Latest / Elon Musk Podcast / Anthropic launches Claude for Small Business

Transcript

Anthropic launches Claude for Small Business · Elon Musk Podcast

• 0:00
  Anthropic has placed its clawed AI directly inside QuickBooks,
• 0:04
  PayPal, and HubSpot via a new desktop toggle called Clawed for
• 0:09
  Small Business. Which targets a segment of the
• 0:11
  economy that mostly just gets ignored by enterprise software
• 0:15
  developers. I mean, similar businesses
• 0:16
  account for 44% of the USGDP. Yeah, it is a huge portion of
• 0:21
  the market. We're talking about roughly
• 0:22
  36,000,000 companies, but up until this point they have
• 0:26
  largely been sidelined by the major AI tools.
• 0:29
  Right, they usually just get an empty chat window.
• 0:31
  Exactly, and they have to figure out the prompts themselves.
• 0:34
  But what they are getting now is a set of prebuilt workflows for
• 0:38
  tasks like payroll and invoicing.
• 0:41
  Which changes the utility of the software.
• 0:43
  It does. The question this all points to
• 0:45
  is what actually happens to a company's operations when an
• 0:48
  external model is granted read and write access to its cash
• 0:51
  flow and customer data. The architecture they are using
• 0:54
  to answer that is highly specific.
• 0:57
  Claude for Small Business operates through Claude Cowork,
• 0:59
  which is a desktop application. It does not run in a web
• 1:02
  browser. That alters the nature of the
• 1:04
  tool entirely. Moving away from the browser
• 1:07
  shifts the AI from being a conversational assistant that
• 1:11
  you just, you know, talked to into a background execution
• 1:14
  engine. Because it lives on the desktop.
• 1:16
  Yeah. So it can navigate local files
• 1:18
  and integrate directly into the software stacks that these
• 1:21
  businesses already run on. The application connectors
• 1:24
  provided in this package cover the core operations for most
• 1:27
  small companies. You have QuickBooks handling the
• 1:29
  finance side, Sure. PayPal managing the settlements,
• 1:32
  HubSpot running the CRM, Canva for visual assets, DocuSign for
• 1:37
  contracts. And then Google Workspace or
• 1:39
  Microsoft 365 for general documents and communication.
• 1:43
  Right. Giving a language model access
• 1:46
  to your payroll data and your client contracts introduces
• 1:49
  severe operational risk though. How so?
• 1:52
  Well, if you run a small logistics company and a model
• 1:55
  has the ability to read and write in your financial
• 1:57
  environments, an unreviewed action could easily damage your
• 2:00
  cash flow. Oh right, or it could send
• 2:03
  incorrect contracts to your vendors.
• 2:04
  They. Handle that through a strict
• 2:06
  permission inheritance model. The system adheres exactly to
• 2:09
  your existing permissions if a specific employee setting up the
• 2:13
  workflow cannot access financial records in Google Drive.
• 2:17
  Like they do not have admin rights.
• 2:18
  Right. Or if they cannot view the
• 2:20
  payroll module in QuickBooks, the model cannot access them
• 2:24
  either. I see the AI just inherits the
• 2:26
  boundaries of the human user. So if you have a junior account
• 2:29
  manager using this and their HubSpot permissions only allow
• 2:33
  them to see their own clients. The AI operates under that exact
• 2:37
  same constraint. It cannot pull data from the
• 2:39
  senior partners accounts to draft an e-mail.
• 2:42
  No, it just mirrors whatever access privileges are tied to
• 2:45
  the login of the person running the app.
• 2:47
  That makes sense. And even with permission
• 2:49
  inheritance, the system still relies on a human in the loop
• 2:52
  safeguard. Which is critical.
• 2:54
  The architecture dictates that the AI only prepares the work.
• 2:57
  Human approval is required before anything sends posts or
• 3:01
  pays. Because the system cannot
• 3:02
  unilaterally execute a sensitive action.
