The Agentic Enterprise. AK · Daily Edition · 8 min read | Friday, November 13, 2026 · Daily Edition Agents need a key to the storeUnited, Delta, American, Walmart, Amazon and JPMorgan Chase now treat AI agents that buy on a customer's behalf in at least four different ways. For a year the agent pitch assumed the open web would let software shop for you. It won't, at least not on the same terms everywhere. The companies that own the checkout are deciding which agents get in, what they can see and who signs off. If you're building agents that buy, book or pay, plan on a separate negotiation with every counterparty. If you sell online, you need a policy of your own. | The Big Story Commerce Block it, log it in, or build it a gate| U | nited updated its contract of carriage on Wednesday to bar AI agents from making, modifying or servicing tickets, including agents that come through travel agencies. Fast Company Agents can still show fares. CEO Scott Kirby gave the commercial reason plainly: "I don't want to create another middleman." PYMNTS |
The rest of the field split. Delta lets an assistant book if it can authenticate the passenger's own credentials, and holds the traveler responsible for approving the purchase. American says some AI tools can buy its tickets, but it monitors automated traffic and blocks it when needed. Fast Company JPMorgan Chase barred outside agents from logging in with customer credentials a month earlier. PYMNTS Amazon blocked Meta's Muse in September with a popup telling shoppers that an unauthorized agent breaks its conditions of use. Adweek Walmart went furthest the other way. Its Agent Gateway lets approved agents search and buy, and Agent Profiles, announced for this month, give each agent its own spending limit, approval rule and purchase history. Agents never see payment details, and Walmart says it will block those that break its terms. TechCrunch Volume is still tiny: LLM traffic was under 1% of room nights at Booking Holdings in the second quarter. So the rules are being written while agents have little volume to bargain with. Airbnb plans its own travel agent for the first half of 2027, and OpenAI announced it is buying Odessia, a booking agent, this week. PYMNTS The agent that shops for you will need a different key at every door. The Spearhead Take Work both sides. For agents you build, design for authenticated, per-merchant access with a human approval at payment, and assume some counterparties will block you. For your own site, decide how outside agents get in, through a governed gate or not at all, and publish the rule before the traffic arrives. |
Sources: PYMNTS · Fast Company · Adweek · TechCrunch | The Obvious & The OverlookedWhat is easy to see, and what is easy to miss. The Obvious Anthropic set IPO terms Tuesday: 825 million shares at $70 to $78 each, ticker ANT on Nasdaq.Anthropic. | Norwest raised a $4 billion fund, its largest ever, with AI as the priority.TechCrunch. | SpaceX plans more than $74 billion for two AI supercomputing campuses, both still awaiting local approval.Reuters via Investing.com. |
|
The Overlooked Prompt-injection defense is moving into the model: IBM's LLMON previews avoid 47% more attacks than the best frontier model tested, IBM says, but aren't production-ready.Hugging Face. | The cheapest agent call may be a yes or no: Google's Decisions API answers up to 1,000 closed questions a request for $0.10 per million input tokens.Google. | Agent vendors are shipping into other people's chat windows: Nectar plans to offer its creator agent over MCP inside assistants like Claude and ChatGPT.TechCrunch. |
|
|
| Moving PiecesNvidia's first earnings without Huang, a lawsuit aimed at the next model, AI debt the Fed can count, a breach two vendors deep and a cheaper Google workhorse. Compute Kress has to sell a forecast Huang madeNvidia reports on Tuesday, November 17, its first results since Jensen Huang resigned on Thursday. Colette Kress, now interim CEO and still CFO, inherits August guidance of $108 billion in quarterly revenue at a 74% gross margin; LSEG consensus sits near $109.2 billion. Shares closed Thursday at $166.19, more than 30% below the October high, and at least nine brokers cut their targets. Watch the fourth-quarter outlook and anything said about supply. A caretaker team that guides cautiously will tell you more about allocation than any roadmap statement. Invezz via TradingView · 24/7 Wall St. via Yahoo Finance | Governance Four nonprofits ask a court to freeze the frontierPublic Citizen and three other nonprofits sued OpenAI, Anthropic, Google, Meta and SpaceX's AI unit on Thursday in San Francisco Superior Court. They argue California's unfair competition and public nuisance laws bar selling products