Applied AI Tooling Roadmap

The stack, organized by the three pillars, split into Current (what we run today — confirm with the team) and Adding (proposed, with the reason and the synergy). Costs for every line live in EFFEN_Cost_Model.xlsx; this doc is the why, not the numbers.

Guiding rule from Applied AI Context & Operating Model: value ≈ AI × min(data, org). Tooling only pays off when it raises a substrate the AI can actually stand on. So every “Adding” below names the substrate rung it lifts and the workflow it unlocks — not tools for their own sake.

Two design principles the whole stack obeys:

  • Everything Claude can reach. A tool earns its place partly by being connectable — a first-party Claude connector or an MCP server. A tool the agent can’t read or act through is a silo, and silos cap min(data, org).
  • Managed for leverage, self-hosted for the flat line. We pay for managed where reliability matters (Workspace, MotherDuck, Claude) and self-host the data plumbing on a VPS to keep Track-B cost flat as brands scale.

Current stack — confirm this with the team

My understanding of what runs today. Correct anything wrong; the “keep / replace” column is the proposal.

CategoryTool todayPillarKeep / replace
Team coordinationWhatsAppOrgReplace — caps org maturity, invisible to Claude
Docs / filesGoogle Drive (consumer)OrgConsolidate on Google Workspace
EmailPersonal Gmail (per person)OrgReplace — move to managed Workspace identity
Project / task tracking(confirm — likely none / in WhatsApp)OrgAdd Notion
Paid adsMeta Ads, Google Ads, TikTok AdsData/ProductKeep + wire managed access
Web / analyticsGTM, GA4, landing pagesDataKeep
Ecommerce / OMSWooCommerce (primary), Fighter.my (1 brand)DataKeep
MarketplacesShopee, TikTok ShopDataKeep + request dev IDs
DesignCanvaProductKeep + connect via MCP
Video editingCapCutProductKeep
Ad research / ideationForeplay (rolling out, 2 seats)ProductKeep
Organic schedulingZernio (rolling out, 2 seats)ProductKeep
AIClaude (individual), MCP experimentsProductUpgrade to Team
ReportingManual exports / direct Claude + MCPData/ProductKeep now → warehouse later

Confirmed: Google Drive for docs/files; email is personal Gmail (not managed); CapCut for video editing. Still to confirm: whether any task tool exists today, and whether Foreplay/Zernio are already live or about to be.


Pillar 1 — Organization Maturity (Track A)

Governs how the org operates and what AI can plug into and act through. Today it’s the weakest substrate: work lives in WhatsApp, which has no structure, no search, no persistence, and no Claude connector. Lifting this rung is the prerequisite for the org-context slice of AI (CS/CRM, institutional memory).

Adding:

ToolRung it liftsWhy
Google WorkspaceTooling adoptionOne first-party home for files, email, docs, sheets — and Chat + Meet to replace WhatsApp’s real-time role without a second vendor.
NotionProcess captureStructured, searchable home for SOPs, calendars, approval logs — the rung an agent can actually take over. Chosen over ClickUp for the stronger Claude connector.
Claude TeamAI literacy + governanceThe licensing + governance substrate: connector scoping, admin, seat elasticity, enterprise search. Premium seat for the builder, Standard for operators.

The killer synergy — Google Workspace + Claude Team

This is the combination that changes the physics. On its own, Workspace is just email and Drive. Wired to Claude Team it becomes the org’s queryable memory:

  • It fixes identity first. Today email is personal Gmail, per person — no admin, no SSO, no shared permission model, and consumer Drive isn’t governable. Workspace gives every seat a managed business identity, the precondition for permission-aware enterprise search and for controlling what each agent may read. You can’t do org data governance on personal accounts.
  • “Ask Your Org” needs somewhere to look. Claude Team’s enterprise search is a live, permission-aware query across connected tools. The instant work moves from WhatsApp and personal accounts into managed Drive, Gmail, and Docs, Claude can read all of it — every brief, sheet, and thread — where before it could see none. Workspace is the substrate; enterprise search is the payoff. This is the org-maturity gating law made concrete.
  • One vendor kills a line item. Workspace bundles Google Chat + Meet, so the WhatsApp replacement doesn’t need a separate Slack subscription. One identity (Google SSO backs the Claude seats), one admin, one bill. (If we later want richer chat + the Claude Slack connector, Slack becomes an option — but it’s not required to get off WhatsApp.)
  • Sheets + Claude = the Tier-1 reporting loop. The fastest compounding win (reporting) reads ad data and writes to Sheets/Docs the team already lives in — no new surface to learn.
  • Drive = the raw-material catalog. Product photos, brand assets, and the creative raw-materials the factory needs sit in Drive, one connector away from the creative agents.

