DEV Community: Or Emergy

Set up a first invoice chasing workflow in Claude

Or Emergy — Fri, 16 Oct 2026 07:39:21 +0000

Updated October 15, 2026: the October 6 change to cloud tasks has happened; dates now read in the past tense.

Chasing overdue invoices is the job that waits for a free hour, and in a small company that hour rarely comes. Two changes at Anthropic make it a sensible first task to hand to Claude. On September 15 the Claude for Small Business plugin grew to 43 ready-made workflows, and on September 16 Anthropic started folding Cowork, its agent for multi-step work, into the regular Claude app.

What changed

Chat and Cowork are merging into one Claude. There is no mode to choose: Claude decides whether a request is a quick answer or a task that needs your files, tools and a plan. Anthropic says people used both, and the frustrating part was deciding where a task belonged. The merged app is rolling out gradually to Pro and Max first. Team organizations keep chat and Cowork separate for now.

The Small Business plugin launched in May with 15 workflows. It now has 43, plus 27 new integrations including Shopify, Salesforce, Xero, Gusto, Square, Stripe and Zapier, next to QuickBooks, HubSpot and PayPal. Anthropic counts more than 900,000 installs.

Since October 6, new Cowork tasks on Pro and Max run in Anthropic's cloud instead of on your computer. Your folders stay put: Claude reaches only the folders you've connected, through the desktop app while it's open, and fetches a copy of just the file a task needs.

What it means for a small business

Picture a design studio of eight people that bills clients monthly from QuickBooks or Xero. Every couple of weeks the office manager pulls the overdue list, decides who gets a polite nudge and who gets a call, and writes each email by hand. The invoice-chase workflow takes over the assembly: it ranks overdue invoices and drafts a reminder for each one, matched to how that customer has paid before. The office manager reads, edits and approves. The judgment stays with a person; the copying of numbers into emails goes away.

The merge matters less than it sounds. The plugin is what changes the week, because a workflow that already knows your business and tools needs a one-line instruction, not a long prompt.

Two things shift around it. Approvals become a small daily habit, since every workflow starts in approval mode. And since October 6, Pro and Max work is processed on Anthropic's servers, including any local file a task opens.

How to set it up

Pick the plan. Pro for one person. For two or more, Anthropic recommends Team: an owner controls the connected tools, and business data isn't used for training by default.
Install the plugin. In the desktop app, open Customize in the left sidebar, then Plugins, then Discover, and add Small Business. On Team, open the Cowork tab first, then Customize.
Run the onboarding. Type /smb-onboard or say "set me up". Claude asks about the business, connects your tools and saves that context for every workflow.
Connect the books. Under Customize, then Connectors, connect QuickBooks, Xero, Stripe or wherever your invoices live. On Team, an owner adds the connector first, then each person signs in with their own account. No connector for your system: upload an export of open invoices, which Anthropic says most workflows can run on.
Run invoice-chase in Manual mode. Read every draft before it goes out, for at least two billing cycles.
Schedule it carefully. Since a desktop update on October 1, scheduled tasks Claude creates for you default to "Automatically approve" where your organization allows it. Switch anything that sends email or touches money back to approval. Team owners can turn off "Allow 'Automatically approve' mode" for everyone.

Costs, limits and risks

Pro costs $20 a month, or $17 billed annually. Team takes 2 to 150 seats; a Standard seat is $25 a month, or $20 billed annually, so five people on annual seats pay $100 a month. Anthropic lists no separate fee for the plugin. These are US prices before tax and vary by region.

Usage runs on a rolling five-hour window plus weekly limits; a Team Standard seat gets 1.25 times Pro's per-session allowance. Agentic tasks use more than chat, and automatic approval more still. A rough estimate: if reviewing a draft takes two minutes instead of ten to write one, 20 overdue invoices a month gives back about 160 minutes. That is arithmetic on assumptions, not a measurement.

