# Are Fortune 500 companies ready for AI agents as customers? | Enrich Labs > We pointed a browser-driving AI agent at all 500 Fortune 500 companies with instructions to buy something, stopping one click short of ordering. 96 let it reach the final order button. Research from Enrich Labs. _Source: https://www.enrichlabs.ai/research/can-ai-agents-buy-from-the-fortune-500_ --- ## Key findings 1. **An AI agent can already reach the point of purchase at 96 of the Fortune 500:** 19% of the full list, and 37% of the 258 companies where an agent-shoppable flow exists at all. At each of the 96 the agent found a product, carted it, completed guest checkout with a test card, and stopped at an enabled final order button by design; it never placed an order.1 2. **The biggest wall is not bot detection. It is the sign-in wall:** 84 companies stopped the agent by requiring an account, more than twice the 32 that stopped it with a visible bot check. Kroger’s only add-to-cart button is “Sign In to Add”; the Google Store’s only checkout is “Sign in & Check Out”; Disney parks tickets cart fine until a required MyDisney login appears. 3. **Nearly half the Fortune 500 has nothing an agent can buy:** 242 companies had no agent-shoppable flow in this attempt: 123 with no eligible online purchase at all, 104 whose flows sit outside a shopping benchmark (insurance enrollment, banking, quote forms), 14 unreachable during testing, and 1 excluded. 4. **Malfunction rivals defense:** 30 companies lost the agent to something nobody chose: blank product listings at Costco, a payment widget that never loads at Ford and Coca‑Cola, a checkout showing “$NaN” totals at UnitedHealth’s Optum shop. Add the 10 payment-validation stops and the last mile loses more buyers than bot walls do: inside checkout itself, page errors stopped 17 companies while bot checks stopped 6. 5. **Infrastructure flips outcomes:** rerunning the original 66 brand sites on different network infrastructure halved visible bot checks (28 to 14) and raised final-order-button reaches by half (20 to 30). Marriott, Hilton, and Vrbo flipped from blocked to buyable without the sites changing anything.5 6. **The audit cost $131 in model usage,** about 26 cents per company. Knowing exactly where your checkout stops an AI agent is now effectively free, and the agents are not waiting for permission.6 Searching Amazon for AA batteries brought up an interstitial saying that access by an unauthorized AI agent violates Amazon’s Conditions of Use. I stopped there and did not try to get past the block. AI The agent’s run log Amazon, Fortune 500 rank #1\* ## Overview Agentic commerce is moving from demo to default: AI assistants now browse, compare, and increasingly complete purchases on a user’s behalf. Which raises a question every large brand will face sooner than it expects: **when an AI agent shows up at your site with a card number and intent to buy, what actually happens?** To find out, we ran a browser-driving agent (starting from a fresh browser each time, with no saved logins) against all 500 companies on the 2026 Fortune 500 list.1,2 The task: confirm the company’s official site, find something a consumer could genuinely buy, add it to the cart, and proceed through guest checkout with placeholder details and a benchmark test card. A hard guard stopped every run **one click short of placing the order**: no sign-ins, no account creation, no CAPTCHA solving, no submitted purchases.3 ✦ Agent ⌂ Site → 258OF 500 **Stage 1:** Find the company’s official site and a flow a consumer can buy. **258** of 500 have one. 242 have nothing an agent can buy: no eligible online purchase (123), insurance, banking, or quote-form flows (104), unreachable (14), excluded (1). ✦ Agent ⌂ Site ◎ Product → 205OF 500 **Stage 2:** Search or browse to a specific, in-stock product. **205** get there. 53 stop at the front door: sign-in wall 29, bot check 19, page error 4, agent stalled 1. ✦ Agent ⌂ Site ◎ Product ◫ Cart → 177OF 500 **Stage 3:** Add it to the cart as a guest. **177** carted. 28 stop on the way: sign-in wall 22, bot check 2, page error 2, unavailable 2. ✦ Agent ⌂ Site ◎ Product ◫ Cart ☰ Checkout → 144OF 500 **Stage 4:** Proceed to guest checkout and fill placeholder shipping details. **144** reach checkout. 33 stop at the cart: sign-in wall 27, bot check 5, unverified 1. ✦ Agent ⌂ Site ◎ Product ◫ Cart ☰ Checkout ▤ Card ■ Guard → 96OF 500 **Stage 5:** Enter the benchmark test card, then stop at an enabled “Place Order” button. The guard never clicks it. **96** reach the final button. 48 stop inside checkout itself: page error 17, payment validation 10, sign-in wall 6, bot check 6, agent stalled 6, unavailable 2, unverified 1. Each row adds one step of the checkout; the circle counts the Fortune 500 companies where the agent completed that step. Every run stopped one click short of buying. The result is a company-by-company map of agentic buyability. 258 of the 500 have a flow an agent can attempt; 96 of those let the agent all the way to an enabled “Place Order” control. The rest stop it with sign-in walls, bot checks, broken pages, or payment validations, and the mix of those stops is the most useful part of the data: **it separates companies that decided to block agents from companies whose checkouts simply fail them**. 