Esto no va de seguridad. Va de quién tiene el kill switch. Una agencia del gobierno de EE.UU. obligó a Anthropic a retirar del mercado Fable 5 y Mythos 5 por un "jailbreak" En el contexto LLM, un jailbreak es un prompt (o secuencia de prompts) que logran generar una respuesta que supuestamente el entrenamiento de seguridad debiera haber bloqueado (ej. como hacer una bomba). El laboratorio más alineado con la narrativa de "AI safety" está argumentando públicamente que un jailbreak puntual no debería retirar un modelo a cientos de millones de usuarios. El argumento es esencial: el indeterminismo de los modelos hace matemáticamente imposible garantizar la inexistencia de jailbreaks menores en un modelo de frontera Si un "jailbreak puntual" justifica retirar un modelo en producción, todo modelo frontera es retirable para siempre. La resistencia perfecta a jailbreaks no existe. Cada producto AI-native ahora carga riesgo de recall gubernamental en su diagrama de arquitectura. Multi-proveedor dejó de ser paranoia esta mañana. Es el nuevo mínimo viable de cobertura. Esa carta es el precedente que va a citar todo founder construyendo sobre rails de IA.

Classical guitar is mostly about what you don't play. A roadmap is mostly about what you refuse to build. Same discipline. Different instrument.

The cert is what the moat looks like before it's named. Regulated CX — federal, financial, healthcare — needs typed surfaces, full audit trails, and reversible permissioned agents. Most "AI chatbot on CRM" stacks can't pass that bar because the architecture wasn't built for it. Sierra's FedRAMP High is the first public proof that agentic CX is leaving the SMB sandbox

Ulta has 46M loyalty members. They drive 95% of US sales. At that ratio, "loyalty program" is the wrong noun. It's the operating system. Same week: Sephora links 80M Beauty Insider members into ChatGPT as a storefront. The agentic discovery surface sits on top of the loyalty graph, not separate from it. Read the two together. The loyalty graph isn't a CRM module bolted onto the retail stack — it IS the retail stack, with merchandising, personalization, and the agentic surface all plugging into it. The procurement question for any retail CMO in H2 2026 has shifted. Not "do we have a loyalty program?" but "what runs on top of it?" Most LATAM retailers still treat loyalty as a CRM module. The gap between that and Ulta isn't a feature gap. It's years of compounded loyalty-graph data acting as the operating layer. The retailers that win the next decade will be the ones whose loyalty layer is callable by every other system — including the agent the customer hasn't met yet.

Pensar problemas en formato solución era un error caro. Ahora es uno absurdo. Las soluciones están a n-tokens. Entender el problema cuesta todo lo demás.

Ocho proveedores en una sola línea de presupuesto. Carrefour comprometió €3.000 millones a su transformación digital y de IA bajo su hoja de ruta 2030. Los socios nombrados: OpenAI, Google Cloud, Microsoft, Bain, Capgemini, Publicis Sapient, Artefact, Centric. Cada uno entrega una pieza. El trabajo de Carrefour es la integración. Pueden pagarlo: €21.100 millones de facturación trimestral, 50 millones de socios de fidelización hoy (meta de 60 millones a 2030), Hopla — su asistente sobre OpenAI — ya en producción, y el catálogo del supermercado disponible dentro de ChatGPT como punto de venta. La tesis es correcta. El costo de armarla es enorme. Cualquier retailer Tier-2 que esté mirando esto tiene una alternativa más limpia. Comprar una sola plataforma AI-native donde fidelización, mensajería, segmentación y la capa agéntica viven en un solo esquema, no en siete contratos. Saltarse el armado. Llegar al mismo destino en un trimestre, no en dos años. El costo de integración era un feature de la década SaaS. En la era AI-native es un bug.

Same primitive one layer up. When the loyalty/CX engine is agent-callable instead of UI-clickable, the operator-side CMO becomes a "small high-agency team" too. Connectors collapse team size twice — once for the founder, once for the marketing org that runs the system the founder buys.

Same primitive missing in loyalty: fidelity events drop silently at the POS handshake. Most platforms don't measure it because they can't see it. The dropped-event problem is universal everywhere data crosses a rail — the recovery layer becomes the moat.

