Bots Are 54% of the Web, and Amazon Doesn't Want Them Buying
8 min read · 15 sources
- Bots account for 54% of global web traffic, up from 51% in 2024, and AI agent traffic grew 8x from August to October 2026.
- US applications for new information-sector businesses are up 60% since 2024, but new employer counts are flat and UK closures (23,645) exceeded openings (22,165).
- Amazon's $68.6B in ads generates all of its $34.4B retail operating income, which is why it blocks AI agents that skip upsells.
- AppSumo's founder found only 4 of 67 shipped changes moved a business number, and 92% of his work wasn't even for the paying business.
- Zendesk opened outcome-pricing early access on November 10, but 58% of 126 surveyed founders still tie AI pricing to the seat.
The software economy just split in two. On one side, bots now drive 54% of global web traffic and AI agents are getting banned from buying things at Amazon and United Airlines. On the other, founders are shipping 738 pull requests with Claude and discovering only 4 of them moved a business number. Both stories are the same story: the AI era makes production cheap and makes the old business models - the ones that lived on upsells and on shipping volume - fall apart.
Here is what broke, what it means for anyone running systems, and what the people who noticed early are doing about it.
Dell’s warranties produce 127% of its operating income, which is why agents that skip the extras get banned.
The software company boom is a registration boom, not a business boom
US applications for new information-sector businesses are up 60% since 2024, and the EU is up 37%. New employer counts are flat. In the UK, 23,645 information businesses closed last year while only 22,165 opened. Mack Grenfell’s analysis argues that “vibe coding” - the flood of AI-assisted development that started in February 2025 - made registering a company trivial and building a sustainable one no easier.
The crowding is concentrated at the cheap end. Median revenue at new subscription companies is down roughly 20% year-over-year, and the drop is steepest for products under $50 a month. UK registered companies grew only 5.5% - far less than the ~25% increase you’d need to explain the revenue collapse via competition alone. So the decline is not just more sellers fighting: buyers’ budgets are moving to OpenAI and Anthropic, which have more than tripled revenue this year. If you sell small-ticket software, your real competitor is a model subscription.
Amazon and United ban bots because their profit is in the upsell
Bots are now 54% of global web traffic, up from 51% in 2024, and monthly AI agent traffic jumped 8x between August and October 2026. Yet Tom Tunguz reports that United Airlines and Amazon are banning AI agents from purchasing. The reason is not security theater. It’s that their base products break even.
Amazon’s $68.6B in ads generates all of its $34.4B retail operating income. United’s fees top its $4.7B operating income. Dell’s warranties produce 127% of its operating income. AI agents skip sponsored listings, seat upgrades, and warranties because they’re optimizing for the cheapest flight or the best-reviewed product. A bot that buys the base product is a customer that costs money.
The options are blunt: raise base prices, charge agents a fee at checkout, or keep blocking. Shopify, which takes a flat 3.1% cut per transaction regardless of who clicks buy, opened checkout to agents in September. If you run a storefront that depends on ancillaries, the agent economy is a direct threat to your margin structure - and the fix is pricing, not a user-agent header check.
Atlassian's Story Factory: 19,440 pixels of stage, locked at 2:51 AM
Atlassian built “Story Factory,” an internal agentic creation studio, to produce the Team ‘26 Europe keynote. The output was a mega-bento wall across three LED screens totaling ~65 meters of stage and one 19,440-pixel-wide image. Content locked at 2:51 AM on version 54, about six hours before the 9am keynote.
The tool turns real front-end code and design tokens into production-grade marketing films, 3D motion, stage visuals, decks, and social assets in minutes, in any language, without screen recordings. The engineering challenges were not pixel-level design: they were scale (hundreds of assets with varied visual languages), change (cards swapped until the night before), and alignment across screens. It’s still in pilot. For anyone who has ever produced a keynote deck by hand, the takeaway is that the bottleneck moved from production to change-management.
Clay's code review is now about architecture, not bugs
Clay reports that coding agents became dramatically more reliable in mid-2026 - rarely writing bugs and producing decent first drafts with minimal guidance. So the company shifted code review goals away from bug-catching toward high-level concerns: “Is this PR solving the right things?” and “Is the architecture good?”
The new review focus areas are data schemas, infrastructure choices, service boundaries, and establishing new code patterns that will be repeated - the decisions that are hard to reverse. For complex changes, engineers discuss the design before reviewing code in detail. If a large PR arrives without that context, reviewers ask for a short design write-up and pause the review. The quiet consequence: a good helper function now helps future agents too, not just future humans. Review is a check on long-term maintainability, because the line-by-line correctness problem is largely solved.
Devin Apps: private web apps on a cloud computer that hibernates
Devin Apps are live, private web applications hosted on Devin’s cloud computer, built from a description and accessible via a stable URL with sign-in required. They run on memory snapshots, so Devin hibernates when unused and wakes on access - first load after sleep takes a few seconds. Billing is based on machine-awake time. They can use any HTTP-serving stack, databases like SQLite or Postgres, and call the Devin API.
