"How to Transform a Company With AI" — Varick Agents

Why this is in the vault

The cleanest external articulation of RDCO's agent-deployer thesis, written as a sellable enterprise methodology — and it independently lands on two disciplines RDCO built this week (self-improving-agent feedback loops + slow/fast state separation). Borrowable frameworks for both RDCO client-scoping and deciding which RDCO-internal workflows to agent-ify. Founder shared cold 2026-05-26.

The playbook (paraphrased)

• Redesign, don't buy. Electricity/assembly-line analogy (same one SDG used today): productivity gains require rebuilding operations around the tech, not swapping the engine and changing nothing. Buying agentic SaaS seats / Copilot licenses rarely moves the needle — transformation is a structural change in people + processes. If the AI doesn't understand the underlying process (and the process-owners aren't brought along), value + adoption both fail.
• Map first. Weeks embedded with teams (AP, procurement, sales, ops); map every workflow end-to-end; compute per-workflow agent ROI; capture tribal knowledge → rules/instructions/decision-logic the agents follow.
• 3-bucket split: deterministic work → scripted automation; judgment work → agents where appropriate; high-risk/high-judgment → stays human. Goal is topline growth + efficiency, not just cost-cutting (better context → faster/better human decisions).
• Workflow-selection — 4 traits: (1) happens often (100s–1000s/mo or touches real $); (2) repeatable decisions (patterns, business rules, exception-routing); (3) context spread across systems (the more humans tab between tools, the higher the agent value); (4) measurable pain (cycle time, error rate, manual hours — before/after).
• Self-improving agents: human-in-the-loop from day one — sandbox → shadow mode → supervised production; log agent output + human correction + context so the system improves (claimed +10% accuracy in weeks → more autonomy).
• Don't disrupt: build on top of existing systems (Salesforce/NetSuite) via APIs or computer-use agents — no rip-and-replace migrations. Keep data in 4 segmented layers: system-of-record / business-rules / raw-intake / agent-memory — so ops can change a rule without an engineer.
• Claimed case study: enterprise-SW sales transformation, large deals crossing 6 teams / 11 handoffs → "$25m in value in the first year."

⚠️ Bias

Lead-gen sales artifact for a $1B+-enterprise transformation consultancy (+ hiring). New account (2 tweets, 1.8k followers) but the article is circulating (857 bookmarks). Treat the dollar figures ("nine-figure," "$25m first year") as case-study marketing, not audited. The methodology is concrete and credible; the numbers are sales.

Mapping against Ray Data Co

• It's the enterprise-scale statement of our agent-deployer thesis — and a competitor/comp data point: Varick targets $1B+ revenue clients with embedded multi-week transformations; RDCO's wedge is smaller / data-teams / the "fractional forward-deployed engineer" frame (see the new curiosity questions). Same playbook, different altitude — useful proof the category is forming with money behind it.
• Self-improving-agents-with-feedback-logging = our validation-gate work. Their "log output + human correction + context → improve" is the same loop as the SkillOpt-derived /improve validation gate greenlit today. Outside confirmation the approach is right.
• 4-layer data segmentation (record / rules / intake / memory) = our slow-vs-fast-state discipline. "Ops can update a rule without an engineer; rules stay separate from memory" is the SkillOpt protected-section invariant + RDCO's CLAUDE.md-hard-rules-vs-working-context split, expressed for enterprise ops.
• Borrowable artifact: the 4-trait workflow-selection filter + 3-bucket (deterministic/agent/human) split is a clean scoping tool — for any future RDCO deployer engagement AND for auditing which of RDCO's own ops should be agent-ified vs left manual.
• "Build on top, don't migrate" matches RDCO's existing posture (computer-use + API over rip-replace).

Related

• [[2026-05-26-seattle-data-guy-ai-consultants-utility-thesis]] — same-day, same electricity analogy, agent-deployer corroboration
• [[2026-05-26-skillopt-self-evolving-agent-skills]] / [[../01-projects/skill-improvements/2026-05-26-improve-validation-gate]] — the self-improving-agent loop, internal version
• [[2026-05-25-joumana-elomar-brand-engineering-legibility]] — adjacent "the methodology is real, the framing is sales" pattern
• [[../00-vault-concepts/agent-deployer-positioning]]