The next model can wait. Fit one job into a workflow that exists now
Anthropic released 43 ready-made small-business workflows on September 15, three days after its own CEO asked rival labs to slow down how fast they make models more capable. The useful move: run one job you already review each week inside a workflow that exists now, with every send and payment held for your sign-off.
In this edition
The decision in 30 seconds
Read this if you read nothing elseRun one job you already review every week, such as the Monday numbers or the month-end close, inside a ready-made workflow in approval mode for two cycles before judging it.
On September 15, Anthropic's Claude for Small Business plugin grew to 43 workflows and 37 connectors, including QuickBooks, Xero, Shopify, HubSpot, Stripe and Gusto, and Anthropic reported more than 900,000 installs since May. On September 12, Anthropic's CEO published an essay asking AI labs to slow capability growth. OpenAI's CEO agreed, the US president called the fears a hoax, and on September 14 the main chip index fell 5.9% while business software stocks rose.
The workflows that already exist cover more jobs and are available without a custom build, and Anthropic and OpenAI have said in public that they intend to slow how fast models gain capability and accept outside review. Whether they hold to it or not, waiting for a better model now has no date attached, and a workflow you can run does.
The pacing debate itself. The September statements did not announce a change to the products your business uses.
Do this week
Only where it appliesInstall Claude for Small Business and run one job
Install the plugin, say 'set me up' or run /smb-onboard, and connect only the one or two tools the chosen job needs. Choose the job by what you already check every week, not by which connector you happen to have. Leave approval mode on: Claude drafts and stages, and nothing sends, posts or pays until you say so. Judge after two runs on one question: did the draft cut your review time?
Limit: Desktop app only, Pro plan and up; Anthropic recommends Team for more than one person. If your team is on ChatGPT the plugin is not available, but the rule still applies: one recurring job, a draft you review, then decide.Check the training setting before a workflow touches real data
404 Media reported on September 14 that OpenAI contractors read real ChatGPT conversations to rate responses; OpenAI says it tries to remove personal details first but, per 404 Media, acknowledged that sensitive details can still get through. Enterprise, Business and Edu plans have training off automatically. Turn off 'Improve the model for everyone' under Settings, then Data Controls, or move client work to a business plan.
Limit: This is reporting, not a breach. The opt-out stops new chats being used for training; it does not remove all human review, and submitting feedback can share the related conversation for training. Anthropic told 404 Media it also uses human review, so this is a plan-and-settings question for every vendor, not a reason to switch.
Ready-made workflows are now broad enough to matter
For owners who want one job off their plate without a build project.
Product releaseConfirmed by Anthropic's announcement and plugin pageClaude for Small Business now covers the back office and the pipelineAnthropic's small-business plugin now has 43 ready-to-run jobs that read your accounting, sales and calendar tools and draft the work for your approval.Applies to: Owner-led businesses on QuickBooks, Xero, HubSpot, Shopify, Stripe, Square or Gusto that run Claude's desktop app
Anthropic launched the plugin in May with 15 workflows. The September 15 release takes it to 43 workflows and 37 connectors, adding Shopify, Salesforce, Xero, Gusto, Square, Stripe, Zoom and Zapier. The named jobs are ones an owner already does by hand: a Monday brief of cash, sales, pipeline and overdue invoices; after-hours lead replies that offer real meeting times and log the lead to your CRM (sales contacts); a proposal priced from past jobs; a week of social posts staged for approval; and a month-end close reconciled against sales settlements, card spend and payroll. For tools with no connector, build-connector links them through Zapier, and build-agent turns a task you repeat every week into a workflow you can run by name or on a schedule.
- Every workflow starts in approval mode: Claude drafts and stages, nothing sends, posts or pays until the owner says so, and each can later be switched to run on its own.
- No connector is required to start: Anthropic says most workflows still run from an uploaded spreadsheet or bank statement, the real on-ramp for owners not yet on QuickBooks or HubSpot.
- It runs only in Claude's desktop app on a paid plan. Half of owners in Anthropic's own survey named data security as their biggest hesitation; existing tool permissions hold, and Team and Enterprise plans do not train on business data by default.
If you already review a weekly numbers page, a lead inbox or a month-end close by hand, run that one job through the matching workflow (/monday-brief, /speed-to-lead or /close-month) and keep approval on. It works because the workflow reads the tools you already check and hands back a draft, so the only new thing to learn is whether that draft is worth trusting. It does not help a job with no clear owner or no weekly rhythm; a workflow makes a messy process faster at being messy.