a company knows are unsafe, and they want an injunction freezing more advanced models. Florida's request to bar OpenAI from new models without outside oversight is still pending. Judges rarely grant orders this broad, but discovery or a settlement could slow releases. Keep a tested second model for every critical workload. Public Citizen · Dataconomy | Markets The Fed puts a number on AI's borrowingMichelle Bowman, the Fed's Vice Chair for Supervision, said Tuesday that AI-related debt is on track to pass $800 billion this year, with other obligations estimated at more than three times that. She attributed about a quarter of above-target core inflation to AI-related price pressures, as demand for chips and servers outruns supply. Her stated risk: if AI revenue lags costs, AI equities could correct sharply and credit could tighten. Before you sign multiyear terms, check how your key AI vendors fund themselves. Federal Reserve | Security Notion finds out its vendor's breach was its ownNotion confirmed that attackers took chat transcripts from Notion AI customers who used its Insights feature, every input and output included, after a breach at inference provider Fireworks AI. About 1% of customers were affected; Notion cut ties with Fireworks on November 3. Fireworks says the attacker breached its cloud provider to reach a development system; TechCrunch counts it as Fireworks' second breach in two months. Your AI data map needs three columns now: the model maker, the inference host, and whoever hosts the host. TechCrunch | Product Google prices its agent workhorse to run all dayGoogle released Gemini 4 Hydrogen on Thursday at $1.50 per million input tokens and $7.50 per million output, flat across a 1 million token context. Google says it finishes tasks about 2.5x faster than Gemini 3.8 Flash, and it pitches Hydrogen as the execution model while Carbon and Argon handle deep reasoning. The benchmarks are Google's own. The routing pattern is worth copying whatever vendor you use: send most agent steps to a cheap, fast tier and save the expensive model for the hard calls. Google |
Sources: Invezz via TradingView · 24/7 Wall St. via Yahoo Finance · Public Citizen · Dataconomy · Federal Reserve · TechCrunch · Google | On the RadarTen signals, one line of context each. Energy US power use will set records in 2026 and 2027, the EIA says. Consumption rises from 4,216 billion kWh in 2025 to 4,370 billion kWh in 2027, driven largely by AI data centers. Reuters via BOE Report |
Policy Treasury Secretary Scott Bessent said the US and China expect to swap best practices for testing advanced models before release. A direct channel for alerting each other to serious AI incidents is also being set up. Reuters via SRN News |
Data Google accepted stricter privacy terms on its $10 million purchase of Spirit Airlines employee records for AI training. Health data is excluded, and the deal still needs the bankruptcy judge's approval. Reuters via KFGO |
Retail J.Crew launched a Claude-powered Style Advisor that builds outfits from its own catalog shoots. It picks from looks J.Crew's stylists already assembled rather than inventing new ones, and is US only for now. The Next Web |
Security Perplexity found that telling agents about earlier failed escape attempts changed what they tried next. An explicit attack hint let seven of 12 models succeed at least once, though none reached the host machine. Perplexity |
Product OpenAI will retire ChatGPT's Advanced and Standard Voice modes on December 11. Teams with voice workflows on them have four weeks to test the replacement. Interesting Engineering |
Commerce A brand-protection firm found AI shopping assistants pointing buyers to risky retail sites. Podqi's September tests found such sites in 15% of Perplexity, 8% of ChatGPT and 4% of Gemini responses. Gadget Review |
Cyber OpenAI put Daybreak Red Team, a fleet of agents that probes web apps for attack paths, into early access. Its Defense Factory drafts patches as merge-ready pull requests that engineers still approve. The Register |
Workforce Open product manager roles are up 75% from their March 2023 low, while engineering openings grew 2.5% in a year. Forward deployed engineer postings rose 4.6x to about 1,500, per TrueUp data. Lenny's Newsletter |
Deals Nvidia's $12.9 billion Hugging Face purchase is still set to close in the first half of 2027. The Justice Department has asked how it would affect competition, Reuters reported in October. TechCrunch |
|
