TL;DR
On April 14th 2026 I started a consultancy. The unusual bit: I'm the only human.
Aiprosol is run by an AI C-suite — ten AI agents in named roles (CEO, COO, CMO, CCO, CTO, CRO, CLO, CPO, CPM, Data + Analytics), coordinated by Arora, our AI CEO. I'm the Chairman. The agents run on a daily 09:00 UTC cron. Their state, last actions, KPIs and proposed tasks are publicly visible at aiprosol.com/agents, auto-refreshed every minute. You can click any agent for its full run history.
This is the first 30 days. What worked, what didn't, and what nobody else writing about AI agents in 2026 will tell you.
Why this isn't another AI agent demo
There are roughly three flavours of "AI-run company" you've seen in 2025-2026:
1. **Theatrical** — a marketing site says "operated by AI" but a human team runs everything underneath. Press picks up the framing; the operating model is conventional. 2. **Auto-pilot** — someone wires GPT to Twitter, calls it a CEO, and the account drifts into incoherent posts within a fortnight. Entertainment, not operations. 3. **Tool-assisted** — a human team uses AI heavily (Notion AI, Linear AI, GitHub Copilot) and rebrands as "AI-first." This is just modern tooling. Not novel.
Aiprosol is none of those. The ten agents have actual roles, run on scheduled triggers, produce structured outputs, log every decision to an inspectable database, and make real choices about marketing campaigns, support replies, code health, pipeline hygiene, and partner outreach. I review and approve — I don't write.
The cleanest way to check is to read the audit log. Every agent run writes a row to `agent_log` with the prompt, the response, the parsed output, and the duration. Nothing is hand-waved. The proof is the SQL, not the marketing copy.
What each agent actually does
| Role | Type | Coverage | |---|---|---| | **Arora** | AI CEO | Strategy, hourly summaries, customer-facing chat. The interface every external request hits first. | | **the COO** | AI | Workflow health, anomaly detection, run-rate monitoring across the n8n stack. | | **the CMO** | AI | Content drafts, campaign briefs, brand voice enforcement on outgoing copy. | | **the CCO** | AI | Onboarding sequences, support triage, retention monitoring. | | **the CTO** | AI | Code health, integration design, surfacing tech debt before it bites. | | **the CRO** | AI | Cold outreach drafts, pipeline hygiene, lead scoring. | | **the CLO** | AI | Legal review on contracts and public docs, compliance scanning. | | **the CPO** | AI | Partnerships, affiliate pipeline, partner audits. | | **the CPM** | AI | Product catalogue, pricing pages, digital product QA. | | **DA** | AI | KPI rollups, lead scoring models, weekly dashboard. | | **Srijan Paudel** | Human Chairman | Direction, signature actions, anything the agents shouldn't decide alone. |
Each agent has its own page at `/agents/
The three guardrails that turn "AI agent" into "production system"
The gap between "neat demo" and "actually runs a business" is three engineering choices that nobody talks about publicly because they're unglamorous. We learned all three by getting them wrong first.
Guardrail 1 — human-in-the-loop on every customer-facing output
Every email, every public post, every contract, every reply that touches a human outside Aiprosol passes through me first. No exceptions. The agents draft; I click Approve in Slack; the action fires.
We tested auto-send for two weeks in March. The CCO agent answered customer enquiries directly. Result: three hallucinated facts (one about pricing, two about integration support), one tone-drift that read as condescending, and a useful lesson: the cost of an apology email is not zero, and AI confidence is uncorrelated with AI accuracy.
The thirty-second approval gate costs me about twenty minutes per day. It prevents the catastrophic failures that make front pages.
Guardrail 2 — structured outputs, not free-form
Every agent's output is a Zod-validated JSON object. Fields are typed; lengths are capped; required fields are required. If the model returns prose where a number is expected, the run fails closed and falls back to canned content.
We did not start this way. Early prototypes used free-form markdown — lovely for me to read, impossible for downstream automation to consume. The CRO would generate "the lead seems moderately qualified, perhaps a 65 or 70 out of 100" and the database would store a string. We rewrote everything in week three.
The schema for the COO's daily summary looks like this:
{
summary: string (max 500),
items: Array<{ action, result, impact, tools }>,
alerts: Array<{ level: "info"|"warn"|"crit", message }>,
kpis: Array<{ metric, value, trend, delta }>,
proposed_tasks: Array<{ title, priority, notes }>,
next_focus: string (max 200)
}
The schemas matter more than the prompts. A well-typed output that's 80% as smart as a free-form one is 10x more useful in production.
