8 Claude API Integrations That Run My Business While I Sleep

By Lazar Milicevic · Published June 10, 2026 · 8 min read

Most founders I talk to are paying for Claude Pro and still spending three hours a day in their inbox. The model is brilliant in a browser tab and completely absent from the workflow. Here are the eight integration patterns I deploy for clients to close that gap — with the wiring, the numbers, and the parts tutorials skip.

1. Inbox triage: Gmail → Claude → Telegram

The pattern is dead simple and the highest-ROI build I do. A script polls Gmail every two minutes, pulls unread messages, sends subject and body to Claude with a classification prompt, and routes the result.

• Leads → drafted reply pushed to Telegram with Approve / Edit / Skip buttons.
• Vendor invoices → forwarded to the extraction pipeline (integration 2).
• Support requests → tagged and queued.
• Noise → archived.

SYSTEM = """You are an inbox router. Classify the email into exactly one of:
LEAD, SUPPORT, INVOICE, NOISE. Return JSON: {"class": "...", "reason": "..."}"""

resp = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=150,
    system=SYSTEM,
    messages=[{"role": "user", "content": f"Subject: {subj}\n\n{body[:2000]}"}],
)

On a real client mailbox doing ~120 messages/day, this runs at about $14/month in API spend and replaces roughly 2 hours of inbox work daily. The founder never opens Gmail to triage again — they live in Telegram and tap buttons.

2. Invoice extraction with vision and a strict schema

PDFs arrive, get rasterized to high-res PNGs, and hit Claude's vision endpoint with an example of the exact JSON shape I want back. Output lands in Postgres.

The three things most tutorials miss:

• Do not send the raw PDF blindly. Convert to images at 200+ DPI first.
• Set temperature=0.
• Include a worked example of the JSON structure in the prompt — not just a schema description.

PROMPT = """Extract this invoice into JSON matching exactly:
{
  "vendor": "Acme GmbH",
  "invoice_date": "2024-11-03",
  "currency": "EUR",
  "line_items": [{"desc": "...", "qty": 1, "unit_price": 0.0}],
  "subtotal": 0.0, "tax": 0.0, "total": 0.0
}
Return only JSON. No prose."""

On the messy, scanned, multi-language invoices small businesses actually receive, accuracy jumps from around 70% (lazy "here's a PDF, extract it" prompts) to 94% with the image conversion + zero temperature + worked example combination. The remaining 6% get flagged for human review.

3. Daily ops summary at 7 a.m.

A cron job fires at 06:55. It hits the client's database — new signups, MRR delta, churn events, open tickets, pipeline movement — packs the numbers into ~800 tokens of context, and asks Claude for a two-paragraph briefing.

# crontab
55 6 * * * /usr/bin/python3 /opt/ops/daily_brief.py >> /var/log/brief.log 2>&1

The briefing lands in Telegram before the founder finishes coffee. No dashboard. No tab. The prompt is boring on purpose:

"Write a two-paragraph operations briefing. Paragraph one: what changed since yesterday. Paragraph two: what needs attention today. Concrete numbers only. No filler."

That last line is doing a lot of work. Without it you get "exciting growth opportunities."

4. Calendar booking via tool use

This is where Claude stops being a writer and becomes an agent. Two tools, one loop.

tools = [
    {
        "name": "check_availability",
        "description": "Return free 30-min slots in the next 7 business days.",
        "input_schema": {"type": "object", "properties": {
            "timezone": {"type": "string"}
        }, "required": ["timezone"]},
    },
    {
        "name": "book_slot",
        "description": "Book a calendar event and send invites.",
        "input_schema": {"type": "object", "properties": {
            "start_iso": {"type": "string"},
            "attendee_email": {"type": "string"},
            "title": {"type": "string"},
        }, "required": ["start_iso", "attendee_email", "title"]},
    },
]

A lead emails asking for a demo. Claude reads the thread, calls check_availability, proposes three slots in plain English, the lead picks one in reply, Claude calls book_slot, calendar invites go out. The founder sees the event appear on their calendar. That's the whole interaction.

5. Lead scoring with a one-sentence reason

Every form submission and cold reply gets enriched (company size, industry, role) and scored 1–10 with a one-sentence justification.

• 9–10 → push notification to the founder's phone.
• 4–8 → sales rep queue, sorted by score.
• 1–3 → automated nurture sequence.

The one-sentence reason is non-negotiable. It makes the score auditable and lets you tune the prompt when you see scores you disagree with. Without it, the model is a black box and you'll never trust it.

6. Follow-up sequencing that doesn't sound like a robot

Claude reads the last three messages in a thread and decides one of three actions: bump, value-add (share a resource), or break-up. Then drafts the email.

