workforce out of a prompt window, and the missing piece is a brain on a server you own that coordinates every employee. Chris describes where it lands. A restaurant has a bartender, servers and a kitchen that each know their jobs. When business becomes software, that registry of jobs and processes becomes files and a little data, and an AI conducts it. It picks up signals like a new sale, starts the right process and tells you when something stalls. Olga’s close: pick something boring and start there, and stop treating AI as a chatbot. Chris’s close: have AI do one thing, and your whole mental model changes.

Resources And Sources

• Two Prices For The Same Work. The Salesforce and HubSpot pricing side by side.
• The episode slides.
• Anthropic on Salesforce in Claude. The 37 sales skills, open beta.
• Anthropic on merging Cowork into Claude.
• OpenAI on Sponsored Agents. Launched September 16.
• Dario Amodei’s essays, including We Must Pace the Frontier.
• iCIMS September 2026 Workforce Report. The 47%, the 22% to 30%, and the flat employer training.
• Apple Newsroom, September 14. The iOS 27 Siri launch.
• AI Funding Report, Week 42.
• Weekly AI Funding Tracker. 42 weeks.
• Episode 56. The theory behind the AI employee.
• PageMotor. The AI-native CMS.
• The email list. PageMotor and Practical AI updates.