The risks are mostly about data:

In Anthropic's own survey, half of small business owners named data security as their biggest hesitation about AI. Permissions carry over: if someone can't see payroll in QuickBooks, they can't see it through Claude.
On Pro and Max, cloud processing since October 6 is not optional. On Team, an owner can switch off "Run Cowork in the cloud", which is on by default.
Team has no audit logs; those are Enterprise only. On Pro and Max, check the privacy setting for model training.
Claude is available in Israel. The interface isn't in Hebrew, but Claude replies in the language you write in.

Should you do it now?

Do it now if your invoices live in a connected tool and you chase more than a handful a month. Keep Manual mode until the drafts stop needing edits.

Wait if you want the single merged app on Team, where it hasn't arrived, or if client files must stay on your own machines. On Pro and Max that is no longer possible in Cowork; use Team with cloud sessions off, or Claude Code desktop, which Anthropic suggests for work that has to stay local.

Skip invoice-chase if late payments are rare, and start with the workflow that matches your most repetitive week, like month-end-prep or the Monday brief.

What I'd do this week: install the Small Business plugin, connect the tool your invoices live in, and run invoice-chase in Manual mode on last month's overdue list. Approve nothing you wouldn't have sent yourself, and count how many drafts went out with only light edits. If it's most of them after two cycles, give it a schedule, with approval still on.

Sources

Drafted with AI and checked against the linked sources before publishing. Originally published at emergyaiops.com.

Gemini skills replace Gems: turn your best proposal into one

Or Emergy — Fri, 16 Oct 2026 03:06:56 +0000

Updated October 15, 2026: Google paused skills in Workspace on October 13; the timing below is corrected.

Most small companies have one document that only comes out right when one person writes it: the proposal, the quote, the client update. Gemini skills are built for that document. Google announced them on September 30 and started rolling them out in the Gemini app on October 13. Skills inside Gmail, Docs and Drive are delayed, and they replace Gems, which business accounts can keep using until March 2027 at the earliest.

What changed

Google calls skills "reusable, custom instructions" for Gemini. In Workspace you call one from the Ask Gemini panel in Gmail, Docs, Slides, Drive or Chat by typing @ and its name, and Workspace Studio flows can use them too. Several can be stacked, and they are plain Markdown files in the open SKILL.md format. A Skill Builder in Docs helps you write one.

Timing changed on October 13: Google delayed the Workspace rollout, due by October 12 for Rapid Release and from October 19 for Scheduled Release, until further notice. Skills in the Gemini app began a gradual rollout the same day, expected to finish by mid-November. The two don't sync: a skill built in Workspace has to be recreated in the Gemini app.

Gems are on the way out. On November 17 they move into Settings. From March 1, 2027 at the earliest, business accounts can no longer create, edit or use them. Workspace Studio already blocks new "Ask a Gem" steps, and existing ones keep working until the cutoff.

What it means for a small business

Picture an architecture studio of 12. Every proposal goes through one partner, who rewrites drafts so they follow the same structure, the same fee logic and the same tone. That knowledge lives in one head. A skill is a place for it to live instead: a page of instructions, the template and two strong past proposals, called with @ inside the Doc where the draft already is. A junior architect drafts in the studio's format, and the partner reviews instead of rewriting.

Skills are also readable: anyone with access can see what the skill tells Gemini, change one line, and every later draft follows. That makes the Gems deadline a chance to rewrite rather than copy; a Gem written in a hurry last year rarely matches how you work today.

There's a second reason to move this work into Workspace now: personal Gemini accounts are getting tighter. From October 9, personal users without a subscription get only the Flash-Lite model. Using skills on a personal account requires Keep Activity on, and with it on, some chats are read by human reviewers; Google asks people not to enter confidential information. In Workspace, Google says business content isn't human reviewed or used to train models outside your domain without permission. Client proposals belong there.

