Applied AI — Product / AI Build Playbook
The concrete, buildable counterpart to Applied AI Tooling Roadmap. The tooling roadmap says what tools; this says how each workflow is actually wired — named tools, data flow, build steps, and where the knowledge base sits in the loop. Ordered least-substrate-first per the sequencing rule in Applied AI Roadmap.
Framing (you’re right): most of “Product / AI” is internal leverage riding on the infra + org substrate — reporting, ideation, creative, social. Only one surface is genuinely customer-facing AI-as-product: the DM / CS bot (§5). That’s why it gets the deepest treatment. Everything here is a Track-C usage line in the cost model; the new monthly costs this research surfaced are collected at the end to fold into EFFEN_Cost_Model.xlsx.
Per-system format: Goal · Architecture · Build · KB-in-the-loop · Needs (rung) · Phase · Cost · Gate.
All prices USD, current as of Jul 2026 (sources at end). Convert at the model’s FX line.
§0 · Full-funnel reporting
Tier 1 — fast reporting, no warehouse (ships first — Phase 0)
- Goal: marketers pull Meta/Google/TikTok performance through chat, zero manual export.
- Architecture:
Cowork/Claude + ads MCP → Google Sheets/Docs. The connected ads MCP reads the accounts; Claude composes the report into the Sheets/Docs the team already lives in. - Build: (1) wire ads MCP with read-only scoping via Claude Team; (2) store the brand↔account map + KPI targets in the vault; (3) save a reporting skill (the standard weekly/daily pull + format); (4) run it, iterate the template.
- KB-in-the-loop: brand list, account IDs, KPI targets, and naming conventions live in the vault as agent context.
- Needs: ads MCP only. Phase 0. Cost: existing Claude seats. Gate: read-only tokens so an agent can never write to live ad accounts.
Tier 2 — unified, MER-based (Phase 3)
- Goal: one source of truth across ads + WooCommerce + Fighter.my, measured on MER not last-click.
- Architecture:
Airbyte → MotherDuck → dbt models → Claude via MotherDuck MCP. Dashboards are conversations, not a BI seat (why Omni/Hex stays deferred). - Build: ingest ads + Woo + Fighter.my → dbt marts (spend, revenue, MER by brand/market) → agent reads via MCP → scheduled narrative reports.
- Needs: the warehouse (Phase-2 fan-out). Phase 3. Cost: warehouse + VPS lines (already modelled). Gate: dbt tests before any number is trusted.
§1 · Ideation agentic workflow (Phase 1)
- Goal: turn what’s working in-market into a steady feed of creative directions.
- Architecture:
Apify (scrape competitor/followed-brand commercial ads) Foreplay MCP (swipe file + Spyder competitor tracking + AI briefs) └──► Claude + angle-library.md ──► pattern/format breakdown ──► ideation output - Key finding: the official Meta Ad Library API returns political/social ads only — useless for commercial ecommerce. Apify actors are the realistic route for competitor commercial ads (~$0.40 / 1,000 ads). Foreplay brings a 200M-ad swipe file, 24/7 competitor tracking (Spyder), and AI briefs — and it’s native-MCP (no API key), so Claude queries it directly.
- Build: (1) maintain a followed-brands list in the vault; (2) Apify scheduled actor → dumps ads to the vault/warehouse; (3) Foreplay MCP connected to Claude; (4) a skill that clusters recurring hooks/formats and proposes angles.
- KB-in-the-loop:
angle-library.md(angle → hook → proof → format) loaded as direct markdown context — for a small, focused library this beats RAG (up to ~95% fewer tokens). - Needs: vault + Foreplay + Apify. Phase 1. Cost: Foreplay Workflow ~45/mo.
§2 · Creative production factory (Phase 1–2)
- Goal: a paid-ads factory outputting finished static image variants + copy for marketers to run manually (explicitly not auto-publishing).
- Architecture:
Claude (picks angle from angle-library.md + performance-ledger.md; drafts copy + brief) ├──► AI image gen: Nano Banana (product-consistent) · Ideogram (text-heavy) · Flux (photoreal hero) └──► Template render: Creatomate / Placid API ──► N variants ──► human review ──► marketers run - Why not Canva API: Canva’s Autofill/Connect API requires Canva Enterprise for the developer and every user — a heavy gate. Creatomate (19/mo) do templated image+video generation via API without the Enterprise lock. Keep Canva + CapCut for manual brand/video polish.
- AI image models: Nano Banana (Gemini Flash Image, ~$0.039/img) for product/character consistency; Ideogram v3 for legible in-image text; Flux 2 for photoreal product shots. All have APIs for agentic use.
- Build: (1) build the raw-materials catalog (product shots) in Drive/vault; (2) branded render templates in Creatomate; (3) skill: angle → copy + image prompt → gen → render → variant set; (4) log results to the performance ledger.
