Turn manufacturer emails into service calls and draft replies for your approval
Have Claude Code read the mailbox, turn manufacturer emails into service calls and draft the replies, with anything that has consequences waiting for the owner's approval on his phone.
What the agent did
Source-reported example · as described by Jorge Garzón (dev.to user satgarzon), second-generation owner of a small family car-wash equipment service company in Spain, writing about his own company
The author, Jorge Garzón, second-generation owner of a small family company in Spain that services and repairs car-wash equipment for two large manufacturers and its own customers (a handful of field technicians and one office; he says nobody there is a programmer), built a set of agents on Claude Code and an MCP server alongside the company's own ERP.
The agents read the mailbox, create service calls from manufacturer emails and draft replies; they prepare emails but never send them. Anything with consequences, such as creating a document or sending something, shows up on his phone in Telegram with two buttons for his approval, and every decision is logged (who, when, result).
The tooling requires the agents to consult the company's written procedures before they are allowed to write anything, and a second model from a different company reviews the first model's work and flags disagreements.
Reported: "I review twenty decisions a day from the phone in a few minutes, and nothing surprising ever happens." This figure covers the whole approval loop, not only service calls and replies. No counts of emails handled, service calls created, time saved or accuracy are given for this job.
Limitations
The post is a short overview of the whole system, not a write-up of this job: it gives no examples of a manufacturer email, a created service call or a drafted reply, and no volumes, time saved or error rates for this job.
The 'twenty decisions a day' figure covers every approval, not just this job. The author's first approach let the agents act and report afterwards, and the company lost a day to a batch of work reports "approved" with the wrong prices (a different task, but the reason everything with money now needs a human click).
He says they measured one long working session re-reading hundreds of thousands of tokens on every step, so they switched to one session per topic with a short hand-over note; anything the AI "remembers" in chat but that is not written in a procedure or file "does not exist the next day".
Local models derail on anything longer than a page. The setup is custom-built (own ERP, MCP server, Telegram approvals, about 1,500 written procedures) and the author does not publish code, configuration or prompts. He warns owners to expect to maintain it themselves.
The post describes other jobs (reconciling manufacturer approvals against work reports, preparing supplier orders, a technician Q&A library, the ERP itself) that are not part of this listing. dev.to labels the post 'AI-assisted'.
Try it in your business
What you'll need
Claude Code access; a business system that holds service calls (the author built his own ERP) plus an MCP server or similar connection that lets the agent read the mailbox and create records in it; an approval step on your phone or desktop (the author uses Telegram with approve/reject buttons and a decision log); written procedures the agent must read before it writes anything (the author's company has about 1,500); your own review and approval of every service call and reply before it is created or sent
The agents run on Claude Code through an MCP server connected to the company's own ERP and mailbox, with a custom Telegram approval flow and tooling that blocks writes until procedures are read; this is self-built software, not a packaged app connection.
The original is a custom-built system. The prompt below is a simpler starting point and won’t reproduce it.
Starting prompt
The author didn't publish their prompt, so this is our adaptation.
How ours differs from the original setup: Same execution agent (Claude Code). The author's agents connect to his self-built ERP through an MCP server and send approvals to Telegram; our prompt assumes you point Claude Code at your own mailbox and job system and an approval queue of your choice.
It keeps everything as drafts until you approve, matching the author's rule to "Separate 'propose' from 'do'", and never sends, orders, deletes or changes records. The service-call fields and the reply content are our additions, since the source does not describe them.
You help me run the office of my equipment service company. Check the mailbox [mailbox] for new emails from the manufacturers we service for [manufacturer names]. Before doing anything, read our written procedures for handling manufacturer requests [procedures file or folder]. For each email that asks for a service visit, prepare a draft service call (do not save it in [our job or service system] until I approve) with the customer site, machine model, fault reported, manufacturer reference and urgency, taken only from the email; mark anything missing or unclear instead of guessing. Then draft a reply to the manufacturer confirming what we understood and asking for any missing details. Put each draft service call and draft reply in my approval queue [where] with a one-line summary. Do not create, change or close any record, send any email, order parts, or delete anything; I will review and approve each one myself.
Replace the [bracketed] parts with your own files and tools.
Get examples like this every Tuesday and Friday
Practical ways to put AI agents to work in your business. We check the original sources, explain the setup and point out the limits. Free.
Unsubscribe any time. Privacy policy
Related use cases
Run a new client's standard audits and draft the kickoff strategy brief
Have Claude Code run a new client's standard first-week audits and draft a strategy brief with recommended priorities for you to review before the kickoff call.
