AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #152

Agents at the Door

Thursday, November 12, 2026 · 2 Developments · 2 Perspectives · 2 Case Studies · 2 Resources

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

Agents are starting to arrive as customers, and companies are deciding whether to let them in. The rest of today's news is about making agents cheaper to run and putting them to work in go-to-market and support.

  • Agents at the gate. United blocks outside AI agents from booking, while Delta admits them through the customer's own login.
  • Cheap decisions. Google's new decision model answers up to 1,000 questions about one input for $0.10 per million input tokens.
  • The B2A economy. a16z argues agents are becoming a distribution channel that chooses and buys products on users' behalf.
  • Levels 9 and 10. Every proposes simulating competing strategies, and rerunning them, before you commit to one.
  • GTM plays. Five teams show marketers building their own agents, from 3-minute ABM launches to more honest attribution.
  • Airbnb's support math. AI now resolves about 40% of issues and helped cut support cost per booking more than 17%.
  • Renewal season. Buyers who start SaaS negotiations 90 days out save 49% on average, per Cledara.
  • Tool costs. How agents connect to tools changed cost per task by about 17x in one OpenRouter test.

This week, check whether an agent can find, judge and buy from you, and decide which actions you will let it take.

Developments

Competition

United Blocks Outside AI Agents From Booking, While Delta Lets Them In With the Customer's Login (November 11, 2026)

United said on Wednesday it will restrict unauthorized third-party AI agents from its app and website, and from booking through travel agencies and aggregators. It cited exposed account credentials, unintended purchases, misused MileagePlus miles, and FAA and DOT rules that require a human to make certain check-in acknowledgments. CEO Scott Kirby told the Wall Street Journal that United does not want to create another middleman. Delta takes the opposite line: assistants can book if they sign in with the customer's Delta login, and the customer is responsible for what they do.

So what

United is the first major brand to say plainly that agents threaten its customer relationship, and every company that sells online will face the same choice. Blocking protects data and margin today but risks losing buyers whose assistants simply go elsewhere. Decide now which of your flows agents may use, how they must authenticate, and who is liable for their mistakes. Delta's approach is a workable starting template.

Confidence: Established fact — The policy and its stated reasons are on the record; the middleman motive is Kirby's own framing, reported by the WSJ.

Go deeper United Airlines newsroom · Forbes · The Neuron

Technology

Google Launches a Decision Model at $0.10 per Million Input Tokens, With Output Free (November 11, 2026)

Google released Gemini Decision 1 through a new Decisions API. Instead of writing text, it answers up to 1,000 yes/no, multiple-choice or rubric questions about one input in a single pass, returning probabilities and a 0-to-1 confidence score in about 150 to 250 milliseconds. It accepts text, images, audio, video and PDFs. Google already uses it to route queries between model tiers and as a safety check on MCP tools in Search, and OpenAI and TypeSafe sell similar models.

So what

Much of what agents do is not writing but choosing: which tool, which model, whether a request passes policy. A cheap, fast classifier for those steps can cut cost and latency in your product, and probabilities are easier to audit than prose. Ask your engineering lead to run one routing or compliance check on a decision model this quarter and compare cost and accuracy against your current LLM call.

Confidence: Established fact — Pricing and specifications come from Google; claims of beating rival decision models rest on Google's own benchmarks.

Go deeper Google blog · The Deep View

Perspectives

Go-to-Market

a16z: Welcome to the "B2A" Economy, Where Agents Choose and Buy the Products (November 11, 2026)

Olivia and Justine Moore of a16z argue that agents are becoming a distribution channel on the scale of the web and mobile, picking tools and subscriptions on users' behalf. When they asked ChatGPT to make a video ad, it pulled in Krea, Descript and ElevenLabs on its own, then talked them into a paid Descript plan. They cite Shopify orders from AI-powered searches growing nearly thirteenfold in a year, and agents generating nearly 40% of Stripe's documentation traffic in June.

So what

Read next to United's move, the question for your company is whether an agent can find, evaluate and buy from you without a human clicking. That requires machine-readable pricing and docs, a connector or MCP server, and a signup path an agent can finish. Have product and marketing walk that path end to end in ChatGPT and Claude this month, and treat every failure as lost pipeline.

Hype check

Some headline claims, such as agents already making up most web traffic, carry no source in the piece, and a16z has investments riding on this thesis. The Shopify and Stripe figures are company-reported.

Confidence: Informed perspective — A credible investor thesis built partly on vendor-reported data.

Go deeper a16z · TLDR Founders

AI Strategy & Transformation

Every's Mike Taylor Adds Two Levels to AI Adoption: Simulate Competing Strategies, Then Rerun Them to Filter Out Luck (November 12, 2026)

Taylor, Every's head of evals, extends the company's eight-level adoption map, which tops out with a manager agent running subagents. At Level 9, several agent teams pursue the same goal with different strategies and are scored against synthetic customers built from real interview data. At Level 10, the contest is rerun many times to see which strategy wins reliably. Every's clone of CEO Dan Shipper makes 26% of the points Shipper actually made on its first guess, but covers 86% across 100 guesses.