• 3:05
  Exactly. You can think of it like an
• 3:06
  administrative assistant who prepares all the necessary
• 3:09
  paperwork, organizes the files, draft the e-mail.
• 3:12
  And then leaves the entire stack on your desk for a final review
• 3:16
  and signature. Right, the model does the heavy
• 3:19
  lifting of the assembly, but you retain the final authority on
• 3:22
  the execution. And the preparation of that
• 3:24
  paperwork specifically targets the administrative tasks that
• 3:28
  business owners tend to dislike the most.
• 3:30
  It does. The package ships with 15 Ready
• 3:34
  to Run agentic workflows alongside 15 reusable skills.
• 3:38
  These are built around the most common bottlenecks small
• 3:41
  businesses face. The invoice chasing workflow is
• 3:44
  a good example of how those connectors actually interact.
• 3:47
  If you've ever run a service business, you know the friction
• 3:50
  of trying to track down late payments without ruining the
• 3:53
  client relationship. It is terrible.
• 3:56
  The mechanics of that specific workflow are highly sequential.
• 3:59
  1st, the model identifies aging balances in QuickBooks.
• 4:03
  It flags the invoices that are 30 or 60 days past due.
• 4:06
  Then it cross references those specific balances with PayPal
• 4:09
  settlements. That cross reference is the
• 4:11
  crucial step. Why is that?
• 4:13
  Well. It ensures the invoice hasn't
• 4:15
  just been paid through a different channel that hasn't
• 4:18
  synced to the main Ledger yet. Oh, right.
• 4:20
  You never want to send an aggressive follow up e-mail to a
• 4:23
  client who actually paid you 3 days ago through a PayPal link.
• 4:26
  Yeah, that was a bad look. So once it confirms the balance
• 4:30
  is actually outstanding across all connected platforms, it
• 4:33
  moves to Gmail or Outlook. And drafts the e-mail.
• 4:37
  It drafts personalized follow up emails, but it matches the tone
• 4:41
  of those emails to the prior correspondence you have had with
• 4:44
  that specific client. So if you usually e-mail a
• 4:47
  particular vendor with very formal, structured language, the
• 4:50
  draft will reflect that, right? But if you have another client
• 4:53
  where the communication is casual and brief, the AI mimics
• 4:57
  that relaxed tone. It handles a high friction, low
• 5:00
  judgment task, but it does so with actual personalization.
• 5:04
  Rather than relying on rigid automated templates that often
• 5:07
  sound robotic. For B2B businesses, automating
• 5:10
  that process frees up an estimated 4 to 6 hours a week.
• 5:13
  That is a lot of time to get back.
• 5:14
  It shifts the burden from manually checking three
• 5:18
  different systems to simply reviewing and clicking approve
• 5:22
  on a batch of drafted emails. The month end close workflow
• 5:26
  operates on a similar principle, but it targets a process with
• 5:29
  much less room for error, right? Closing the books usually
• 5:32
  involves sitting down with your bank statements and your
• 5:35
  accounting software and just, you know, hunting for
• 5:37
  discrepancies. Under this workflow, the model
• 5:40
  flags reconciliation errors between the bank statements and
• 5:43
  QuickBooks. It identifies the stray
• 5:45
  transactions. It writes a plain English profit
• 5:48
  and loss summary and then exports a complete close packet
• 5:52
  that you can just forward to your external accountant.
• 5:54
  The reality of small business data hygiene comlicates that
• 5:58
  particular workflow, though. What?
• 5:59
  Do you mean? If a company's QuickBooks is a
• 6:01
  mess of uncategorized expenses and mixed personal and business
• 6:05
  transactions, the model's output is going to reflect that chaos.
• 6:09
  A language model cannot fix structural accounting errors if
• 6:13
  the underlying data is disorganized, like if you are
• 6:16
  running your grocery bills through the business card and
• 6:19
  tossing everything into a generic expense category.
• 6:22
  The AI is just going to summarize your poor accounting
• 6:24
  habits, right? The system actually addresses
• 6:27
  that through the Business Pulse workflow.
