BizOS
Rule-based AI orchestration for automating small business operations
AI Brain for small businesses. A rules-driven multi-agent system: the owner writes business rules once in simple files, and AI staff (support, sales, accounting, HR onboarding) follow them. One inbox for every channel, hard money/safety limits enforced in code not by the AI. Design doc is ready; we build the working demo tomorrow. 5. Agent Specifications5.1 OrchestratorLoads company.rules.md only. Routes; never answers domain questions.Classification contract: must output the verdict JSON above; confidence < 0.7 → asks ONEclarifying question on the same channel; still unclear → human_now .Rules editing is a separate, privileged path (v1.1): it can only be started from an authenticatedowner session, and the customer-message pipeline has zero write access to rules. Flow: ownerasks → a proposed diff is shown → owner approves → only then the version bumps and publishes.An agent that reads untrusted customer input must never be the agent that can rewrite companypolicy — this is the prompt-injection firewall.5.2 Specialist agents (accounting / sales / support / media)System prompt = company.rules.md + own rules file + few-shot examples tagged with therulesversion they were approved under — examples expire automatically on version bump, so_stale examples can never override new policy. (Context isolation: the sales agent physically cannotleak supplier costs it never sees.)Tool access is per-domain (accounting → invoice API read; media → post scheduler; support →order lookup).Every answer must cite rule IDs it relied on. No matching rule → answer is held as draft + flagged"rules gap" so the owner can add the missing rule. The rulebook grows from real traffic.5.3 Web Agent (public-facing)Separate agent, loads only web.rules.md — a hand-curated public-safe subset (prices, hours,FAQ). Hard rule: if it's not in the file, say "let me connect you" and open a ticket through the sameenvelope.This is the security boundary: the internet-facing brain holds zero internal knowledge. And it worksboth ways (v1.1): when an internal ticket's result flows back to the web channel, it passes the sameweb.rules.md filter on the way out — a clean first prompt alone is not the boundary.5.4 HR / Academy Agent (employee onboarding & Q&A)Same rules files, opposite audience: answers employees, not customers.New hire asks "how do refunds work?" → answers from support.rules.md +accounting.rules.md , citing rule IDs → the handbook that's never out of date, because it IS thelive rulebook.Onboarding mode: generates a day-1 checklist and a 10-question quiz from the rules files; logscompletion.Flywheel: when an employee asks something no rule covers → "rules gap" ticket → owner adds rule→ next hire never asks again.6. Human-in-the-Loop & TrustTier What happens Exampleauto_reply Agent answers, logs "What are your hours?"draft_for_approval Reply queued, owner taps ✓/✗ on phone Refund $120human_now Push notification, agent holds "My lawyer will call"Owner's daily surface = one approval queue + one audit feed. The promise is not "AI replaces you"; it's"you only see exceptions."Approval SLA (v1.1): every held item has a priority and a timeout. No owner response in 2h → thecustomer gets a polite hold message and the reminder escalates (push → SMS); critical items ring thephone. An unanswered queue never silently drops a customer.Honest expectation: in week one, rule coverage is low, so the queue is busy — but approving a readydraft is ~10x faster than composing from scratch, and every rejection writes a rule, so the queue shrinksweekly. Success metrics: owner-minutes per day, time-to-resolution, error/reopen rate — not just theauto-reply percentage.7. Hackathon MVP (buildable in one day)Cut to the bone — demo path:1. /api/inbox webhook (one endpoint, envelope above) — 2 hrs, pattern already exists in mynotification service2. Orchestrator classify via Claude → route JSON — 1 hr3. TWO specialists only: Sales + Support, each with a real rules file written live on stage — 2 hrs4. Academy agent answering "I'm the new hire, how do refunds work?" from the same files — 1 hr,pure prompt5. Simple inbox UI (teammate) + approval queue — parallel
⏺ Read(summary.md)
BizOS lets an owner write their business rules once in simple text files, then has AI helpers for customer support, sales, accounting and new-hire training follow those rules. Messages from every channel land in one inbox, and anything risky, like a $120 refund, waits for the owner to approve on their phone.
Best for: Small businesses of any kind where the owner personally handles customer messages, approvals and staff questions and wants to only deal with the exceptions.
Summary by Claude from the team's submission and README. Find more free AI tools for your business →