What to build with AI customers: 10 production builds

Developers can build 10 kinds of production system on AI customers, agents that stay customers of other companies, from sign up and login code tests on every deploy to competitor feeds, prospect facts in Clay and AI disclosure checks. Each calls a few of 5 verbs, lookup, facts, evidence, become_customer and notify, and sends dated facts with their evidence to a test run, a webhook, Slack, a CRM or Clay.

Engineers and technical founders who want agents to sign up, wait and report what companies send their customers need a customer that lasts. What AI customers are defines the term, why developers need AI customers covers the engineering behind them, and AI customers for developers prices the parts.

Each build has 4 things a team decides before writing code: what fires it, which calls it makes, which facts it returns and where those facts land.

What an AI customer gives an agent

An AI customer is an AI agent that becomes a real customer of a company, with its own inbox, phone number and browser, and keeps what the company sends and shows it as evidence. An agent on its own works for minutes inside a session. An AI customer adds the parts a customer relationship needs.

• An identity: an inbox, a phone number and a browser that a company treats as a customer, created for the job.
• Time: it stays a customer for weeks or months, and wakes when the company sends something or a check comes due.
• Evidence: every email, text, page and transcript kept as it arrived, with the time and, for email, the DKIM result checked on arrival.

An AI customer system exposes this as a few calls. Obsession’s developer page names 5 verbs, offered over MCP, an API and an SDK, and each of the 10 builds uses some of them.

• lookup: which companies the shared record already covers, and since when.
• facts: what a company did, as facts, each with its evidence.
• evidence: the raw email, page or transcript behind a fact.
• become_customer: start AI customers, or join ones already there.
• notify: new facts pushed to your webhook, Slack or CRM as they land.

An agent reaches them through MCP, the open source standard for connecting AI applications to external systems, and code reaches them through the API or a webhook. The same build can then serve a CI pipeline, a product feature or a sales agent. How AI customers work explains the sign up, the sleep and the evidence step by step.

What to build with AI agents: 10 builds

1. Sign up, login codes and magic links on every deploy

A deploy hook starts a fresh AI customer with a new inbox and a real number. It signs up, enters the login code from the text, follows the sign in link in the email and reads every email the release sends, then posts a pass or a fail the pipeline can act on before the launch email goes. It meets the product as a new customer does, outside the team’s test accounts. Run it against production after each deploy as well as staging, because each environment can send through its own email and SMS providers.

Facts: “Login code by text in 8 seconds”, “Sign in link arrived in 6 seconds and opened the account”. A pass: every code within 60 seconds and every link resolving. Website audit runs this after every release, and Delivery monitoring follows the first 14 days in each country.

2. Shopping agent readiness

AI shopping agents now read stores directly. Shopify’s help centre says agentic storefronts are active by default for eligible stores, and OpenAI’s shopping help describes an Instant Checkout option for some eligible products and merchants. The build checks what an agent meets: each top product’s page, price, stock and returns policy as a customer sees them, and whether a basket reaches checkout.

AI answers about price can drift from the page. In a study Product.ai ran and Business Insider reported, 85% of 913 checkable price answers matched the current or listed price. Facts: “Page price £52, AI answer £64”, “Basket reached checkout”. A pass: page, feed and answer agree on price and stock. Agentic commerce for merchants covers what to fix.

3. Price and offer agents

A pricing agent that responds to rivals needs what rivals show customers, including the codes they send only to subscribers. Each morning the build checks each rival’s prices, offers and delivery, and the AI customer on each rival’s list wakes when a code arrives. The agent drafts a response and a person approves it, and each proposal carries the email or page that prompted it.

Facts: “Pro plan cut from $49 to $39 a month”, “Subscriber code SAVE20 arrived, 20% off”. A pass: every change reported the day it happens, with a before and after. Price watch runs it as a recipe.

4. Competitor feeds in a marketing tool

In a marketing tool, rival tracking becomes a product feature. Users name their rivals, and the app shows every email, text, offer and price change those rivals send. One AI customer per rival, shared across every user who watches it, gives a new user the history on day one. The app asks lookup first, reads facts for the history, and starts AI customers for the rivals nobody watches yet. A notify webhook then posts each new fact as JSON, keyed so a redelivered event is stored once.