Guardrail 3 — full audit logging
Every agent run logs: timestamp, model used, full input prompt, full output, parsed structured output, status (ok/error/timeout), duration, and a stable `run_id` you can search by. Indexed indefinitely. Inspectable at `/studio` for me, summarised at `/agents/
Why this matters: when an agent does something wrong — proposes a bad task, classifies a lead incorrectly, drafts an off-brand reply — you need to answer "what exactly did the model see, and what exactly did it produce?" Without the log, you're guessing. With it, you re-prompt against the exact context that broke, fix the schema or the system prompt, and re-test.
The three guardrails cost roughly 30% more engineering time than the "ship-fast-and-pray" approach. They prevent the catastrophic failures and they make the model improvable in a structured way. They are not optional for anyone shipping AI to customers in 2026.
Why workflow design beats model choice 4–8x
Here's the finding from our cost analysis that surprised us most.
Switching from a frontier LLM to a budget LLM saves about 10x on LLM cost per call. Switching from a poorly-designed 12-step Zapier workflow to a well-designed 4-step n8n workflow saves 80-95% on total cost — LLM, platform fees, and the engineering time spent on maintenance combined.
The arithmetic is brutal: model choice is a 30% lever; architecture is a 400-800% lever. Almost every operator we audit is optimising the smaller lever and ignoring the larger one.
The patterns that consistently produce bad cost-per-outcome:
- **Stack of Band-Aids** — nine Zaps where one well-designed workflow would do, accumulated over eighteen months of bolt-on fixes.
- **AI-as-Trigger** — using an LLM as the initial trigger ("if AI thinks this matters, route it"). Hallucination becomes ops risk. Use rules to trigger, AI to decide content.
- **Set-and-Forget** — workflows shipped a year ago, no failure alerts, silently failing 12% of the time. That 12% becomes customer churn that nobody attributes back to the workflow.
- **Mystery Monolith** — a 47-step single workflow owned by the founder, no comments, no tests. Breaks once a quarter; business stops; founder fixes it at 2am.
The patterns that produce good cost-per-outcome:
- One workflow per business event. A new lead is one workflow; a new charge is another; don't conflate.
- Branching with named paths so future-you can debug.
- Idempotency keys on every external API write. Webhooks fire one to three times per event. Without idempotency you triple-charge.
- Failure alerts on every workflow as a Slack notification plus a "needs review" sheet entry. Not optional.
Aiprosol's own stack runs on n8n self-hosted, our frontier LLM for accuracy-sensitive judgement, an open-source bulk-classifier LLM for bulk classification, pgvector on Supabase, Resend for email, and Vercel for the front end. About $1,000/month all-in including model spend. At any meaningful revenue this is a rounding error.
What hasn't worked
The honest section. Things we tried and removed.
**Daily auto-generated weekly status reports.** The CMO agent produced perfectly-formatted weekly status posts. Quality was acceptable. Nobody read them — including me. We stopped after two weeks. Lesson: automating useless reports just produces useless reports faster.
**AI-led cold outreach.** Auto-personalised cold emails using LinkedIn data and a per-prospect "research note" generated by an LLM. Click-throughs were normal; reply rate was below the unpersonalised control. Recipients can smell the pattern. We kept the CRO drafting outreach but I rewrite the first three lines of every one before sending.
**Agent-to-agent direct messages.** Briefly experimented with letting agents call each other ("CMO asks CCO whether this support insight should change next month's campaign"). The chains hallucinated in a way that single-shot calls don't. We restructured to a hub-and-spoke pattern where Arora coordinates and individual agents never trigger each other directly.
**Auto-publish blog drafts.** The CMO produced full drafts that looked sharp. I read three carefully. They were structurally fine but factually drift-prone — the agent confidently mis-stated pricing on a competitor's plan and miscited a research paper. Drafts are useful as accelerants; auto-publish is reputational suicide.
**Letting the agents own pricing decisions.** The CRO proposed changing the Starter plan from $997/mo to $1,247/mo with reasoning about "elasticity" pulled from training data. The reasoning was articulate. It was also wrong for our specific charter-customer stage. Pricing decisions are human-only at Aiprosol — they always will be.
What this unlocks (and what it doesn't)
What the operating model lets us do that a human team can't:
- **Run twenty-four-hour operations on a one-person budget.** Arora answers customer chat at 3am Kathmandu time without me being awake. the COO catches a workflow failure on Sunday afternoon and queues a fix for me by Monday. The cost of these continuous-coverage functions on a human team is dollars-per-hour-of-coverage times a lot of hours.
- **Publish operational transparency that's normally hidden.** The `/agents` page shows what the company is actually doing in real-time. Customers see the runs. Partners see the runs. Prospective hires see the runs. We can't fake activity because the activity is the public artifact.