Engineers should note the constraints: apps are single-machine, suited for internal tools and dashboards, not high-traffic production. Updates create new versions with rollback if a version fails to start, and automations can redeploy on a schedule or branch change. This is the “describe an internal tool, get a URL” workflow made real - but the single-machine ceiling means you shouldn’t point a marketing site at it.
Cloudflare gives PayGo users a kill switch for Workers spend
Cloudflare introduced Spend Management in Beta for pay-as-you-go customers, starting with Workers and Durable Objects. Users set spending thresholds and Cloudflare pauses activity when limits are reached. Controls work via dashboard, cf CLI, or REST API, with near-real-time cost visibility (under 10 minutes) and an audit log of changes.
Billing cycles start when the paid subscription begins, not on deployment. Spend Management works alongside AI Gateway spend limits for per-user and per-model budgets on Workers AI. For anyone who has received a surprise cloud bill at 3am, this is the guardrail that should have existed years ago - and the API surface means it can be automated.
Ask Claude which of your shipped work moved a number
AppSumo’s founder built 738 pull requests since September with Claude. When asked Claude to grade the work, only 4 of 67 shipped items significantly improved the business, and only 8% of the work was even for AppSumo, the business that pays for it. Two winning tests sat for two weeks because nobody made the call. His new rule: nothing ships unless he knows what number it’s supposed to move.
The prompt: “Look at everything I shipped since [date] and every message I sent you. For each change: what number was it supposed to move, did it move, and what did I never check? Rank what I should stop, keep and finish.” Building with AI feels productive because something ships every day. This is the audit that reveals whether shipping is the same as progress.
HubSpot refounds as the "Customer Harness Company"
HubSpot’s founders announced they are refounding as the Customer Harness Company, and SaaStr - with 3 humans and 20+ AI agents - describes what that means in practice. They are 100% headless on their CRM: agents use MCP and API, no human logs in. They moved off traditional marketing automation in the last 60 days, with agents building templates and personalizing emails across a 450k list, eliminating journey builders.
The argument: for B2B software, value is shifting from the UI to the data and actions behind it, and seat-based pricing is over for a growing share of customers. Agents handle CRM reads/writes and marketing execution. The UI part of software value “is headed to zero” for agent-first customers. If your product’s value is a nice interface, agents will not see it - they’ll see your API.
Zendesk tests outcome pricing, but 58% still pay per seat
Source: hsdigital.substack.com ↗
Zendesk opened early access to AI plans described as “Access without per-seat charges” on November 10. But a Q4 2026 survey of 126 vertical software founders found 58% still tie AI to the seat, and 69% list “agents that replace headcount” as part of their AI strategy. The pricing split: 44% include AI in existing seats, 14% add a per-seat uplift, 15% charge on outcomes, 15% on usage, 9% take revenue share, 3% don’t charge yet.
The catch: 28% say annual contract value hasn’t improved and 55% say net revenue retention is unchanged. Zendesk’s live price list still defaults to seats - 5-10 automated resolutions per agent per month, extras at $1.50 committed or $2.00 pay-as-you-go. Intercom charges $0.99 per outcome for Fin. The seat remains the default even among vendors testing alternatives.
Dumb ways to manage inference spend
Source: supercompanies.substack.com ↗
Section spends roughly $1,000 per month on AI inference per nontechnical employee - about $300K total, with engineers using 2-3x more. The projection: $120M annually at a 10,000-employee company as agents become more common. The post criticizes two common controls: giving everyone the same AI budget (limits productive users while wasting budget on light users - set budgets by work complexity with at least 2x for engineers), and shutting off access when employees hit spending caps (disrupts work mid-month). Inference spend is becoming a major operating expense, and flat budgets or hard caps are blunt tools that penalize the most valuable AI usage.
Most AI moats are just fuel
Source: adlrocha.substack.com ↗
The author revisits his March argument that context was the moat in AI, now concluding that intelligence has become a commodity and context is no longer durable. Subscription tokens are discounted because users’ corrections and decisions feed labs’ training loops - users are effectively selling their know-how. The real moat may shift to the “harness”: vertical integration of model plus tools. Context shared through subscriptions becomes the labs’ property, not the user’s.
Build when customer requests cluster
A founder at a company past $100M run rate says they still don’t have a product roadmap. Instead, if 5-10 customers from completely different verticals ask for the same thing in the same week, they build it immediately - which has happened at least once a month since starting. No quarterly commitments, no annual planning beyond this cluster-based prioritization. It’s a demand-driven development model that prioritizes cross-vertical signals over scheduled roadmaps.
You May Also Like
SaaS vendors add agent tollbooths while foundation models break seat pricing
Salesforce, HubSpot, and Atlassian are aggressively moving away from per-seat subscriptions, hitting automated agent workflows with annual API surcharges …
The New Valley: Why $5M-$25M ARR Is Where AI Startups Go to Die
Series A financing for AI startups has hit a wall. Carta data shows just 15.4% of Q1 2022 seed companies reached a Series A within two years, down from 30.6% …
Agents now outnumber humans online - and they're buying
AI agents now make up the majority of web traffic, write over half of new articles, and generate most production code, according to a16z. Shopify reports orders …