- KB-in-the-loop:
angle-library.md+performance-ledger.csv(angle, brand, CTR/ROAS/CPA, date, recheck-by) — the agent ranks angles by past performance. Ad-run duration is the prioritisation signal. - Gate — compliance: for the supplement brand (Lipidri, MeSTI + Halal), every copy claim must be drawn from an approved substantiated-claims list in the vault. The agent may only use claims on that list.
- Needs: raw materials + angle library + performance ledger + image gen + render API. Phase 1–2. Cost: Creatomate 80/mo (≈2k imgs) + Foreplay/Apify (shared with §1).
§3 · Social media management (Phase 2)
- Goal: automated organic publishing across FB, IG, TikTok, Threads — a growth loop (IG/TikTok) + an operational track (FB/Threads), gated by human approval.
- Architecture:
scrape top content ─► Claude (analyse winners, draft copy + carousel/Reel) ─► Gate 1 (strategic calendar review, in Notion) ─► Gate 2 (finished copy/carousel review) ─► Zernio (MCP, ISO-scheduled) ─► IG / FB / TikTok / Threads orchestration: n8n - Platform truths (important): IG + FB = genuine hands-off auto-publish (IG ~100 API posts/24h). TikTok direct-posting is gated behind TikTok’s content audit — using Zernio’s already-audited client sidesteps building/passing your own. Threads has no native scheduling — any “schedule” is a queue/cron the tool runs, publishing immediately at fire-time (~250/24h). Zernio is a good fit: MCP-native (~280 tools), covers all four, ISO-timestamp queue, approval-friendly.
- Build: (1) Zernio connected (MCP) with human-approval-before-publish; (2) two-gate approval in Notion (calendar → finished assets); (3) n8n orchestrates the scrape→draft→stage loop; (4) Claude drafts from brand-voice + performance context.
- KB-in-the-loop: brand voice, content pillars, and the performance ledger in the vault; the agent stages drafts into Zernio, humans approve, Zernio publishes.
- Needs: vault governance + Zernio + approval workflow. Phase 2. Cost: Zernio plan + n8n ~$20/mo.
§4 · Conversational AI — DM + CS automation ⭐ (the customer-facing surface — Phase 2–3)
This is the direct answer to “can we build DM automation on TikTok/IG/FB using our AI + knowledge base?” Short version: yes for Instagram, Facebook Messenger, and WhatsApp; no for TikTok (no official DM API). And yes — it runs on your own Claude + knowledge base, not a generic canned bot.
What’s possible, per platform
| Channel | DM automation? | How |
|---|---|---|
| Instagram DMs | ✅ Yes | Official IG Messaging API. Reactive automation to user-initiated DMs, comment→DM, 24-hr window (+7-day Human Agent tag for real humans). |
| Facebook Messenger | ✅ Yes | Messenger Platform. Same model; bots must reply within 30s and disclose automation. |
| ✅ Yes (key for Malaysia) | WhatsApp Cloud API. 24-hr service window free; per-message pricing since Jul 2025 (Marketing billed, Service free in-window). | |
| TikTok | ❌ No (honest) | No general TikTok DM API. Business Messaging is gated (and excluded for US/EEA/UK); only TikTok Shop seller messaging exists, for approved Shop sellers. Third-party “bridges” violate ToS — don’t risk brand accounts. |
Recommended architecture — ManyChat + your own Claude+KB
The trick: a no-code platform owns the hard, compliance-heavy plumbing (Meta app review, webhooks, the 24-hr window, human handoff), while your endpoint owns the intelligence (Claude grounded in your vault). ManyChat’s “External Request” is the bridge.
Customer DM (IG / FB / WhatsApp)
│
▼
ManyChat Pro ($29/mo) ── owns: Meta app review · 24-hr window · inbox · human handoff
│ External Request (webhook: message + username + context)
▼
Your API (serverless — Vercel / Cloud Run)
│ ├─ retrieve from OpenKnowledge vault via its native MCP / hierarchical-RAG
│ │ (products · FAQs · policies · APPROVED substantiated-claims list)
│ └─ Claude answers, grounded ONLY in retrieved KB; claims list = hard guardrail
▼
JSON reply ──► ManyChat sends the DM + lead/contact ──► CRM
│
└─ escalation ──► human agent (Meta Human Agent tag, 7-day window)
Build steps
- Set up IG Business/Creator + FB Page and a Meta app; submit for App Review (
instagram_business_manage_messages,pages_messaging) — start early, it takes weeks. - Stand up ManyChat Pro; connect IG + FB (+ WhatsApp later).
- Build one HTTPS endpoint (serverless): embed the incoming message → retrieve from the OpenKnowledge vault (its native MCP + RAG is exactly this retrieval layer) → call Claude with a strict system prompt: answer only from retrieved context; for the supplement brand, only claims on the approved list; escalate medical/edge questions to a human.
- In ManyChat, add an External Request step posting the message to your endpoint and mapping the JSON reply back into the DM.
- Wire human handoff (ManyChat live chat + Human Agent tag) and lead capture to the CRM.