Our suggested starting prompt — not the author's original · untested
You'll need: Claude Code; read access to the client's website and, for the analytics and CRM reviews, the client's Google Analytics/Tag Manager and email/CRM tools; a written audit checklist; a brief template; the owner's review before anything is sent
We're onboarding a new client: [client name], website [URL], industry [industry], main goal [goal] (from their intake form in [file]). Please prepare a first-week audit and a draft strategy brief for me to review. 1) Technical SEO: check the public site for crawlability, page titles and descriptions, broken links, page speed and mobile issues. 2) Analytics: from the export or read-only access I've given you [path/access], note whether Google Analytics and Tag Manager are set up and tracking key actions. 3) Competitors: for these competitors [list], compare the main keywords and topics they appear to target. 4) Email/CRM: from my notes [file], summarize their current setup and gaps. Save each audit as a separate file with findings, evidence and what you couldn't check. Then draft a one-to-two page strategy brief with the top 3–5 recommended priorities, each tied to the audit findings, and list open questions for the kickoff call. Keep everything as drafts in [folder]: don't send anything to the client, don't change any site, analytics or CRM settings, don't buy any tools and don't delete anything. Ask me before using any login or access I haven't listed.
Onboard a new client end to end: draft the contract, set up their Slack channel and podcast feed, and build website mockups for your review
Have Claude run a client-onboarding skill that drafts the client's contract, creates their shared Slack channel, adds them to your client database, and prepares samples, website mockups and a podcast feed for your review, while you send the invites, grant access and write the client messages yourself.
Our suggested starting prompt — not the author's original · untested
You'll need: Claude with skills (or saved instructions) describing your onboarding process; your internal docs, past contracts and the client's correspondence; access to your documents (for example Google Drive and Docs) and, if you let it do setup steps, to your team chat (the author used Slack via Claude's browser use) and client database; your own review of the contract, every account, channel or access change, and every client message. Building the skills was a one-off investment of about a day of the author's time, by his estimate.
Help me onboard a new client for my [type of service] business. Work in stages and stop for my approval at the end of each one. Stage 0 (first time only): Read my internal docs and past client notes in [folder]. Write a short overview of what my business does and how it works, then draft a step-by-step client onboarding checklist. I'll correct it. Once I approve it, save it as reusable instructions I can run for future clients. Stage 1, contract: Using the client emails and call notes I've attached and my past contracts in [folder], draft a service contract for [client] as a document. Look up their public details (legal name, registered address, registration number) and list where you found each one. Highlight anything that differs from my usual contract or that you're unsure about. Don't send it or upload it to a signing service. I'll review it and send it myself. Stage 2, setup: List every channel, record and account this client needs (for example a shared [Slack] channel that follows my naming pattern, or an entry in my client database), and wait for my OK before you create each one. Don't invite the client, grant access, change permissions, or enter any password or payment details. Give me the invite link or the exact page to open, and I'll do those steps. Stage 3, tailoring and samples: Review [the client's website or materials], list the special cases my service needs to handle for them, and propose a rule for each. After I approve the rules, prepare [sample deliverables] for me to check, and flag anything you're unsure about. Stage 4, mockups and delivery setup: Make mockups of [how my deliverable would look on their website], using their public branding, and draft [the feed, page or other delivery setup] privately. Don't publish anything until I approve it. Stage 5, client update: Don't write or send messages to the client. Give me a bullet list of points to cover in my update, then check my draft for clarity and for anything I've missed. Finish with a checklist of what's done, what's waiting for my approval, and what I need to do myself.
Project inventory and recommend when to reorder each product
Have Claude Code pull inventory and sales data, project the next twelve months of sales and stock by product, and recommend when to reorder each one, using written notes on supplier lead times and holidays.
Author's example instructions
You'll need: Claude Code; sales and inventory data by SKU (the author pulled it from Amazon with scripts Claude Code wrote); written notes per brand covering supplier lead times, holiday closures, off-platform stock and past stockouts; an instruction file or command listing the projection steps. Not disclosed: The scripts' code and how the Amazon credentials were connected; the full text of the instruction file and per-brand context files; the number of brands and SKUs; how accurate the projections turned out against actual sales.
1. Check to see if the sales/inventory data we have locally is >5 days old, and if not, pull fresh data from Amazon. 2. Check the context notes for the brand it's working on. 3. Pull inventory and sales data from the local cache. 4. Display the trailing twelve month sales data by SKU and by month. 5. Describe any factors it's going to use to make adjustments to TTM sales data when projecting NTM sales. 6. Review its stated adjustment factors, then project NTM sales by SKU on a monthly basis, and display the amount of stock remaining at the end of each month based on those projections. Use Python for the math here. 7. Recommend when to reorder each SKU, factoring in CNY and other holidays.