• 6:29
  OK, that tool continuously monitors for anomalies and
• 6:33
  unexpected expense spikes across the connected apps.
• 6:36
  So it catches things early. Yeah, by constantly surfacing
• 6:40
  those discrepancies in a daily or weekly dashboard, it forces
• 6:44
  the business owner to develop better data habits over time.
• 6:48
  It creates a feedback loop where messy data is immediately
• 6:52
  visible rather than hidden until tax season.
• 6:54
  If a recurring software subscription suddenly double s
• 6:57
  in price, or if a travel expense is categorized under Office
• 7:00
  Supplies, the system flags it immediately.
• 7:03
  You correct it in the moment, which slowly cleans up the.
• 7:05
  Ledger exactly. Collecting money and organizing
• 7:08
  the books is just one side of the equation.
• 7:10
  Generating revenue requires an entirely different set of tools
• 7:14
  and processes. Right, and the system connects
• 7:16
  the financial data directly to the marketing execution.
• 7:19
  Why do they do that? The campaign manager workflow
• 7:21
  illustrates how these systems communicate.
• 7:24
  The AI detects a slow revenue stretch by analyzing the cash
• 7:29
  flow in QuickBooks. It looks at the historical data
• 7:32
  and identifies that you are entering A seasonal dip or that
• 7:36
  sales have been flat for three consecutive weeks.
• 7:39
  It takes that financial trigger and moves over to the marketing
• 7:41
  stack, right? It analyzes past campaign
• 7:44
  performance metrics in HubSpot to see what has worked
• 7:47
  previously. It looks at open rates.
• 7:49
  Click through rates and conversion data from your
• 7:52
  previous emails or promotions. Based on that data, it drafts a
• 7:56
  new promotional strategy and generates the necessary visual
• 7:59
  assets. Canvas specifically partnered
• 8:02
  with Anthropic to build this integration.
• 8:05
  Which is interesting because the generated assets are fully
• 8:08
  editable and they remain on brand based on your company's
• 8:11
  style guidelines. They are not just flat
• 8:13
  generative images that you cannot alter.
• 8:15
  Right. If the AI generates A
• 8:17
  promotional banner, you can open it in Canva, move the text
• 8:21
  around, change the colors, or swap out the logo.
• 8:24
  That reduces the need for a dedicated design budget or an
• 8:26
  outside agency. Particularly for solo founders
• 8:29
  who are running lean, you get a complete campaign drafted,
• 8:33
  designed and ready for approval based on a financial trigger in
• 8:37
  your accounting software. The AI is doing the assembly and
• 8:40
  the execution based on a brief, but the human owner still has to
• 8:43
  provide the creative direction and the strategic intent.
• 8:46
  Yeah, the tool removes the friction of creating the assets
• 8:50
  and draft crafting the copy, but it does not remove the necessity
• 8:53
  of having a good idea in the first place.
• 8:55
  The AI does not know what your customers actually want to buy,
• 8:59
  it just knows how to format the campaign once you tell it what
• 9:01
  the offer is. The focus on assembly and the
• 9:04
  emphasis on ease of use highlights exactly who this tool
• 9:07
  was built for, especially when compared to Anthropic's other
• 9:10
  offerings. Anthropic maintains a strict
• 9:12
  division between Claude Co work and Claude code.
• 9:15
  Co Work provides A graphical interface designed for
• 9:17
  non-technical users. Claude Code is a command line
• 9:21
  interface built specifically for developers.
• 9:24
  The technical differences dictate how they are used and
• 9:27
  who uses them. Co work abstracts the execution
• 9:30
  layer you describe an outcome in natural language and the model
• 9:33
  plans and executes the necessary steps within an isolated virtual
• 9:37
  machine on Mac OS. The virtual machine acts as a
• 9:41
  secure sandbox right when the AI is moving files around or
• 9:45
  interacting with local data, it is doing so in a contained
• 9:48
  environment that is separated from your core operating system.