| Quick HitsThe wider field, one line each. Values are as reported. Funding | 01 | UFORCE raised a $700 million Series A at a $6 billion valuation, co-led by Valor Equity Partners and the UK National Wealth Fund.UFORCE | | 02 | Agile Robots raised $500 million led by SoftBank, with Nvidia and Hitachi, at a reported $8.5 billion valuation.Reuters via Global Banking & Finance Review | | 03 | Edelbyte came out of stealth with $70 million from Sequoia and Quiet Capital for large-scale data search and analytics.Globes | | 04 | Millionways disclosed $6 million and released Thorsten-4, a model built from about 7,000 psychologist-run interviews.Business Wire via 01net |
|
Models and tools | 05 | ElevenLabs says Scribe v3 cuts word errors 27% versus v2 across 102 languages.ElevenLabs | | 06 | MiniMax released open weights for H3-Image, which it says renders 4K images in under 10 seconds.MiniMax | | 07 | Apple's Claris is testing Harbor, an app that pools memory across Thunderbolt 5-linked Macs to run larger open-weight models privately.MacRumors | | 08 | Google Labs opened a US waitlist for Tabletop, which builds shared trackers, polls and itineraries from a prompt.The Next Web |
|
Spend and adoption | 09 | Anthropic is paid by 44.8% of businesses on Ramp, OpenAI by 40.6%.The Decoder | | 10 | The average price per million text tokens fell 21% in October, per Ramp.The Decoder | | 11 | Customer support rose to 13.3% of tracked AI spend in October, from 0.8% in June.The Decoder |
|
Surveys | 12 | Four in 10 game developers now use generative AI at work, GDC's 2026 survey of more than 2,300 professionals found.80 Level | | 13 | ThinkYoung found 79% of European secondary students use AI for schoolwork, but only 47% have been shown how.EdTech Innovation Hub |
|
| From Spearhead · AI Bootcamp Get your leadership team AI-ready, in a day.The AI Bootcamp is a one-day, in-person session designed for your industry, your organization, and your use cases. Your leaders understand AI, build AI agents, envision outcomes, and start building a new work culture. |
| The Number 80% The share of OpenAI and Anthropic revenue on Ramp's platform that came from the top 1% of companies in October. That top 1% spends a median $10,200 per employee per month on AI. The typical business spends $16.32 per employee per month. Adoption is wide, and the spending is concentrated in very few companies. The Decoder · Ramp AI Index | Counter-Signal Governance A human sign-off can make the work worseWhere the new rules let agents buy, they end in the same place: a person approves the purchase. Delta puts that on the traveler, and Walmart lets shoppers approve each order or set a limit. Helsing, presenting a military targeting platform this week, says every handoff to a weapon carries the name of the human who approved it. The Decoder A study published Tuesday in PNAS Nexus should temper that comfort. Wharton's Kartik Hosanagar and the University of Georgia's Daehwan Ahn had 285 students write stories with AI under three designs. When people only approved what the AI produced, the stories rated below those written with no AI, and satisfaction trailed the designs where people kept more control. When people set the ideas first, quality beat writing alone. AI help cut writing time 36.2% overall. EurekAlert A confirm button is the thinnest kind of oversight. If that's your control, give the human an earlier job, such as setting the budget and the approved merchants. Sources: The Decoder · EurekAlert | From the Field Every agent we've put into production eventually hit a system that didn't want it there. Sometimes it's a vendor portal with a CAPTCHA. Sometimes it's an ERP that only trusts a named user, or a supplier whose fraud team reads automated orders as an attack. What fixes it is usually an agreement that names the agent, sets its permissions and says who answers for its mistakes. United, Delta and Walmart are writing those agreements in public this month, and they don't match. That's normal for a market this young, and it will stay messy for a while. The practical consequence is that an agent's reach depends on the permissions it can earn, one counterparty at a time. That changes how we'd budget an agent program. Put access work in the plan from day one, the way you would for any new trading partner, with an owner and a timeline. In our experience, access found late is the most common reason a pilot stalls. In most of our projects, the access agreement takes longer than the build. Let's create value with AI, AK | The Agentic Enterprise Know more about AI than 95% of your peers. Every morning. Anthropic, OpenAI, Google, Microsoft/Azure, AWS, NVIDIA, Databricks and Snowflake are Spearhead technology partners; United, Delta, American, Walmart, JPMorgan Chase, Public Citizen, Meta, SpaceX, Notion, Fireworks AI, IBM, Ramp, Helsing and the funded companies named are not. Produced using Anthropic’s Claude with editorial direction from AK. © 2026 Spearhead. All rights reserved. |
|