The second synergy — Notion + Claude MCP

Process capture is the org-maturity rung that does the quiet heavy lifting: an agent can only run a process that’s been made explicit. Notion is where “explicit” lives, and its Claude connector makes it two-way:

  • Claude reads Notion for context (the SOP, the angle guidelines, the approval policy) and writes back status, drafts, and logs.
  • The two-gate social calendar runs here: Claude proposes the calendar in Notion → human approves at Gate 1 → Claude drafts copy/carousels → human approves at Gate 2 → Zernio publishes. The approval trail is native, searchable, and agent-legible.

To realize the Claude ↔ Workspace/Notion synergy, those connectors must be authorized in Claude settings (claude.ai → Connectors). Until then Claude can’t read them — the tools exist but the bridge is down.


Pillar 2 — Data Maturity (Track B)

Governs what the org can know and how fast. Target rung is Growth (warehouse + dbt + a BI/agent layer). Built slowly, in step with adoption — a warehouse nobody queries is the classic failure.

Adding:

ToolRoleWhy
GitBackbone / syncSource of truth + version history; every doc is plain markdown you own. Sharing and ownership stay in Git.
OpenKnowledge (by Inkeep)Agent-native KB + retrievalOpen-source, local-first, git-backed markdown editor + knowledge base — the tool we’re authoring this vault in. Native MCP, built-in agent skills, and agentic hierarchical-RAG search; humans and agents co-author the same files (real-time CRDT). This is the agent brain: the persistent, retrievable context both AI surfaces read from and write to.
ObsidianHuman vault viewExisting personal browsing view over the same markdown. Overlaps with OpenKnowledge — see open decisions.
MotherDuckWarehouseFlat, predictable cost; native MCP; DuckDB runs free on the VPS for dev. Default over BigQuery (cheaper but per-scan, risky under broad agent queries).
Airbyte (self-host on VPS)IngestionFlat compute cost that steps with brand tiers, not rows — the sub-linear line. Connector Builder covers Shopee / TikTok Shop.
Dagster + dbt-core (self-host)Transform / orchestrateDeclarative Python assets + official agent skills → the most Claude-Code-friendly orchestrator. Same agent builds workflows and maintains the pipeline.
Managed ad accessIngestion substrateMeta System User token, Google MCC, TikTok Business Center — stable multi-account access that survives constant new-account churn.

The synergy — the warehouse is an agent, not a dashboard

The context doc’s law is dashboards are the product; insight velocity is the metric. The stack is built so the “dashboard” is a conversation:

  • MotherDuck’s native MCP means Claude queries the warehouse in natural language. No BI seat needed yet — that’s why BI (Omni/Hex) stays deferred; Claude + MCP is the BI layer until scale demands more.
  • Read-only scoping via Claude Team lets an agent read the warehouse (and Meta Ads) without write access to live systems — governance that only Team provides.
  • The vault is the agent’s long-term memory. OpenKnowledge makes it agent-native — native MCP + hierarchical-RAG search mean Claude/Codex retrieve the right doc (strategy, angle library, policy) rather than grep for it; Git holds history; Dagster/dbt hold the procedural pipeline. Declarative context vs procedural pipeline, cleanly split.
  • Naming conventions are the join keys that link produced assets ↔ Ads Manager ad names ↔ performance exports — the thread that makes the whole loop measurable. This is a convention, enforced in the vault, not a tool you buy.
  • Self-hosting is the thesis in hardware. Airbyte + Dagster + dbt share one VPS, so Track-B cost is a flat step-function while output climbs — exactly the sub-linear line the cost model proves.

Pillar 3 — Product / AI (Track C)

The multiplier, pointed at output. Ships fast (ship → test → iterate), gated per-item by the substrate rung it needs. Two surfaces: internal leverage (all near-term weight) and AI-as-product (thin for now).