• 9:51
  It provides A buffer between the AI's actions and your critical
• 9:55
  system files. Cloud Code exposes the execution
• 9:58
  layer directly in the terminal. It manages code repositories,
• 10:01
  execute scripts and interacts with the system at a much lower
• 10:04
  level. It does not use that same same
• 10:06
  virtual machine abstraction. That alters the learning curve
• 10:09
  entirely. Setting up Co work takes menace
• 10:11
  and relies on clicking through an installer and typing natural
• 10:14
  language instructions. Setting up code requires
• 10:16
  terminal familiarity, configuring environment
• 10:19
  variables, and having deep technical context about the
• 10:22
  operating system you are running it on.
• 10:24
  That technical difference connects directly to operational
• 10:26
  risk. How so?
• 10:27
  Co work is safer because it's folder access is structurally
• 10:30
  limited by that virtual machine framework code operates with the
• 10:34
  user's full account permissions. Oh, I see.
• 10:38
  An unreviewed instruction in clawed code could execute
• 10:41
  terminal commands that cause system wide issues or delete
• 10:44
  critical files. Yeah, if you tell a command line
• 10:46
  agent to clean up a directory and you are not explicit about
• 10:49
  the parameters, it has the authority to wipe out your
• 10:53
  entire project folder. The abstraction layer and Co
• 10:56
  work acts as a safety net for non-technical users who do not
• 10:59
  want to monitor terminal outputs.
• 11:01
  Running these continuous multi step agentic workflows, whether
• 11:04
  in cowork or code, requires an immense amount of background
• 11:07
  computing power. It does.
• 11:09
  Anthropic introduced the Max 5X and Max 20X subscription tiers
• 11:14
  specifically for heavy users who need that sustained compute.
• 11:17
  When an agent is working through a task like the month end close,
• 11:21
  it is not just generating a single response.
• 11:24
  It is calling an API, reading a spreadsheet, analyzing the data,
• 11:29
  formatting a summary, and exporting a file.
• 11:32
  Each of those steps requires processing power.
• 11:35
  It is a continuous loop of compute.
• 11:37
  The pricing structure reflects that resource intensity.
• 11:40
  The standard Pro plan is $20.00 a month.
• 11:43
  The Max 5X tier costs $100 a month and provides five times
• 11:48
  the usage capacity of the Pro plan, and the other one, the Max
• 11:51
  20X tier, costs $200 a month and provides 20 times the capacity.
• 11:56
  That structure allows professionals to run extensive
• 11:58
  document analysis, long coding sessions, or complex Co work
• 12:02
  automations without hitting rate limits that disrupt their
• 12:05
  workflow. Because when you are relying on
• 12:07
  an agent to process hundreds of invoices or analyze a massive
• 12:10
  code base, hitting a rate limit breaks the entire utility of the
• 12:14
  tool. You cannot have your automated
• 12:16
  assistant clock out halfway through drafting your marketing
• 12:19
  campaign because you ran out of tokens.
• 12:21
  Exactly. The mechanics of those limits
• 12:23
  operate on a specific schedule. The capacity resets every five
• 12:28
  hours on a rolling window rather than a daily reset.
• 12:31
  Under that structure, the Max 5X plan allows for roughly 225
• 12:36
  messages per window. The Max 20X plan allows for
• 12:39
  about 900 messages per window. Open AI's ChatGPT Pro also
• 12:43
  charges $200 a month, but Open AI offers unlimited usage of
• 12:47
  their advanced models. Anthropic maintains finite caps
• 12:50
  even at the $200 tier. Providing large context windows
• 12:54
  and continuous agentic execution is highly expensive from a
• 12:58
  compute standpoint. Anthropic trades unlimited usage
• 13:01
  for system sustainability to ensure the infrastructure
• 13:03
  doesn't collapse under the weight of heavy users.
• 13:06
  They are enforcing a ceiling to maintain performance stability
• 13:09
  across the network. The cost of that computing power
• 13:12
  is also forcing Anthropic to restructure how they build
• 13:15
  developers who are using third party tools.