Sources: Anthropic: Claude for Small Business release, September 15, 2026, Anthropic: Small Business plugin page, Anthropic: original Small Business launch, Claude Academy: plugin install guide
The industry argued about slowing down, and the money moved toward deployment
For executives whose AI plan is to wait for the next model.
Industry directionConfirmed events; Sidekick inference on the implicationOne lab CEO asked to slow capability growth, a second agreed; Washington and the markets answered within two daysAnthropic's CEO said labs should slow how fast they make models more capable, OpenAI's CEO agreed, the US president called the fears a hoax, and chip stocks fell while business software stocks rose.Applies to: Any executive whose AI plan for 2026 is to wait for a clearly better model
On September 12, Dario Amodei published 'We Must Pace the Frontier', writing that 'we must slow the pace at which we improve the capabilities of AI models' and committing Anthropic to third-party evaluators with employee-like access. Within a day Sam Altman wrote, 'I agree with Dario that we need to pace the frontier', and that OpenAI 'will do the same' on evaluators. On September 14, President Trump wrote that AI 'destroying Humanity, and all other things bad, is a HOAX'. That day the PHLX semiconductor index, the main chip index, fell 5.9% and Nvidia 3.4%, while ServiceNow, Adobe and Workday rose between 4% and 7.4%; Reuters attributed the chip selloff to the AI warnings.
- Amodei's essay warns that within '6-12 months' an agent swarm could be 'capable of taking over the entire internet with a persistent botnet'. It says nothing about current products or customers.
- Not all of the selloff was about AI: the 10-year Treasury yield briefly passed 5% the same day.
- The president dismissed the concerns on Truth Social; the White House gave Axios no comment.
If your plan for AI in 2026 is to wait for a clearly better model, replace it with one job you can run now: Anthropic and OpenAI have said they intend to slow capability growth and accept outside review, and the same day the money moved from chip makers toward the companies that put AI into everyday business software. The limit: this is a stated intention and one trading day, not a product change. The September statements did not announce a change to the products your business uses.
Sources: Dario Amodei: We Must Pace the Frontier, September 12, 2026, SiliconANGLE: Altman and Musk back Amodei, September 2026, Axios: Trump on AI safety and Amodei, September 14, 2026, Reuters via Investing.com: chip stocks fall, September 14, 2026
Further reading
- Apple: Siri AI begins rolling out in beta
A capable assistant with access to mail, messages and calendar began rolling out in English on supported iPhones on September 14; not yet in the EU or China.
Confirmed - OpenAI: ChatGPT data controls
The training setting and what it controls. Turning it off applies to new chats and does not remove all human review.
Confirmed - Microsoft AI: draft Humanist AI code of conduct
A readable statement of what 'remains under human control' means in practice; useful as a template for an internal AI-use policy. A six-week consultation opened September 14.
Confirmed - Bryna Kra on Terence Tao's blog: when AI makes the output cheap
A mathematician's argument that explaining and standing behind work becomes the scarce skill when producing it is easy. It maps cleanly onto reviewing AI drafts.
Confirmed - RubyGems: what is known about the May spam attack
RubyGems confirmed registry abuse but could not determine whether AI agents created or published the packages.
Confirmed
Sources and further reading
- Anthropic: Claude for Small Business release, September 15, 2026
- Anthropic: Small Business plugin page
- Dario Amodei: We Must Pace the Frontier, September 12, 2026
- Axios: Trump on AI safety and Amodei, September 14, 2026
- Reuters via Investing.com: chip stocks fall, September 14, 2026
- 404 Media: Inside Project Lily, September 14, 2026
Evidence note: Sources checked September 25, 2026. This edition covers September 12 to 15, 2026 because the September 11 edition closed the prior window. Prepared September 15, 2026; sources rechecked September 25, 2026. Anthropic's workflow, connector and install figures are its own; an install is not a business running a workflow. Market figures are September 14, 2026 closes as reported by Reuters. The Project Lily account rests on 404 Media's reporting. That value is moving toward deployment is Sidekick's reading of the week, not a claim any source makes.
Get the next Skim.
New editions by email when they clear the evidence and usefulness checks.
Bring one decision. Leave with a next step.
Use the 30-minute call to work through that specific model, workflow, or risk choice.