So what

Few companies are anywhere near Level 8, so treat this as a direction, not a to-do. The usable idea now is cheaper: before a pricing or launch decision, have AI draft several competing plans and stress-test them against synthetic buyers built from your own call notes, then bring the survivors to the real team. Heed Taylor's caveat that a flawed simulation run 50 times just produces a confident wrong answer.

Confidence: Informed perspective — Experimental practice at one company, supported only by internal results.

Go deeper Every · Every: The eight levels of AI adoption

Case Studies

Go-to-Market

Five GTM Teams Building Their Own Agents: incident.io Launches ABM Campaigns in 3 Minutes, and Vercel Triples Events Without Hiring (November 11, 2026)

Kyle Poyar's Growth Unhinged profiles five plays. incident.io CMO Tom Wentworth replaced the company's ad agency with a Claude skill, connected to Salesforce and the ad platforms, that builds account-based campaigns in about three minutes instead of two to three days. Clay built an attribution agent in about two hours that found a $200,000 deal credited to a Google ad had actually begun with a free trial two months earlier. Vercel went from one or two field events a month to three to six, with no new headcount, using an internal app built with Claude Code.

So what

The pattern is marketers, not engineers, building narrow tools on top of their own CRM and call data. The Clay example should give pause to any CEO allocating budget on last-touch attribution. Pick one recurring GTM workflow that consumes days of agency or ops time and ask a marketer to prototype an agent for it, with a person approving output before anything goes live.

Confidence: Established fact — Results reported by named executives about their own teams; not independently audited.

Go deeper Growth Unhinged · TLDR Marketing

Operations & Finance

Airbnb's AI Assistant Now Resolves About 40% of Support Issues and Helped Cut Cost per Booking More Than 17% (Q3 2026, Reported November 10)

On Airbnb's third-quarter earnings call, CEO Brian Chesky said support cost per booking fell more than 17% year over year, "driven in large part by improvements to our AI assistant." The assistant resolved about 40% of all customer issues without a human agent in the quarter, and Chesky said customer service NPS is as high as it has ever been. It launched about 18 months ago and is now expanding into voice support in the US.

So what

This is one of the cleaner public data points on AI support at scale, and the metric choice is the lesson: unit cost paired with satisfaction, not tickets deflected. It also took roughly 18 months to get here. If you are funding AI support, set cost-per-transaction and NPS targets up front and budget for a multi-quarter ramp rather than a one-quarter win.

Confidence: Established fact — Figures stated by the CEO on an earnings call and reported by CX Dive.

Go deeper CX Dive · TLDR AI

Resources

Reference

SaaS Renewals Rise 8% to 12% a Year, and Buyers Who Start 90 Days Out Save 49%: A Renewal Checklist (November 11, 2026)

Software spend manager Cledara compiled published price changes for 14 vendors, including Microsoft 365 increases of 5% to 33% by plan from July 1, 2026, and Canva's 317% Teams increase in 2024. Its 2026 benchmarks put typical renewal increases at 8% to 12% a year, and 15% to 25% for aggressive vendors. Companies that negotiate about 90 days before renewal save 49% on average, versus 19% for those that start 30 to 90 days out.

So what

Use this two ways. As a buyer, ask finance to list every contract with its notice deadline, start negotiations a quarter early, and ask for increase caps and true-down rights. As a seller, assume your customers' procurement teams now run this same playbook, so any price increase tied to bundled AI needs a clear value story before the renewal conversation begins.

Confidence: Informed perspective — Cledara sells software management and does not publish its sample size or method.

Go deeper Cledara · TLDR Founders

Tools

How Agents Connect to Tools Changed Cost per Task About 17x in OpenRouter's Test of MCP, CLI and Skills (November 11, 2026)

OpenRouter ran one GitHub task across four models. On Claude Sonnet 5.5, the GitHub MCP server with its 46 default tools used 59,417 input tokens, against 1,421 through the command line and 3,315 with a skill, or roughly $127 versus $7.50 per 1,000 tasks. Limiting MCP to the two tools needed and capping result size cut usage to 5,357 tokens.

So what

If your product or internal agents rely on MCP connectors, unused tool definitions and oversized results may be a large share of the AI bill. Ask your team to measure tokens per task on your top agent workflows, load only the tools each one needs, and trim what tools return. The authors tested one server on one task, so benchmark your own setup before re-architecting anything.

Confidence: Established fact — A published test with its method disclosed, limited to a single task and server.

Go deeper OpenRouter · TLDR AI

Sources in this edition: The Neuron / Mindstream, The Rundown AI, The Deep View, Every (AI & I / How I AI), TLDR Founders, TLDR AI, TLDR Marketing

← Latest Editions and Search