OIAS is a small B2B studio in Bengaluru. Three teams — Labs (AI and software), Studio (brand and design), Marketing (growth) — and one operating decision that shapes everything: the company itself is the first client. Before we automate anyone else's operations, we automated our own. What follows is the system as it runs today, written down because almost everyone publishing about "AI-run companies" is describing a demo. This one has a cron schedule.
The shape of it.
Five components, each boring on its own. The system is the interesting part.
- 01The site captures leads. A visitor submits the contact form; a serverless function validates it (with a honeypot for bots) and writes the lead into Zoho CRM. If the CRM is ever unreachable, the lead degrades to a logged fallback — no inquiry is ever lost to an outage.
- 02The CRM is the ledger. Zoho holds every lead and its status. One system of record. Agents read it, agents write it, the founder audits it. No shadow spreadsheets.
- 03An agent does outreach — daily, on a clock. Every morning at 09:00 IST a cron wakes, pulls queued prospects from the CRM, and for each one Claude writes a unique short email — researched, specific, under 120 words. No templates, no mail-merge blanks. Gmail sends it; the lead is marked contacted. Every prospect is contacted exactly once unless they reply.
- 04A dashboard shows the founder everything. A password-gated admin panel aggregates the pipeline: lead counts by status, latest inquiries, outreach stats. Numbers first, narrative separate.
- 05A daily CEO brief sets priorities. Each morning an agent reads the pipeline, the outreach results, and the backlog, then writes the founder a short brief: what happened, what's stuck, what matters today.
The guardrails are the design.
An AI company that sends email autonomously is one bad morning away from being a spam operation. So the guardrails aren't an afterthought — they're most of the engineering.
- ARamped sending. The outreach agent is capped at 5 emails a day in week one, then 10, 15, 20. Domain reputation is built the way trust is built — slowly, on purpose.
- BAbort on first failure. If any single send fails, the entire run stops. No "keep going and hope." A human looks at the logs before the machine gets another turn.
- CAuthenticated mail, properly. SPF, DKIM, and DMARC configured before the first cold email went out — the unglamorous work most senders skip and then wonder why they land in spam.
- DTwo classes of output. Outreach is full auto-send — that decision was made deliberately, with the ramp as its counterweight. Everything else customer-facing (site copy, replies to inbound, this article) is draft-and-approve: an agent writes, a human ships.
- ENo invented proof. Standing rule across every agent: no fake testimonials, no inflated numbers. Where we don't have proof yet, the copy says less. It costs us conversions today; it's the only version of this that survives contact with reality.
"The interesting question isn't whether AI can run a company. It's which decisions you still refuse to delegate."
What we learned building it.
The failure paths are the product. The contact form's most important feature isn't the happy path — it's that a CRM outage silently degrades to a log entry instead of a lost lead. The outreach engine's most important feature is the abort. If you're building agent systems, budget more time for "what happens when it breaks" than for the demo.
Rate limits will find you. Zoho throttles OAuth refresh-token exchanges; parallel agent calls each requesting fresh tokens will lock the whole org out of the API for minutes. One shared token cache fixed it. Every SaaS API your agents touch has a version of this landmine.
Uniqueness beats volume. Twenty researched, individually written emails a day outperform a thousand templated ones on every axis we care about — deliverability, reply quality, and not being the kind of company we'd refuse to work with.
The founder's job changes shape. With agents running the loops, the human work concentrates into three things: deciding what the machine may do alone, reviewing what it drafted, and choosing what to build next. That's the actual meaning of "AI-run" — not zero humans, but humans only where judgment lives.
Why we're telling you this.
Because it's the pitch. This stack — lead capture, CRM ledger, agent outreach, founder dashboard, daily brief — is what OIAS Labs builds for clients, including NGOs who need their operations to run this lean. We sell what we use. If you're deciding whether to buy, build, or skip this kind of system, we wrote an honest guide: what to buy, what to build, what to skip →
And if you'd rather learn this by doing it — we take interns. You'd be working inside the machine described above.
Tell us what's eating your week →
Architecture as described is live as of July 2026. When it changes, this page changes.