The founder approves a batch once per day. A three-person agency I built this for went from ~90 minutes/day on follow-ups to about 8 minutes. The drafts are good enough that maybe one in five needs an edit.

The trick is putting the agency's actual past follow-ups (the ones that worked) into the system prompt as examples. Generic "professional but friendly" tone instructions produce generic output. Five real examples produce on-brand output.

7. Model routing: Haiku for plumbing, Sonnet for thinking

You do not run every call through the most capable model. That's the single most common money leak I see.

• Haiku → classification, routing, short structured outputs, lead scoring, JSON extraction with clear schemas.
• Sonnet → drafting replies, summarizing complex threads, handling ambiguous edge cases, tool-use loops with reasoning.

def pick_model(task_type: str) -> str:
    if task_type in {"classify", "score", "extract_simple", "route"}:
        return "claude-haiku-4-5"
    return "claude-sonnet-4-5"

On a workload of ~12,000 API calls/month, splitting the routing this way takes the bill from roughly $60 to around $18. Quality on the outputs that matter is identical, because the calls that need reasoning still get reasoning.

8. Prompt caching: one parameter, 30–50% off the bill

If you have a long system prompt — tone of voice, product context, rules, examples — cache it. Anthropic charges a fraction of the normal rate for cached tokens.

client.messages.create(
    model="claude-sonnet-4-5",
    system=[
        {
            "type": "text",
            "text": LONG_SYSTEM_PROMPT,  # 4000+ tokens of context
            "cache_control": {"type": "ephemeral"},
        }
    ],
    messages=[{"role": "user", "content": user_msg}],
)

That's it. One parameter. On a high-volume workflow it shaves 30–50% off the bill because every subsequent call in the cache window reads from the cached copy at a deep discount. Most people never flip it on.

What this actually looks like running

On a live deployment I checked this morning: the 7 a.m. ops summary in Telegram, three drafted email replies waiting for approval, one invoice extracted clean from a scanned PDF (vendor, amount, due date), two calendar bookings closed overnight while the founder was asleep. Server logs show 240 API calls in the last 24 hours, total spend $0.41.

That is the entire operation: a founder who opens Telegram, taps a few buttons, and has their morning back.

Why bizflowai.io helps with this

These eight patterns are the backbone of what I deploy for bizflowai.io clients. Inbox triage with human-in-the-loop approval, invoice extraction into accounting tools, daily ops briefings, calendar-booking agents, and the model-routing plus prompt-caching work that keeps monthly costs under $20 for most small operations. I build the pipelines on the client's own infrastructure or my home server, hand over the code, and stay on for maintenance — no SaaS lock-in, no per-seat pricing.

Frequently asked questions

What is the gap between Claude the chatbot and Claude the worker?

It's the difference between using Claude Pro in a browser tab and actually integrating it into business workflows. Many solopreneurs and small agencies subscribe to Claude Pro but still spend two to four hours daily on email, hand-key invoices, and chase leads manually. Closing this gap requires wiring Claude into tools like Gmail, calendars, and databases using platforms like n8n or Python scripts, not just chatting with it.

How do I automate inbox triage with Claude?

Set up a script that pulls unread Gmail messages every two minutes and sends each subject and body to Claude with a classification prompt categorizing it as a lead, support request, vendor invoice, or noise. Real leads get drafted replies pushed to Telegram or Slack for one-tap approval, noise gets archived, and invoices route to extraction. On 120 messages daily, this costs about $14/month in API fees and saves roughly two hours.

How do I improve Claude's accuracy on invoice extraction?

Convert PDFs to high-resolution images before sending them to Claude's vision endpoint, set temperature to zero, and include an example of the exact JSON schema you want (vendor, date, line items, totals, tax, currency). This single change boosts accuracy on messy real-world invoices from around 70% to 94%. The structured output can then flow directly into Postgres or an accounting tool.

When should I use Claude Haiku vs Sonnet?

Use Haiku for classification, routing, and short structured outputs where speed and cost matter more than reasoning depth. Reserve Sonnet for calls that require real reasoning, such as drafting replies, summarizing complex email threads, or handling edge cases. On a workload of about 12,000 API calls per month, routing tasks this way significantly cuts costs compared to running every call through the most expensive model.

How does Claude book calendar appointments automatically?

Using tool use, you define functions like check_availability and book_slot that Claude can call during a conversation. When a lead emails asking for a demo, Claude reads the email, queries the calendar API, proposes three available slots, and once the lead picks one, books it and sends confirmations. The founder only sees the finished calendar invite appear, with no manual scheduling involved.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.