Replicate every worker with AI.
Observe a human for one week.
Deploy an AI worker on day seven.
Demo
Watch Zoral replace a real worker.
One recording. Three phases. Full replication.
The last mile problem of AI automation
Every AI tool requires a million integrations. We took the opposite approach. Instead of bringing human tools to the AI, bring the AI to the human tools.
One API changes and the whole thing breaks.
Same computer. Same apps. Same email. Same browser.
If a human can use it, the agent can use it.
Replicate first. Optimize second.
Same I/O surface as the human
Once the agent perfectly replicates what the human was doing, optimizations apply on top. Why type in TextEdit when you can write a file with one terminal command? Why click through a UI when you can hit an API directly? Those AI vertical tools (your CRM, your analytics platform, your outreach tool) become part of the optimization layer. The agent uses them the same way a human power-user would.
From the outside, the system still looks like a human worker. Imagine being at work and having no idea whether your coworker is an AI or a human.
How it works
Same architecture. Three modes. The kernel predicts, observes, and learns.
Training Mode
Days 1-6The system shadows a human worker. On every event, the kernel predicts what it would do, watches what the human actually does, and learns from the delta.
Click to learn more
Supervised Mode
OptionalFor high-stakes roles (finance, legal, client-facing). The AI controls I/O but externally-visible actions queue for manager approval before execution.
Click to learn more
Autonomous Mode
Day 7+From the outside, nothing changes. The AI worker uses the same Slack, Teams, email, and tools your human employee used. It types on the same keyboard, clicks the same buttons, responds in the same channels. No extra setup. No integrations. It is the worker.
Click to learn more
Scroll down to explore the full architecture
Our rollout
Zoral is rolling out in two phases. Phase 1 is in beta now.
Phase 2 waitlist opens soon.
Phase 1: Inherit
BetaAI runs on the departing employee's actual laptop. Same email, browser, network, device.
All 9 anti-abuse checks pass by default because everything is real.
- Browser fingerprint
- IP reputation
- Email domain (MX/SPF/DKIM)
- Email verification flow
- Phone line type
- Phone reputation
- Card BIN
- Address AVS
- Behavioral monitoring
Phase 2: Create
Waitlist opening soonAI workers get their own full corporate identities, created from scratch.
- >M365/GWS email on customer tenant
- >Real eSIM (not VoIP)
- >Bank-issued cards
- >Cloud VM with matched browser fingerprint
- >VPN through office IPs
Anti-abuse check status
- Browser fingerprint
- IP reputation
- Email domain (MX/SPF/DKIM)
- Email verification flow
- Phone line type
- Phone reputation
- Card BIN
- Address AVS
- Behavioral monitoring
Consolidation
Your AI agent needs to sleep.
Every animal with a nervous system sleeps. Dolphins sleep with half their brain at a time so they don't drown. Evolution tried everything to eliminate sleep and failed every time. There's a reason for that.
NREM
Memory consolidation. Replay important memories, globally downscale weak connections, promote recurring patterns from episodic events (L2) to semantic knowledge (L3). Signal-to-noise cleanup.
REM
Cross-domain association. Find connections between completely unrelated memories. Detect gaps and inconsistencies in the agent's mental model. Promote stable patterns from semantic knowledge (L3) to procedural rules (L4).
Garbage Collection
Prune dead edges. Archive dormant nodes. Update behaviour.md (the agent's behavioral profile) and push it to the execution engine.
The OS equivalent? Defrag, cache eviction, cron jobs. Same operations, same reason: they can't run while the system is active.
“We keep building AI agents like software. Maybe we should build them like brains.”
Built on
All adaptive systems minimize surprise by maintaining generative models, predicting sensory input, and updating from prediction error.
Karl Friston, 2010 ↗Hebbian Learning and Associative Memory for LLM Agents. Validated 4-layer hierarchy with power-law decay and spreading activation.
Sparse Attention with a "lightning indexer" that determines which tokens deserve full attention before the attention mechanism runs.
Conditional memory via scalable lookup. Separates static knowledge retrieval (O(1)) from dynamic reasoning.
Spreading activation theory of semantic processing. Memory retrieval via activation propagating through associative links.
Allan Collins, Elizabeth Loftus, 1975 ↗Synaptic Homeostasis Hypothesis. Memory consolidation during sleep: replay, selective strengthening, global downscaling, pruning.
Giulio Tononi, Chiara Cirelli, 2014 ↗Continuous cross-platform screen capture + OCR + audio transcription. Local SQLite storage. 2-5 GB/day.
Open Source (MIT, Rust), 2024 ↗Power-law forgetting curves match behavioral data better than exponential decay.