How AI customers work: from sign up to evidence

An AI customer works in 6 steps: it gets its own inbox, phone number and browser, signs up through a company’s public forms, sleeps until the company sends something or a check is due, wakes, keeps what arrived as evidence, and reports each fact with a link to that evidence. Because it costs nothing while it sleeps and most sign ups run as programs, one system can hold millions of AI customers at once.

An AI customer has to do what a person does over weeks: hold an address and a number, sign up, wait, notice what arrives and remember it. The hard part is doing that at thousands of companies at once without paying for thousands of people, or for thousands of AI models thinking all day.

The design that makes it work is simple to state. An AI customer is asleep almost all the time, and the outside world wakes it: an email arrives, a text lands, a scheduled check comes due. Asleep, it costs nothing. That is why one system is built to run 100,000 companies at once, and built for millions. The term itself is defined in what AI customers are.

The life of an AI customer

1. Identity. It is given its own inbox, phone number and browser, and later the accounts it opens.
2. Sign up. It joins through the company’s public forms, the same ones any customer uses, and confirms by email or text where the company asks.
3. Sleep. It waits, with no process running, for as long as the watch lasts.
4. Wake. A new email, a text or a scheduled check wakes it.
5. Keep. It stores what arrived exactly as it arrived, with the time and, for email, the sender’s authentication.
6. Report. It turns what arrived into facts, each with a link to its evidence, and sends them where its owner works.

The identity: an inbox, a number and a browser

Each AI customer belongs to 1 company and 1 watch. Its inbox receives what a subscriber receives, and a Gmail inbox also shows which tab Gmail chose for each email. Its phone number receives texts and calls. Its browser opens pages, fills baskets and reaches checkout.

The parts already exist as services built for AI agents: AgentMail gives agents their own inboxes, and Browserbase runs browsers for them. An AI customer adds the parts those services leave out: a customer’s life at one company, kept over months, and a record of what arrived.

The identity is created for the job and plainly belongs to whoever runs it. It never borrows a real person’s name, address or card.

One AI customer per journey

A customer who left a basket gets different emails from one who bought, and both differ from one who only joined the list. So each journey gets its own AI customer: one for the welcome, one for the basket, one that buys where the owner funds an order, and one for each market where prices and messages differ. Each keeps a clean history, so every finding names the journey it came from.

Signing up, once per platform

An AI customer signs up where any customer would: the pop up, the footer form, the checkout tick box, the trial form. Where the company uses double opt in, it opens the confirmation email and clicks the link, as a subscriber must before being added.

Most companies sign customers up on the same few platforms, so a sign up learned at one company runs at every company built the same way. The first sign up on a platform is worked out step by step and kept as a program. Every company after it runs that program in seconds, and only an unusual form needs an agent to think.

Waiting costs nothing

A customer’s most useful information arrives late: the reminder a day after the basket, the win back email after 60 days, the renewal notice after 11 months. A system that keeps a process running for each customer pays for every hour of waiting. An AI customer pays for none of it.

Serverless runtimes make this possible. Cloudflare’s pricing for Durable Objects states that objects that are idle and eligible for hibernation are not billed for duration, and lists requests at $0.15 a million on the paid plan beyond the first million. An AI customer that wakes 20 times a month is 20 requests.

The count of AI customers is then set by sign ups and storage; model time barely moves it. Obsession can run millions of AI customers this way, and a company another customer already watches starts with its history on day one.

Waking: what it checks

Each wake is a comparison. The AI customer holds what the company promised (the pop up’s offer, the flow’s timing, the policy page) and checks what arrived against it.

Silence is a finding. The last row is the one no internal report produces. An email platform logs what it sent, and nothing logs the email a broken trigger never sent. An AI customer knows what it was promised and when, so a missing message becomes a dated fact.

The clock comes from the company’s own settings where they are public. On Shopify, a checkout counts as abandoned once it stays incomplete for more than ten minutes after the customer gives an email, the merchant chooses when the reminder goes, and a second abandoned checkout automation defaults to a 10 hour wait. An AI customer that left a basket at 9:00 knows when a reminder was due, and records when one came.

Keeping the evidence

Each email is stored exactly as it arrived: its headers, body, images and the time it landed. Bulk senders to Gmail must authenticate their mail under Google’s sender guidelines, and a DKIM signature, checked when the email arrives and kept with its key, shows which domain signed the message and that it wasn’t changed after sending (RFC 6376). The same details are visible to any Gmail user under Show original (Gmail’s header help).

• Emails and texts: the message as it arrived, its time, and for email its tab and verified sender.
• Pages: a capture of what the page showed, with the time and the address.
• Checkout: each step up to payment, with the total after codes.
• Chats and calls: the full transcript, with the time of each turn.

From evidence to facts

Evidence becomes facts a person or an AI agent can use: 1 sentence each, with a link to what proves it. "Welcome email arrived 2 minutes after sign up, Promotions tab." "Code IVY10 refused at checkout on the sale price." "Chat said 60 days for returns, the returns page says 30." A fact without evidence is never reported.

Where the facts go

Facts are useful where people already work. An AI customer system sends them to Slack, a CRM, a sheet, a report or a webhook, and to the owner’s own AI agent through the Model Context Protocol, the open standard that lets AI assistants call tools.

Obsession exposes 5 verbs over MCP and an API: lookup, facts, evidence, become_customer and notify. An agent asks which companies are already covered, reads their facts and evidence, starts AI customers at the companies that aren’t, and gets told when something new arrives (AI customers for developers).

Limits

• A bot wall or a CAPTCHA stops it. An AI customer never gets round one, and names the step it couldn’t test.
• A new customer sees what new customers get. Engaged, long standing customers can receive other emails and other inbox placement.
• Countries differ. Prices, offers and delivery messages vary by market, so a watch needs a number and an address in each market it covers.
• Some journeys need a real purchase. Delivery and returns messages arrive only after an order, which needs a card the owner funds.
• It reads outputs. Why a company sent something sits in its own systems.

Questions about how AI customers work