500 companies run end to end, every run completed and reviewed 96 reached the final order button; all verified by screenshot or manual review 84 stopped the agent with a sign-in wall, the single largest blocker $131 total recorded model usage for all 500 companies, about 26¢ each ## The funnel: where each company stops the agent Treat the 500 companies as one funnel and the shape is blunt: **half the list falls away before shopping even starts**, because there is nothing a consumer-facing agent can buy: industrial suppliers, insurers, energy producers, B2B service firms whose sites end in quote forms. Of the 258 companies with an assessable flow, the agent found the item at 205, carted it at 177, reached checkout at 144, and reached the final order button at 96. Exhibit 1 * * * From 500 companies to 96 final order buttons: most of the loss is structural, not defensive. **Companies remaining at each funnel stage,1** count and share of the Fortune 500, October 2026 1Each company is placed at the furthest stage its best reviewed attempt reached. Blocker counts under each stage name the reason the companies stopped there. “No eligible online purchase” describes this attempt, not proof that no purchase is possible anywhere in the company. Source: Enrich Labs agent-readiness audit, October 2026 Enrich Labs * * * Two readings of the same funnel matter. For agents: even against America’s largest companies, with no special integrations, a general-purpose browser agent gets to the point of purchase at **more than a third of everything that is buyable**. For brands: of the 162 companies that had a shoppable flow but stopped the agent somewhere inside it, fewer than half did so on purpose. ## What stops the agent: chosen walls vs. broken glass Sort the 258 assessable companies by outcome and the blockers split into two families. **Chosen walls**: sign-in requirements (84), visible bot checks (32), and payment validations such as 3-D Secure (10). **Broken glass**: pages and checkouts that error for no visible reason (23), flows the agent could not finish inside its budget (7), items that turned out unavailable (4), and two runs whose final state could not be verified. Exhibit 2 * * * Sign-in walls stop more than twice as many agents as bot checks do. **Outcome of the furthest attempt at the 258 companies with an assessable flow,1** count of companies 1One outcome per company, taken at the stage where the agent stopped. Payment validation (e.g., 3-D Secure or card pre-authorization the guard will not complete) is counted as a chosen wall. “Agent could not finish” = ran out of its 80-turn / 15-minute budget without a terminal wall. Source: Enrich Labs agent-readiness audit, October 2026 Enrich Labs * * * The sign-in wall deserves the attention. It is rarely an anti-bot decision; it is a loyalty and data decision made years ago, and it now quietly reads as “closed” to agents: **every account wall is an agent wall**, whatever its original intent. Bot checks, by contrast, behave exactly as designed: Walmart’s and Target’s press-and-hold challenges let the agent see the product and blocked the cart; Tesla’s Akamai config returned a 403 before the homepage rendered; Amazon alone stated the policy in writing, naming unauthorized AI agents in an interstitial. The broken-glass column is the one nobody is managing. The agent’s logs read like an unusually patient QA tester: Costco’s search and category pages rendering as blank placeholders; Ford’s guest-checkout payment widget failing repeatedly with a wheel-lock kit in the cart; Coca‑Cola’s payment step never loading; UnitedHealth’s Optum shop reaching guest checkout and then showing **a tax line and order total of “$NaN”**. These companies are not refusing agent revenue. They are dropping it, and most would fail a determined human the same way. ## The infrastructure effect: half of “blocked” was the plumbing The audit’s first edition, 66 brand sites run months earlier on our original network setup, looked much more hostile: 28 of 66 sites threw bot checks. Rerunning **the exact same shopping tasks on the exact same 66 sites** on upgraded network infrastructure cut visible bot checks to 14 and raised final-order-button reaches from 20 to 30.5 Exhibit 3 * * * Same sites, same tasks: changing only the network infrastructure halved bot checks and lifted completions by half. **Paired results on the original 66 brand sites,1** first-edition run vs. current rerun 1Both denominators are the same 66 brand sites (58 parent companies) repeating their original task text. The historical run was earlier and used different retries and earlier grading, so the deltas are directional rather than a measured causal effect of infrastructure alone; the current cohorts share the same setup and final-state review criteria. Source: Enrich Labs agent-readiness audit, October 2026 Enrich Labs * * * Travel shows the flip most clearly. In the first edition, Marriott, Hilton, and Vrbo all walled the agent; in the rerun, all three let it configure a real multi-night stay, enter the test card, and stop at the booking button. Expedia and Hotels.com threw their “Bot or Not?” challenge both times; Booking.com completed both times. The lesson cuts both ways: **much of today’s “agent blocking” is infrastructure inference rather than policy**, and any brand that believes it has decided to block (or allow) agents should verify that its anti-fraud stack agrees. ## Walls, errors, and open doors What “buyable” means varies wildly across the list, and the agent’s paths through the conglomerates are the most telling. At Berkshire Hathaway it worked from the holding company’s link page to See’s Candies and carted an $8 milk-chocolate turkey. At Exxon Mobil it bought the only consumer thing on offer, a $25 physical gift card. At General Motors and Boeing it reached order buttons only in official merchandise stores: a $6 keychain, a $2.99 sticker. Of the 69 expansion-cohort order buttons, **15 are company merchandise stores rather than the core business**, and the audit labels them separately.4 The clean completions are instructive precisely because they are boring. Apple, Home Depot, Best Buy, Nike, Microsoft, CVS, and Starbucks all let a guest assemble a cart, enter a card, and stop at an enabled order button with correct totals, in minutes, with no special treatment. **Agent-readiness today is mostly indistinguishable from a fast, guest-friendly checkout**: the same property that converts an impatient human converts a bounded agent. And one company answered the question nobody else has put in writing. Amazon’s interstitial did not challenge the agent to prove humanity; it told the agent that its access violated the Conditions of Use. That is a policy position on agentic commerce from the company with the most to lose from disintermediated shopping, and it will not be the last one. ## What should brands do? If you run marketing or e-commerce at a brand of any size, this audit suggests a short, concrete list: 1. **Decide your agent policy on purpose.