Adding:

ToolWorkflowWhy
Claude (Team seats + API)The multiplier itselfRuns every workflow; API meters usage that scales with volume. Team governs it.
ForeplayIdeationScrapes followed brands’ live ads (MCP/API) → the raw material for pattern/format breakdown.
CanvaCreative productionTemplate autofill + asset hosting via MCP → finished image assets at volume.
ZernioSocial managementOrganic scheduling + analytics + hosted MCP, human-approval-before-publish.
Higgsfield (+ HeyGen / Arcads)Video generationGeneration layer paired with CapCut editing (already in use): an ad-focused engine + a UGC/avatar specialist for spokesperson volume. Exploratory, Phase 3.
Scraping (e.g. Apify)Ideation / researchFollowed-brand and market scraping beyond Foreplay’s coverage.
Operator benchBuild + runCodex and Claude Cowork are the core — where the builder builds and operators run — alongside Obsidian, GitHub Desktop, and Antigravity.

The synergy — the creative factory is one MCP pipeline

Each tool is a hop, and because every hop is MCP-connected, Claude orchestrates the whole chain rather than a human shuttling files:

Foreplay (scrape live ads)  →  Claude (break down angles/formats)
      →  Canva (autofill finished assets)  →  Zernio (schedule, human-approved)
      →  performance ledger in Obsidian  →  feeds the next Foreplay ideation
  • The loop closes in the vault. Angle-level results land in the Obsidian performance ledger, so the next round of ideation is data-informed, not guesswork — ad-run-duration as the prioritization signal.
  • Compliance is a gate, not an afterthought. For the supplement brand (Lipidri, MeSTI + Halal), copy claims must draw from an approved substantiated list — enforced as a gate in the pipeline before anything ships.
  • Higgsfield/HeyGen bolt onto ideation. Video generation pairs with the same angle library, so UGC-style variants are produced from the same strategic context, not a separate brief.

Phasing map — tooling by pillar × phase

Aligned to Applied AI Roadmap. Waves, not gates — all three move at once.

PhaseOrganization (A)Data (B)Product / AI (C)
0 · Foundation (now)Google Workspace + Notion + Claude Team stood up; begin WhatsApp → Workspace/Notion migration; operator benchGit repo + Obsidian vault; core MCP servers; managed ad tokens (Meta/Google/TikTok)Reporting Tier 1: Claude + ads MCP; Canva (in use)
1 · First loopsTrain marketers on reporting + ideationMotherDuck (1 brand) + Airbyte + Dagster + dbt on VPS; MotherDuck MCP; SupermemoryIdeation (Foreplay); creative production in draft (Canva MCP)
2 · Scale factoryOperators own creative + social; support channelWarehouse fan-out all brands; Shopee + TikTok Shop dev IDs + ingestCreative production matured; social management (Zernio)
3 · Depth & exploratorySelf-sustaining Own→Iterate→Sharedbt tests, scheduled reports, anomaly alerting; optional BI (Omni/Hex) if neededReporting Tier 2 (unified, MER); explore video (Higgsfield/HeyGen) + CRM/CS

Open decisions (pick as data comes in)

  • Warehouse: MotherDuck (default — flat, agent-safe) vs BigQuery (cheaper at tiny scale, per-scan risk). Model both as a toggle in the cost model.
  • Ingestion: self-host Airbyte (default — flat cost) vs Airbyte Cloud (zero ops, per-row cost).
  • Orchestration: Dagster OSS self-host (default) vs Dagster+ managed (if ops burden bites).
  • Chat: Google Chat (bundled, free) vs Slack (richer + Claude Slack connector, +cost).
  • BI: none now (Claude + MCP) vs Omni/Hex later at scale.
  • Task tool: Notion confirmed over ClickUp — reconfirm it fits the team’s habits.
  • Knowledge base: consolidate on OpenKnowledge (agent-native — MCP + RAG + git-backed sharing) vs keep Obsidian for personal notes, or run both over the same markdown. OpenKnowledge is the stronger “agent brain”; Obsidian is the incumbent view.

The one-line summary

Two managed anchors the org can’t run without — Google Workspace (the org’s memory) and Claude Team (the multiplier + governance) — bridged by connectors so Claude can see and act; a self-hosted data plane (MotherDuck + Airbyte + Dagster on a VPS) that keeps cost flat as brands scale; and an MCP-wired creative factory (Foreplay → Canva → Zernio, +video) that compounds through the vault. Every tool either raises a substrate Claude stands on, or is a surface Claude acts through. Nothing is bought that the agent can’t reach.