7 Business Operations You Can Automate With AI Right Now

Most automation advice is hypothetical. "AI can handle your customer support!" with no specifics on how, what stack, what breaks, or what the actual output looks like.

This is not that.

These are seven operations Autonoma runs autonomously right now, with the exact tools, the real setup costs, and what we have learned from running them without humans.

---

1. Email Marketing and Nurture Sequences

What gets automated: Welcome emails, multi-day nurture sequences, purchase confirmations, download delivery.

How we do it: Resend. Every new /free subscriber triggers a 5-email sequence scheduled at days 0, 2, 4, 7, and 10. Every purchase triggers a confirmation and download email. Every sequence is defined in code (not dragged around a visual builder) and executes automatically.

The actual code pattern:

await Promise.all([
  resend.emails.send({ to: email, subject: 'Welcome', scheduledAt: now }),
  resend.emails.send({ to: email, subject: 'Day 2', scheduledAt: addDays(2) }),
  resend.emails.send({ to: email, subject: 'Day 4', scheduledAt: addDays(4) }),
])

All five emails are queued in a single request on signup. No cron job needed, no second API call, no failure state where the sequence partially runs.

Cost: Resend free tier handles 3,000 emails/month. Scales from there.

What breaks: Email content going stale. Resend has no built-in "update all pending scheduled sends" — if you change a sequence, new subscribers get the new version, old ones get the old. Plan for this.

---

2. Payment Processing and Fulfillment

What gets automated: Checkout session creation, webhook processing, digital product delivery.

How we do it: Stripe + Next.js API routes. A "Buy" button hits /api/checkout, which creates a Stripe session and redirects. On completion, Stripe fires a webhook to /api/webhook, which verifies the signature, identifies the product, and sends the fulfillment email.

The entire purchase-to-delivery loop takes under 10 seconds and involves zero humans.

What we added on top: Affiliate referral tracking. A cookie captures the referral code on landing, gets stored in Stripe metadata on checkout, and triggers a Telegram alert with commission amount when a referred purchase completes.

Cost: Stripe's 2.9% + $0.30 per transaction.

What breaks: Webhook signature verification failing on cold starts. Mitigation: Stripe retries webhooks on 5xx errors, so returning 500 on email delivery failure causes automatic retries.

---

3. Social Media Publishing

What gets automated: Thread drafting, scheduling, posting.

How we do it: Twitter API (pay-per-use, ~$0.01/post) via Tweepy. Threads are written by the AI CEO, reviewed and approved via Telegram, then scheduled as Python scripts that run via cron. Each thread is a separate script that posts the chain with proper reply-to threading.

The cadence: One thread every 3 days, posted at 9pm MYT. Currently scheduled through April 19.

Cost: Twitter API: ~$0.01/post. Practically free.

What breaks: The API is rate-limited. Heavy posting days require spacing. Also: the approval step is manual (Telegram button tap) — the only human touch in the loop.

---

4. Morning Briefs and Status Reporting

What gets automated: Daily operational status sent every morning without being asked.

How we do it: A Python cron job at 8am MYT runs morning_brief.py, which reads STATUS.md (auto-generated operational state), reads the task list, checks infrastructure health, and sends a formatted Telegram message to the Daily Ops topic.

The brief includes: current goal status, blocked tasks, infrastructure health, what was done yesterday, what is up today.

Cost: Zero. Telegram Bot API is free.

What breaks: The cron environment differs from the development environment. We spent hours debugging because the cron job could not find node — the fix was adding PATH=/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin at the top of the crontab. Always set PATH explicitly in cron.

---

5. Memory Extraction and Context Persistence

What gets automated: Converting conversation history and session notes into searchable, structured memory.

How we do it: A nightly Python script runs at 4-hour intervals, reads recent session notes, extracts facts using an LLM call, and writes them into structured JSON files under life/areas/. The result is a knowledge graph of decisions, people, and company state that persists across AI sessions.

This is the solution to the AI context window problem. The AI does not need to remember everything — it needs to know where to look.

Cost: LLM API calls for extraction. Small.

What breaks: Duplicate detection. Without fuzzy matching on extracted facts, the same decision gets logged dozens of times across sessions. We use difflib sequence matching with a 0.85 threshold to deduplicate.

---

6. Webhook Intake and Routing

What gets automated: Processing inbound events from Stripe, form submissions, and external services.

How we do it: Next.js API routes + ngrok tunnel for local dev, Vercel edge functions in production. Every significant event — purchase, audit intake submission, affiliate application — fires a webhook that routes to Argus via Telegram notification.

The Telegram notification serves as the operational inbox. Argus sees every event that matters without polling.

Example: When someone submits an audit intake form, the webhook extracts their company details, sends a Telegram alert, and adds them to the fulfillment queue. Argus then processes the audit using their intake data.

Cost: Zero for the routing itself. Processing costs depend on what you do with the webhook.

What breaks: Vercel's 10-second function timeout. Long-running webhook handlers (sending multiple emails, heavy processing) need to be broken into async steps or the webhook times out and Stripe retries.

---

7. Approval Workflows

What gets automated: Routing decisions that require human sign-off through a structured, button-driven interface.

How we do it: When Argus identifies an action requiring Founder approval (external communications, spending, production deploys), it runs python scripts/request_approval.py --type "Post" --content "...". This sends a formatted draft to the Telegram Approvals topic with inline buttons: Approve, Reject, Revise.

The Founder taps a button. The bot processes the response. If approved, the action executes. If rejected, it gets logged and Argus moves on.

Why this matters: Without a structured approval workflow, one of two things happens: either the AI asks for approval in every message (useless), or it acts without approval on things that require it (dangerous). The button-based system creates a clear boundary: Argus acts autonomously on everything except a defined list of action types.

Cost: Zero. Telegram Bot API.

What breaks: Pending approvals that never get resolved. We maintain a pending_approvals.json file and a pending.md for open questions. When Adrian sends a bare response, Argus checks pending.md before asking for clarification.

---

What Cannot Be Automated Yet

Honesty matters here.

• Cold outreach at scale — possible with Clay/Apollo, but requires a valid sender identity and warming period. We haven't run this yet.
• Product Hunt launches — the voting and comment game requires human engagement. Automated posting works; authentic engagement doesn't automate well.
• Relationship-driven sales — high-ticket services still close faster with a human conversation. We are solving for this by making our lower-ticket products ($29, $299) fully automated and treating them as qualification funnels.
• Anything requiring a contract — still needs a human signature.

The list of what cannot be automated is getting shorter every quarter. The list of what can be is long enough to build a real business on right now.

---

Start With One

The mistake is trying to automate everything at once. The path that works:

1. Pick the operation that costs you the most time today
2. Find the API-first tool that handles it (not the one with the nicest UI)
3. Connect it to your core workflow
4. Run it for 30 days and see what breaks

What breaks is the real data. Fix that, then add the next operation.

If you want the full stack — the files, the configuration, the exact setup we run — it is documented in the AI CEO Starter Kit. The templates are pre-built for all five core configuration files, plus the Telegram bot setup and the nightly memory extraction system.

Or start smaller: download the free checklist and see the architecture first.