** Today most brands’ stance toward shopping agents is whatever their bot-mitigation vendor defaults to, as the infrastructure flip shows. Whether you choose to welcome agents or wall them, make it a decision, not an accident of infrastructure. 2. **Treat guest checkout as the agent on-ramp.** Sign-in walls stopped 84 of the Fortune 500, more than every bot defense combined. If an account is required to buy, agents (and the humans sending them) will route to a competitor whose checkout is open. 3. **Fix the last mile first.** Inside checkout itself, broken pages stopped nearly three times as many companies as bot checks (17 vs. 6). An agent that reaches your payment step and watches it fail is logging, in plain language, a bug your human customers are silently hitting too. 4. **Run this audit on yourself.** At roughly 26 cents per company, pointing an agent at your own purchase flow, and your top competitors’, is the cheapest mystery shopping ever offered. The run logs double as QA evidence. 5. **Prepare for buyability the way you prepared for findability.** Being recommended inside AI answers was the first half of agentic readiness; letting the resulting agent complete the purchase is the second. The brands that win the channel will do both. If you want to compare notes on agent-readiness, see the underlying company-level evidence, or run the same audit against your own funnel, write to [seijin@enrichlabs.ai](mailto:seijin@enrichlabs.ai): questions, pushback, and collaboration ideas are all welcome. ## Method and limits An AI shopping agent visited each company’s public website in a fresh browser, with no saved logins, and tried to complete a purchase within a fixed time budget.2 It confirmed it was on the right company’s site before shopping and could follow official subsidiary and store links. Safeguards stopped it before any real order, inquiry, or account was created, and only placeholder details and a test card were used. The universe is the 2026 Fortune 500 list; the original 66 brand sites repeat their fixed shopping tasks from the first edition, while the other 442 companies first had to discover an eligible purchase flow, so results describe **observed coverage rather than a like-for-like ranking**.1 The honest limits: reaching a final order button does not test a live payment or guarantee a transaction would succeed. “No eligible online purchase found” describes one bounded attempt, not the company’s full commercial surface. Fourteen companies could not be assessed at all because their sites failed to load during testing. A small number of runs that tripped our safeguards were excluded from successes and reviewed before testing resumed.3 ##### Footnotes 1. Universe: the 2026 Fortune 500 list; rankings include ties. The original 66 brand sites map to 58 parent companies; each company row takes the furthest reviewed result among its tested brands, so the 66 sites do not add to the 500-company denominator. Data last updated October 19, 2026. 2. Each company received a limited number of attempts, and reports retain the furthest reviewed attempt. 3. No sign-in, account creation, CAPTCHA solving, or final order submission was allowed at any point. Financial applications and investments, insurance enrollment, clinical treatment, utility enrollment, regulated products, and excluded contracts are outside the benchmark; sales, demo, and quote forms and referrals to unrelated retailers do not count as checkout. 4. Expansion-cohort final-order paths: 42 product or standard service, 5 gift card, 15 merchandise or company store, 7 booking or reservation. Merchandise stores are labeled separately because they do not demonstrate that the company’s main industrial or financial business can be bought online. 5. First-edition run: 20 final order buttons, 31 checkouts, 28 bot checks of 66 sites. Current rerun: 30, 44, and 14 of the same 66. The runs used different testing conditions and grading, so the change is directional rather than a measured causal effect. 6. Total recorded model usage was about $131. All 96 final-order states are supported by screenshot checks or manual review, with zero awaiting review at publication. \* Quotes attributed to the agent are drawn verbatim from its run logs, with browser screenshots retained as evidence for every stage claim. They reflect each site as tested between September and October 2026. Written by Enrich Labs, from a 500-company agentic purchase-flow audit run between September and October 2026, including a paired rerun of the 66 brand sites from the audit’s first edition. The full company-level table, evidence screenshots, and CSV are available on request. Previously in this series: [When marketing runs itself](/research/when-marketing-runs-itself) and [Project FSD: Can an AI agent run a Shopify store?](/research/project-fsd-can-ai-run-shopify-store) Want to know where your own checkout stops an AI agent? [Get on the Agent Conversion Optimization waitlist](#waitlist).