• 13:18
  If you are building software or running code editors that lean
• 13:20
  heavily on Anthropic's models in the background, the billing
• 13:24
  mechanics are shifting. Anthropic is splitting their
• 13:26
  subscription billing into two distinct pools.
• 13:30
  There is one pool for first party tools like the Clod
• 13:33
  desktop app, Clod Co work or their official web interface.
• 13:37
  And then there is a separate pool for 3rd party agent context
• 13:40
  protocol usage. The Agent Context protocol is
• 13:43
  the standard used when running clawed inside external tools.
• 13:46
  So if you are using the Zed code editor and you have clawed
• 13:50
  integrated into it to help write or debug your code, that
• 13:54
  connection runs through the ACP. They are implementing a new
• 13:57
  credit system for that third party usage using clawed through
• 14:01
  an ACP connection no longer draws from your standard Pro or
• 14:04
  Mac subscription limits. Oh really?
• 14:06
  Yeah, instead it utilizes a monthly agent SDK credit.
• 14:09
  That credit is $20.00 for the Pro plan, $100 for Max 5X, and
• 14:14
  $200 for Max 20X. Once that specific credit pool
• 14:18
  is depleted, any further third party usage is billed at
• 14:21
  standard API rates. That increases costs for
• 14:24
  developers who rely heavily on 3rd party agents.
• 14:26
  Significantly, under the previous structure, the flat
• 14:29
  rate subscription essentially subsidized this heavy agent
• 14:32
  usage. Developers were getting a return
• 14:34
  of 15 to 30 times the value compared to what they would have
• 14:38
  paid if they were being billed at standard API rates for the
• 14:42
  same volume of requests. You essentially had power users
• 14:45
  running huge code generation tasks through third party
• 14:49
  editors and pulling all that compute out of a flat $20
• 14:52
  monthly fee. The unit economics of that were
• 14:55
  entirely unsustainable for the provider.
• 14:58
  There are practical workarounds for developers dealing with this
• 15:01
  change though. Like what?
• 15:02
  You can still run Anthropics official command line interface
• 15:06
  inside Z terminal window to utilize your standard
• 15:09
  subscription limits. Because the official CLI is
• 15:12
  counted as a first party tool, right?
• 15:14
  It requires adjusting your workflow though.
• 15:16
  You are bypassing the smooth native integration of the editor
• 15:21
  and running the official tool alongside it.
• 15:23
  Alternatively, you can route your built in editor agents to
• 15:26
  different providers utilizing models from a Llama or DeepSeek
• 15:29
  instead. You keep the editor integration,
• 15:32
  but you swap out the engine running it.
• 15:33
  The underlying economics for the AI providers are continuously
• 15:37
  shifting as they try to catch U with the compute demands of
• 15:41
  automated multiste agents. Agents consume resources
• 15:45
  differently than standard chat interfaces.
• 15:47
  I mean, a human types a question, waits for an answer,
• 15:50
  reads it and types another question.
• 15:53
  An agentic workflow fires off 30 queries a minute as it cross
• 15:56
  references databases and formats files.
• 15:59
  AI is moving from being a conversational tool where you
• 16:02
  ask a question and get an answer to functioning as a background
• 16:05
  operating system that actively manages the administration of a
• 16:08
  small business. Right.
• 16:10
  The lingering question is whether traditional accounting
• 16:12
  firms and marketing agencies will adopt these exact same
• 16:16
  agentic tools to scale their own service offerings.
• 16:19
  Or if they will find themselves competing directly against the
• 16:21
  software that their clients are now installing on their own
• 16:24
  desktops. If you're not subscribed yet,
• 16:25
  take a second and hit follow on whatever app you're using.
• 16:29
  It helps us keep making this. We appreciate you being here.
• 16:32
  Also, check out our YouTube channel for more business and
• 16:34
  tech updates. There's a link in the
• 16:35
  description.