- Add WhatsApp Cloud API for CS (Malaysia’s dominant channel) — same endpoint, new channel.
Alternatives (pick by need)
- Respond.io ($79–279/mo) — stronger shared team inbox + omnichannel + CRM-grade workflows; same “HTTP request → your Claude+RAG” pattern. Choose it if CS volume needs a real agent console.
- Botpress / Voiceflow — LLM-first, native Anthropic + built-in Knowledge Bases (RAG); more control, more build. Choose if you want the whole bot in one dev platform.
- ManyChat — fastest to live (1–3 weeks) and cheapest; the recommended start.
- Fully custom on Meta Cloud API — cheapest per-message at scale but 6–12+ weeks of build + you own review/compliance. Only once volume justifies it.
KB-in-the-loop & Gate
This is where OpenKnowledge earns its keep — native MCP + hierarchical-RAG means the bot retrieves live from the same vault humans author, so answers stay current and on-message. Gate: the bot must never state an unapproved health claim — retrieval for Lipidri is restricted to the approved substantiated-claims list, and medical questions escalate to a human. This is compliance-critical (MeSTI/Halal, ad-claim rules).
- Needs: KB (OpenKnowledge) + Meta business accounts + app review. Phase 2–3. Cost: ManyChat 20 + WhatsApp per-message + Claude API usage.
Exploratory
Video generation (Phase 3)
Generation (Higgsfield general + HeyGen/Arcads UGC avatars) paired with CapCut editing; draws on the same angle-library.md so UGC variants come from the same strategic context as static creative. Pairs directly with §1 ideation.
Marketplace APIs — Shopee + TikTok Shop (Phase 2–3)
Request developer access (Shopee Open Platform, TikTok Shop Partner Center) → ingest orders + performance into the warehouse for fuller per-brand MER. Note: TikTok Shop seller messaging is also the only TikTok DM path (ties to §4).
Enablement & Support (continuous — Track A)
Hands-on training on the bench (OpenKnowledge/Obsidian + GitHub, Codex, Claude Cowork, Antigravity), full environment setup, and the onboarding path for everything above; plus a standing support channel. In-house, this is ongoing — not a billable line.
New cost lines to fold into the model
From this research — add to Track C on EFFEN_Cost_Model.xlsx in tomorrow’s pass:
| Line | Tool | ~USD/mo | Scales with |
|---|---|---|---|
| DM platform | ManyChat Pro (→ Respond.io at scale) | 29 | contacts |
| DM/CS hosting | serverless endpoint (Vercel/Cloud Run) | 20 | usage |
| Cloud API | per-message | conversations | |
| Ad scraping | Apify | 45 | ad volume |
| Ideation | Foreplay Workflow | 149 | flat (5 seats) |
| Creative render | Creatomate (or Placid $19) | 41 | variants |
| AI image gen | Nano Banana / Ideogram / Flux | ~80 | creatives/mo (≈ brands) |
| Orchestration | n8n cloud | 20 | flat |
Most are flat or per-seat; only image gen + Apify + WhatsApp scale with output — consistent with the model’s “video/API/scraping are the linear lines” finding.
Build sequence
Reporting Tier 1 → Ideation → Creative production → Social management → DM/CS bot → (Video, Marketplace APIs, Reporting Tier 2). Each ships when its substrate rung exists; DM/CS waits for the KB (OpenKnowledge) to be solid enough to ground answers safely.
Sources (accessed 8 Jul 2026)
- Instagram Messaging API · Messenger Platform policy — developers.facebook.com/docs/instagram-platform / …/messenger-platform/policy
- WhatsApp Cloud API pricing (per-message, Jul 2025) — developers.facebook.com/documentation/business-messaging/whatsapp/pricing
- ManyChat External Request + pricing — help.manychat.com (Dev Tools: External request) · manychat.com/pricing
- Respond.io custom channel + pricing — respond.io/help/custom-channel · respond.io/pricing
- TikTok: no DM API / Business Messaging gating — developers.tiktok.com/doc/webhooks-overview · business-api.tiktok.com/portal/docs
- Instagram Content Publishing API — developers.facebook.com/docs/instagram-platform/content-publishing
- TikTok Content Posting API (audit requirement) — developers.tiktok.com/doc/content-posting-api-get-started
- Threads API (no native scheduling) — developers.facebook.com/docs/threads
- Foreplay MCP/API + pricing — foreplay.co/mcp · foreplay.co/pricing
- Meta Ad Library API (political-only) — facebook.com/ads/library/api
- Apify ad-library scrapers — apify.com/apify/facebook-ads-scraper
- Canva Connect / Autofill (Enterprise-gated) — canva.dev/docs/connect/autofill-guide
- Template APIs — creatomate.com · placid.app · bannerbear.com/pricing
- AI image model pricing (Nano Banana / Ideogram / Flux) — ai.google.dev/gemini-api/docs/pricing