How to set it up

Check the plan. Google's comparison shows Gemini in Docs, Slides and Drive on Business Standard and Plus, not Starter. The Docs route below needs Standard or Plus; on Starter, build the skill in Workspace Studio and use it in Gmail.
Check the admin side. In the Admin console, Account settings, then Preferences, shows your release track. Skills are switched on under Apps, then Google Workspace, then Workspace Studio, then Skills; changes can take up to 24 hours. Google's admin page still calls skills a Gemini Beta feature. With the Workspace rollout paused, they may not appear yet.
Write the page first. On one page, describe how the proposal is done: sections, fee logic, tone, what never goes in. Gather the template and two strong past proposals.
Build the skill in Docs. Once skills reach your domain, open a Doc, click Ask Gemini, then Tools, then Skill builder, and describe the skill using your page. Review what it drafts, click Make this skill in Studio, then Turn on.
Share it. In Studio, open the skill, click Share and choose My organization. The copy includes reference files and linked flows, and anyone with the link can see the creator's email address.
Test it on a real case. In a Doc, open Ask Gemini, type @, pick the skill, and run it on a proposal you already sent. Compare the two.
Rebuild the Gems that matter. List the Gems your team uses, keep the two or three worth keeping, and recreate them as skills by hand. At the cutoff, leftover Gems become draft skills in the Gemini app, not in Workspace.

Costs, limits and risks

On annual plans, Business Starter costs $7 per user per month, Standard $14 and Plus $22; flexible plans cost more. A 12-person studio on annual Standard pays $168 a month. Viewed from Israel, the pricing page shows euros: €13.60 for Standard on an annual plan. Google lists skills on all three Business plans.

Limits worth knowing:

Google lists Gmail, Docs, Slides, Drive and Chat for skills, not Sheets.
The Gemini side panels in Gmail, Docs and Drive support Hebrew, but Gemini in Workspace Studio doesn't list it. Test a skill on a Hebrew proposal before trusting it.
Workspace users under 18 can't create or share skills.
A skill sees only what its user can access: Drive and the web by default, plus Gmail, Chat and Calendar if checked. If someone can't open a file, Gemini won't retrieve it for them.

The bigger risk is quality. A skill encodes judgment only as well as the page behind it: if the page is vague, every draft comes out vague in the same way, only faster.

Should you do it now?

Do it now if one recurring document still routes through one person: write the page from step 3 today, and build the skill in the Gemini app once it reaches your account. Build one skill, not five.

Wait if you need the skill inside Docs and Gmail, where it is most useful for a proposal: Google has not set a new date for Workspace. There's no rush on Gems themselves: they keep working until March 2027 at the earliest.

Skip it if no document in your business depends on one person's judgment. A skill adds little to work anyone on the team already does well.

What I'd do this week: write one page on how your proposal should be done, gather the template and two good past proposals, and turn them into a skill with Skill Builder once skills reach your domain. Share it with the person who drafts the most. Next Friday, compare their draft with what you used to rewrite, and count the changes you still had to make.

Sources

Drafted with AI and checked against the linked sources before publishing. Originally published at emergyaiops.com.

Zapier moved its agents into the Zap: triage quote requests with one AI step

Or Emergy — Thu, 15 Oct 2026 20:30:06 +0000

Quote requests are the email that waits. Someone, often the busiest person in the office, has to read it, check the price list, work out what the client wants and write a first answer. On October 6, Zapier confirmed that its standalone Agents product now lives inside the regular Zap editor as one AI step. The useful part is the design: the AI now sits inside an ordinary automation, with fixed steps around it.

What changed

Zapier Agents used to run on its own site, agents.zapier.com, outside the rest of the platform. According to Zapier's announcement and its migration guide, the same four parts (a prompt, tools, a trigger and permission to act) now live in an AI by Zapier step you drop into any Zap.

What comes with the move:

Per-tool approval. Each tool has a "Require approval before running" toggle. Agents was all or nothing.
Model tiers. Standard, Advanced or Premium, or a named model, each with its own task cost.
Structured output. You define the fields you want back, such as a category or a subject line, and map them into later steps.
One bill and one history. AI runs use the same tasks as every other Zap, and every run shows in Zap History with the tools called and the tasks used.
A runaway brake. A run that hits its task limit, 75 by default and adjustable up to 500, pauses and waits for a person.

