Case Study · June 2026 · Holding Company × AI Executive Operator
How AmA Invest's CEO moved past his DIY proof of concept — and built an autonomous operating layer across 25 businesses.
25 businesses. 3 industries. 1 founder. A case study on how a technically sophisticated operator outgrew his own PoC — and chose WORKERON.ai to build the execution layer his holding company actually needed.
The before-state
I want to send a short task in Telegram so as not to open separate SaaS tools and work panels to start a process. I want to get a finished result — not general advice. The task should end with a file, a link, a report, a draft message, or a created workflow.
What Arkadiy tried before WORKERON.ai:
- Openclaw.ai PoC on AWS — a custom Telegram bot connected to ChatGPT, built in-house
- Perplexity for research (standalone — not integrated into any execution flow)
- Google Workspace for email and docs (manual, no agent access)
- ClickUp for task tracking (not connected to the AI layer)
- Human assistants for coordination across business units
Why the PoC wasn't enough
I want to store passwords, tokens, and API keys in a protected Vault so the agent can work with services without revealing secrets in chat or logs.
The PoC proved the concept but exposed the gaps. Security ran on system prompts alone. There was no long-term memory between sessions. The agent sometimes failed to use available tools mid-task. Running 25 businesses through the same context window — with no isolation between them — was a risk he wasn't willing to take at scale.
Day-to-day · 4 WORKERON use cases
Telegram Command Center
"I want to send a short task in Telegram and get a finished result — a file, a link, a report — without opening ClickUp, Perplexity, and Gmail separately."
Replaced: opening 5+ disconnected tools to start any task. Now: one Telegram message → Prompt Engine enriches the command → best-fit model executes → result delivered.
Board Council — multi-model stress test
"I want investment and marketing decisions analyzed from multiple perspectives — not just one model's take. Like having a critic, an investor, and a chairman in the same room."
Replaced: founder reviewing strategic decisions solo. Now: a multi-model ensemble runs structured debate across roles — Chairman (deep reasoning), Critic, Financial Investor — and returns a consensus with a Verifier check on all sources.
Granola Mode — meeting oversight
"I can't attend every meeting across 25 businesses. I need to know if a task got lost between departments — without being the one to chase it."
Replaced: post-meeting follow-up calls and manual transcript review. Now: agent monitors meetings he doesn't attend, analyzes the transcript, and flags tasks that risk falling through the cracks.
Autonomous contractor identity
"I want to confirm risky actions before they execute — the agent should not send a message, make a call, or publish content without my explicit control."
Replaced: founder as the communication bottleneck between contractors and business units. Now: agent operates via its own ProtonMail and WhatsApp Business number as "Ark's Agent" — with human-in-the-loop confirmation for all outbound actions.
The PoC-to-production difference
| Need | DIY PoC (Openclaw.ai + tools) | WORKERON.ai |
|---|---|---|
| Security | System prompt only | Vault + AI Firewall + DLP |
| Memory across businesses | Resets each session | Namespaced long-term memory |
| Task execution | Text output only | Playwright, codex-cli, Lovable |
| Context switching (25 domains) | Manual, cognitive lag | Expertise Layers per domain |
| Maintenance | Founder builds and fixes it | WORKERON team owns it |
The vision · Operating system for a holding company
This is not an external SaaS product. It's an internal R&D layer that must prove performance fast and become an operational amplifier — one that handles the cognitive switching I physically can't do across 25 industries at once.
The architecture being built: each business domain gets its own isolated agent with a separate memory namespace — construction, cannabis, and financial services in strict separation. Arkadiy's personal agent acts as the orchestrator, routing queries to domain specialists and returning a unified answer. Specialized modules (HealthOS for medical analysis, LinkedIn Copilot for brand growth, a security Vault for credential management) layer on top. The Web Admin Dashboard gives a live visual audit trail of every agent action, token cost, and workflow status — replacing manual status checks.
Honest gaps
- Trust threshold for autonomous action not fully set.
"I want to confirm risky actions before they execute — the agent should not send a message, make a call, or publish content without my control." The approval-flow guardrails exist, but the exact boundary between what the agent can do freely versus what needs a confirmation is still being calibrated.
- Data isolation is architecture — not yet battle-tested.
The namespace separation between 25 businesses is built in by design. But running all of them through a single operator for the first time means edge cases around cross-domain context contamination are still being discovered and patched in real use.
- Tool-routing reliability carried over from the PoC.
The original system "sometimes failed to use available tools like Perplexity." The new architecture addresses this at a structural level, but reliable tool-selection across a multi-model matrix — at the pace and volume of 25 businesses — is still being refined in production.