Operators shipping AI-native products see this debate from below. The "fictions that organize collective work" today are APIs, schemas, contracts, audit trails. That is where the framework actually lives — built or unbuilt, line by line. Most of the new age is being decided in code nobody is debating.

Mastercard just turned billions of transactions into a foundation model and put it behind an API. Your loyalty stack is either a consumer of network-scale models, or it is a museum. AI-native means being able to call the network's model. AI-bolted-on means you are still asking your data scientist to train one on a fraction of the data.

Taco Bell 8% comp. Wendy's -7.8%. Same quarter, same category. Q1 2026 was the quarter the digital loyalty gap stopped being invisible on the P&L. Weather got the blame. The blame is structural.

Not CRM-with-AI. Post-CRM. Bret Taylor just raised $950M at $15B for Sierra. 40% of the Fortune 50. $150M ARR in seven quarters. One in three of the world's largest banks. The category arrived while the procurement docs were still being written.

I spent three years as a Senior PM at Mercado Libre. The Q1 2026 number that matters isn't the 49% revenue. It's the 85% of customer service running on agents. 85% AI handling at MELI's volume isn't a productivity proof point. It's the category-arrived signal for LATAM. The cohort that was still asking "will AI customer service work at scale" got an answer the size of Latin America. And the pairing is the part most readers will miss. Same quarter as the 85%, revenue grew 49% — fastest since Q2 2022. The framing isn't AI-as-cost-saver. It's AI-as-growth-engine. The agentic layer didn't replace the operators. It freed the operators to compound the rest of the business. For every CMO and CPO looking at this from the outside: the question isn't whether to ship an AI-native CX layer. It's whether your data is shaped to make 85% structurally possible. Most companies' isn't. That's the architecture problem this quarter just made visible.

For four years, Starbucks loyalty members were the only cohort transacting more. Last quarter, the rest of the brand caught up. The "loyalty as discount channel" frame doesn't survive that data. The program wasn't the cost. It was the floor.

If your loyalty members keep transacting and your non-members don't, that's not a problem with the loyalty program. That's the program holding the brand up. Starbucks just published four years of data on this. Q1 FY26 — for the first time since Q2 fiscal 2022, transaction growth came from members AND non-members. For nearly four straight years, the loyalty cohort was the only group transacting more. Through pandemic recovery. Through inflation. Through staffing crises. The "discount channel" frame for loyalty programs was always backwards. The discount is the surface artifact. The cohort is the asset. The asset compounds; the artifact decays. Most retail and QSR operators look at loyalty as a margin cost. Starbucks' last four years say it's the floor that kept the brand transacting through the softness — and the cohort whose return-to-growth signals the brand has cleared it. The program led the recovery. The brand followed.

The Italian word bottega doesn't mean workshop. It means a place where mastery transfers through proximity. The bottega never scaled. Most things worth doing don't.

The hardest engineering problem in loyalty isn't loyalty. It's getting the loyalty engine to fire inside the 200 milliseconds between authorization and settlement, every single time, without becoming the reason the transaction failed.

AI replaces the easy half of a job. What's left exposes how much of the rest was judgment nobody priced in.

The agentic commerce stack just locked in. Four open protocols. Each ratified by serious infrastructure in the same window: - MCP (Anthropic) — agents calling tools - A2A (Google) — agents talking to agents - UCP (Google Shopify) — agents transacting with merchants - AP2 (Google) — agents paying under bounded authority The endorsers tell the whole story. On the payments side: Visa, Mastercard, Amex, Stripe, Adyen. On the merchant side: Walmart, Target, Home Depot, Macy's, Best Buy, Etsy, Wayfair. The two sides of the commerce balance sheet ratified the same standard inside a single news cycle. Two things follow. First, protocols stop being a differentiator. They become infrastructure, the way REST stopped being a differentiator around 2010. The retailer who ships UCP and AP2 cleanly gets parity with the rest of the market. Nothing more. Second, the competitive rent moves up the stack. The retailer who knows what to do with the agentic transaction once it lands — which customer to recognize across surfaces, which reward to fire at the moment of payment, which signal to capture about an agent-driven purchase that a human never sees — wins the next decade of margin per customer. Every agentic purchase is a customer telling you exactly what they want, what they decided to spend, and what their bounded authority is. That's the cleanest behavioral signal in the history of commerce. Most retailers are about to receive it and throw it away. That gap is the next category split.