Existing agents convert into a Zap with a trigger and one AI step. Zapier has set no shutdown date for the old site and says it will email users first.

Where the AI belongs in your workflow

Picture a renovation company of 10. Quote requests reach a shared Gmail address, some clear, most vague, and one partner answers each one at night from the price sheet.

The useful split is this: let the AI do the reading and the reasoning, and keep everything that touches the client fixed. The AI step reads the email, looks up the matching items in a Google Sheets price list, and returns structured fields: job type, rough size, missing details, urgency and a draft reply. Two ordinary steps follow: one creates a Gmail draft, never a sent email, and one logs the request in a sheet. The partner starts the morning with drafts, not a blank page.

Zapier's announcement argues the same: reasoning where it adds value, predictability everywhere else. It also caps the damage. The AI can misread a request, but it cannot email a client a wrong price, because sending is not one of its tools.

How to set it up

Create a Gmail label called Quotes and a filter that applies it to quote requests, or forward your website form there.
Build the Zap. Trigger: Gmail, new email matching the Quotes label. Add an AI by Zapier step.
Pick the tier on purpose. Zapier's help center says new steps on paid plans default to Premium, at 5 tasks per unit. Tools need Advanced or Premium, so switch to Advanced (3x) unless your tests show you need more.
Add one tool. Under Tools and Knowledge, add a Google Sheets "lookup row" action on your price list. Leave approval off for this one, since it only reads. If you ever add a tool that writes or sends, turn Require approval before running on for it. It is off by default.
Write the prompt and output fields. Describe your services, what a complete request contains, and the reply you want: short, in the client's language, no prices beyond the sheet. Define the output fields above.
Add the fixed steps. Gmail "Create draft" using the draft field, then Google Sheets "Create row" with the job type and missing details.
Test on ten old requests. Use Preview, then Test Zap. Check the tools used, the tasks and every price before you publish.

What it costs and what to watch

Zapier's formula is: tasks per run = (1 × model rate) + (tool calls × model rate). On Advanced with one Sheets lookup, that's 3 + 3 = 6 tasks. The draft and the log row add 1 task each, so about 8 tasks per email. Triggers don't use tasks.

Assume, as an estimate, 150 quote requests a month: 150 × 8 = 1,200 tasks. Zapier's pricing page puts Professional with 1,500 tasks at $39 a month billed annually, or $58.50 billed monthly. Left on the Premium default, the same run costs 5 + 5 + 2 = 12 tasks, or 1,800 a month, which moves you to the 2,000 task tier at $49 annually or $73.50 monthly. If you pass your limit, Zapier switches you to pay-per-task billing at a higher rate unless you turn that off.

The risks to check:

Data. Every quote email passes through Zapier to an AI model. Zapier's help center lists Claude models as the defaults: Claude 4.5 Haiku for Advanced and Claude 4.6 Sonnet for Premium. Read Zapier's AI data terms before sending client details.
Admin settings. On accounts where Agents was blocked, tool calling starts off. An admin can also require approval before any Zap with an AI step is published.
Hebrew. None of the Zapier pages cited here mention Hebrew. Test with real Hebrew requests and check the draft's direction and tone in Gmail.
Gaps. Knowledge sources from HubSpot, Asana, Zendesk, Zoom and Jira are not yet supported in the new step. Collab365 warns that a converted agent is not automatically production ready.

Should you do it now?

Do it now if quote or inquiry emails pile up behind one person, your price list already lives in a sheet, and you're on a paid Zapier plan.

Wait if you run agents on agents.zapier.com that work today. With no shutdown date yet, migrate one at a time and add an approval gate as you go.

Skip it if you get only a handful of requests a week. Reading five emails takes less time than maintaining an automation.

What I'd do this week: build the Quotes Zap on the Advanced tier with one read-only price lookup, output to Gmail drafts only, and run it on ten past requests. Success is at least seven drafts you would send after light edits, at under 10 tasks per email in Zap History.

Sources

Drafted with AI and checked against the linked sources before publishing. Originally published at emergyaiops.com.