Agentic AI in Marketing & Customer

1 · The funnel becomes a loop -- why marketing moved first

Marketing has always been the fastest-iterating function in the enterprise, so it is no surprise it is where agentic AI hit the mainstream first. The classic linear funnel (awareness → consideration → purchase → loyalty) is being replaced by a continuous loop in which an agent senses intent, generates the creative, buys the media, personalises the offer, closes the sale, and feeds the outcome back into the next decision -- all without a human touching each step. AI-driven traffic to retail sites grew nearly 700% over the 2025 holiday season and those shoppers convert at 31% higher rates; AI-assistant retail traffic is up 119% year-over-year. The channel is no

Two structural shifts make this work. First, the storefront is moving inside the assistant: consumers increasingly start in ChatGPT, Gemini, Copilot, or Perplexity rather than a search bar or a brand site, so the brand has to be reachable as an MCP tool, not just a web page. Second, the buy-side and sell-side of advertising are both becoming agents that negotiate with each other -- the human marketer sets objectives and guardrails, and a fleet of agents executes against them in real time. So what: If your product cannot be discovered, compared, and purchased by an agent acting on a customer's behalf, you are invisible to the fastest-growing acquisition channel of 2026.

2 · Four live deployments rewriting the playbook

The pattern: Every player is shipping two agents -- a customer-facing one (Sparky, Klarna app) and a brand-facing one (Marty, Agentforce campaigns). The marketer's job shifts from running campaigns

3 · The continuous agent loop, stage by stage

Sense: a demand-sensing agent fuses first-party data, social signal, weather, and live pricing into a rolling view of intent. Create: a creative agent generates and adapts assets per format (vertical video, pause ad, email, SMS) -- exactly the Netflix creative-adaptation tool. Buy: a media-buying agent purchases inventory across DSPs against brand objectives, the Netflix and Walmart Marty pattern. Personalise & close: a storefront agent (Klarna, Sparky, Agentforce Commerce) answers questions, compares products, and completes checkout inside the assistant via MCP + a payments protocol (MPP / x402). Learn: every outcome writes back to episodic memory (Write-Aside pattern, Day 44) so the next loop starts smarter. This is the same Sense/CreatefiBuy/ClosefiLearn cycle Salesforce and Adobe have hard-wired into their 2026 clouds.

4 · Trust is the bottleneck -- disclosure, conflict, consent

Consumer-facing agents inherit every governance obligation from the enterprise stack, plus new ones unique to marketing. EU AI Act Article 50 (T-64 days, Aug 2 2026): any AI-generated or AI-personalised creative must be disclosed to the consumer and watermarked (C2PA / SynthID); an agent transacting on a shopper's behalf must identify itself as an agent. Conflict of interest: Netflix selling a platform-owned agent that autonomously buys inventory on Netflix's own platform is the open question media agencies are pushing back on -- transparency and independent verification are unresolved. Autonomous-purchase caution: 50% of consumers remain wary of fully autonomous buying, so human-on-the-loop approval and clear spend caps are table-stakes UX, not optional. KYA for consumer agents: the Know Your Agent identity + audit standard (Day 54) extends to any agent shopping with a consumer's stored payment method -- scoped tokens, spend limits,

Watch this: The first Article 50 enforcement actions in Q3 2026 are most likely to land on consumer marketing -- undisclosed AI creative and non-transparent agentic ad-buying are the highest-visibility,

5 · Reference architecture -- an agentic CX stack

Brain: Opus 4.8 / GPT-5.5 for planning + creative reasoning, Sonnet 4.6 for copy and conversation, DeepSeek V4 Flash ($0.14/M) for high-volume classification (intent tagging, audience bucketing). Surface: expose the catalogue as an MCP server (Klarna's Product Search MCP is the canonical example) so ChatGPT / Gemini / Copilot can discover and transact. Orchestration: Salesforce Agentforce or Adobe CX Enterprise as the control plane wiring the Sense/CreatefiBuy/Close loop; AG-UI generative-UI surfaces (Day 48) for the human supervisor. Memory: Write-Aside -- Valkey L1 + pgvector L2 with per-customer namespace isolation; Memory IDs for GDPR erasure. Payments: MPP / x402 with scoped tokens + spend caps. Guardrails: Article 50 disclosure + C2PA watermark on every generated asset, AG-UI approval gate on any spend above threshold, T1-T4 kill switch, and a WORM audit trail of every agentic ad buy for conflict-of-interest verification.

6 · Viral AI app of the day

Klarna Shopping Search in ChatGPT -- launched May 20, the clearest consumer proof-point yet of agentic commerce. The app lets a shopper search products, compare live prices, check availability, and view offers from multiple retailers without leaving the chat, powered by Klarna's Product Search MCP server spanning 100M+ products and 400M+ merchant listings across 13 markets. It turns ChatGPT into a price-comparison storefront and turns Klarna into the connective tissue between assistants and merchants -- the exact 'be the agent, expose the MCP entry point' move every brand now has to make. Meta's competing Instagram shopping agent (codenamed 'Hatch', influenced by OpenClaw) is expected later in 2026, and OpenAI projects $25B in advertising and commerce revenue by 2029, signalling how big this surface becomes. Why it matters: Klarna didn't build a chatbot -- it built the merchant-discovery layer for the assistant era. Whoever owns the product-search MCP that assistants call becomes the new shelf, and the new