Most corporate AI failures don’t stem from the model being dumb. It fails because the model is blind — disconnected from how the company actually runs.
This episode is the executive companion to Production-Grade Agentic AI: why Ran wrote Company-Scale Agentic AI (free PDF/EPUB via varops.com/books; print on Amazon), and what “agentic” has to mean at org scale.
Blind, not stupid
Execs and engineers blame intelligence: weak model, unoptimized workflow. Sometimes that’s fair. More often the system has no access to real operations — WhatsApp decisions, phone-call judgment, the thing that only lives in Jane’s head after fifteen years.
Advanced people stuffing ChatGPT/Claude projects with policies and transcripts is a start. It still isn’t a company brain. It’s scattered context in favorite chat companions. And outside startups, “AI use” is often still: open ChatGPT, ask one question, move on.
The five-part loop
Ran’s frame (from the book):
- Observe — install a “ghost in the machine”: resident on Slack/Teams (not DMs unless invited), mail (BCC or server hooks), WhatsApp/iMessage groups, conference and phone calls (e.g. Plaud/Pocket-style capture), Drive. Absorb how work actually happens — company-wide, and eventually per team / per person.
- Understand — distill a knowledge graph of processes, relationships, and real bottlenecks (not the ones management thinks they have).
- Build — propose what to automate; report cracks and delays; humans or the system start shipping skills/workflows.
- Run — execute with a readiness dial. Start gated.
- Compound — every fix to a bad draft, every cancelled action, feeds the brain. Don’t patch outputs forever; fix root cause in the skill/process.
Automation vs agent
Don’t sprint to fully autonomous agents. Business work is mostly fixed-path automation with fuzzy islands:
- Deterministic: pull CRM history, account owner, products bought — API, not LLM roulette
- Fuzzy: interpret the request, draft the reply, decide refund edge cases — that’s where AI belongs
Throwing a ticket at an LLM with a pile of MCP tools and hoping is not a strategy.
Gate on purpose
In business, one wrong send can cost tens of thousands to millions. Start with human-at-the-gate: draft email, filled order awaiting review. The gate’s job isn’t “edit forever.” If two drafts miss company voice or policy, go fix the skill — and let the loop learn from cancels and corrections until trust earns more autonomy.
Role-based access (the hard part)
Org-wide AI dies when everyone connects as the CFO. You need middleware that adds/removes tools and data grain based on who is talking (Slack, email, chat) — percentages for sales forecasting, line items for finance, or nothing for HR. Ran’s pattern in MUXI: a filter between user and brain. Painful (WorkOS, Active Directory, etc.) — and non-negotiable at hundreds or thousands of employees.
Fit AI to where people already work
Don’t force a 20-year Outlook/Slack habit into a consultant’s shiny tool. Meet people in email, Teams, Slack; bridge “cool kids” tools to the rest. New-tab Chrome extension that refreshes a live brief (Slack threads, mail) — tasks only clear by doing the action or delegating the agent — is one shape of not letting work fall through the cracks.
Leave-with
Observe everywhere important decisions actually happen. Understand. Build. Run gated. Compound. Optional future episode: company-specific LLMs on top of that brain — out of scope here.
Free book when it’s up at varops.com/books