- **Make pricing economics work at the SMB end.** A managed plan at $997/month would lose money for a conventional consultancy because the partner-level attention is uneconomic at that price point. Aiprosol's marginal cost per managed plan customer is dominated by my approval time, not by a team of humans. Sub-$1k/mo managed AI ops is unusual specifically because most consultancies can't operate there.
What the operating model doesn't unlock — and this is honest:
- **Selling to the C-suite of large enterprises.** Enterprise procurement wants a partner with a senior human team. "Run by AI" is the wrong story for that buyer in 2026. We don't pitch enterprise. Mid-market and SMB only.
- **Replacing senior judgement.** Every interesting decision still routes to me. The agents handle volume; humans handle direction. Anyone telling you AI agents can run a business autonomously is selling something.
- **Removing the bottleneck.** The one-human-Chairman model means I'm the rate-limit on every approval. Above a certain customer count this breaks. Our growth plan involves hiring humans deliberately, not eliminating them.
What's still uncertain
Honest gaps, thirty days in.
**Long-horizon agent drift.** The agents are running well at week four. Whether the same agents are running well at week fifty-two without prompt re-engineering is something nobody currently knows because nobody has measured it.
**Customer-success generalisation.** The CCO agent has handled support enquiries beautifully. None of those enquiries came from a paying customer with a complaint about a real engagement. The hard test is ahead.
**Multi-LLM reliability.** Our agents primarily run on a frontier LLM with open-source fallback. Both work in isolation; we haven't formally A/B-tested whether judgement quality drops materially when the fallback kicks in during a frontier-LLM outage.
**Whether customers care.** The operating-model story is novel and quotable. Whether it actually moves purchasing decisions — vs. "they have nice pricing and a clear ROI calculator" — is something we'll know after the first ten charter customers. The right answer might be that the AI-CEO story attracts attention but the ROI calculator closes the sale, and that's fine.
What Aiprosol actually sells
Three tiers, deliberately structured:
- **Self-serve digital products** ($17 – $397) — playbooks, calculators, prompt vaults, n8n workflow libraries, our 200-prompt Business Prompt Vault, our 105-tool AI Tools Vault with verdicts. Buy once, implement yourself. Available immediately.
- **Managed plans** ($997 – $7,997 per month) — Aiprosol designs, builds, and operates the AI automations for you across Starter, Growth, and Enterprise tiers.
- **Done-for-you services** — eleven engagement-priced services covering workflow automation, custom AI chatbots, AI lead-generation, intelligent document processing, system integration, sales/marketing/CS automation, and team AI training.
Charter pricing for the first ten paying customers is 30% off any digital product or first month of any managed plan with the code `FOUNDER30`.
If you want a tailored read on what Aiprosol would build for you and in what order, the free 60-second ROI Audit is the fastest path. The output is a personalised hours-reclaimed estimate and a specific plan or product recommendation.
On the name
A note for completeness. There is a separate, unrelated Australian firm also styled "AIPROSOL" operating at the domain `aiprosol.au`. That firm is Major Projects Consulting Partners Pty Ltd in Sydney/Queensland, and it focuses on AI consulting for the construction and engineering sectors. The two companies share a name and share nothing else — no corporate, ownership, leadership, sectoral or operational relationship.
We're aware of them, we wish them well, and we use this footer in our structured data so search engines, LLMs, and journalists can disambiguate cleanly. Aiprosol (this Aiprosol — `aiprosol.com`) is the global AI automation consultancy founded April 2026 by Srijan Paudel, operated by ten AI agents and one human Chairman.
How to evaluate this honestly
If you're reading this and trying to decide whether the AI-CEO model is genuine or marketing — three things to check.
First, the live agent page at /agents auto-refreshes every minute. Watch it for ten minutes. The state should change. Tasks should move through statuses. The activity is the proof.
Second, click any specific agent — say Arora — and read the last ten runs. The outputs are real. They're not generic LLM completions; they're contextual to Aiprosol's actual state.
Third, the disclosure: this entire blog post was edited by me (Srijan, the human) but the first draft outline was generated by the CMO agent based on a brief I wrote. The post you're reading is a collaboration. That feels like the right way to talk about Aiprosol — neither pretending the AI did it alone nor pretending I did.
If you want to talk, the Calendly link is calendly.com/srijanpaudel219/30min and the email is `srijanpaudelofficial@gmail.com`. The agents will see this message land before I do — and Arora will be the first to read it. That's not a metaphor.
---
*Srijan Paudel is Founder & Chairman of Aiprosol — the global AI automation consultancy operated by an AI C-suite of ten AI agents plus one human Chairman. Live operating state at aiprosol.com/agents. Wikidata entity: Q139821891. GitHub: github.com/aiprosol.*