Facts: “4 emails this week, 2 with offers”, “Welcome offer changed from 10% to 15%”. A pass: each new email in the feed within minutes of arriving. Competitor tracking is the recipe behind it.

5. Prospect facts in a sales agent or Clay

Agents already act as buyers to test companies. Clay’s team ran agents that submitted 6,346 demo and contact forms and waited to see who replied. The build puts the customer side of each account on the record: for every new row, an AI customer joins the newsletter or starts the trial, keeps what arrives, and the gap lands next to the contact.

Facts: “No welcome email in 7 days”, “Opted in to texts, none arrived”. A pass: a verdict only once the sign up is confirmed, the spam folder checked and the full window watched. Prospect intelligence writes the verdict and a proof link back to Clay.

6. Customer-side tests of your own support bot

The UK’s Competition and Markets Authority says a business is responsible for what its AI agent does in the same way it is responsible for an employee. The build asks the bot what customers ask, on chat, email and the portal, each morning and after every update to the bot, its prompts or its knowledge base, and scores each answer against that morning’s policy page.

Facts: “Chat said 60 day returns, the policy says 30”, “Handoff to a person in 4 minutes”. A pass: every answer matches the policy page. Support bot check runs it daily, and how to test your store’s chatbot gives the questions for a check by hand.

7. Inbox placement monitoring

Email teams need to know where their own sends land. AI customers subscribed to the company’s list on Gmail inboxes record the Gmail tab each campaign and each account email arrives in, with its DKIM result. Gmail’s sender guidelines require senders of 5,000 or more messages a day to Gmail accounts to authenticate with SPF, DKIM and DMARC, require every sender to keep the spam rate reported in Postmaster Tools below 0.3%, and recommend keeping it below 0.10%.

Facts: “Black Friday launch: Promotions, DKIM pass”, “Order confirmation: Primary, DKIM pass”. A pass: authentication passes and each send lands in the tab it was meant for, with a notify call the day one doesn’t.

8. Partner and affiliate code checks

Partners keep sharing codes and banners after an offer ends. Each Monday the build visits every partner’s page as a customer, follows the link to the store, checks the partner’s tag survives the redirect, and tries the code in the basket.

Facts: “SPRING15 expired 12 days ago, still on 3 partner pages”, “Link lost its affiliate tag on redirect”. A pass: every live code works and every link keeps its tag. Partner checks drafts a note to each partner for the team’s OK.

9. Trial teardowns

Product and sales teams want to see how a rival brings a new user in. The build starts each rival’s free trial with its own inbox and browser, signs in like a new user and logs every onboarding email, prompt and offer until the trial ends. Where a trial asks for a card, it uses one the team funds.

Facts: “Day 3: 20% off annual if upgraded within 48 hours”, “Day 13: trial ending email”. A pass: every email of the trial logged in order, to its last day. Trial teardown is the recipe.

10. AI disclosure checks

Article 50 of the EU AI Act requires AI systems that talk to people to be designed so those people are told they are dealing with an AI, unless it is obvious, and the Act applies from 2 August 2026. Each morning the build asks every channel, chat, phone, email and text, “Am I talking to a person?”, and times the wait for one.

Facts: “Chat: AI notice in the first message”, “Phone: no notice before the first answer”. A pass: an AI notice first, a truthful answer when asked and a person within the target time. AI disclosure check keeps a dated record of every check.

Ways to build them, compared

Obsession makes AI customers, so every other option is described from its provider’s own pages as of 24 October 2026.

Testing tools and synthetic monitors suit builds 1 and 6 on a team’s own product, run from its own test code. Archives suit browsing rivals’ campaigns. Builds 3, 4, 5 and 9 need a customer at many companies for months, and the waiting and evidence they rely on are covered in why developers need AI customers.

Limits

• A bot wall or a CAPTCHA stops an AI customer, and the report names the step it couldn’t test.
• A new customer sees what new customers get. Long standing customers can receive other emails and other inbox placement.
• Some builds need a real purchase, on a card the owner funds.
• It records what a company sends and shows. Why the company sent it sits in that company’s own systems.

Where to start

1. Pick the build closest to a decision someone makes every week, such as holding a release or answering a rival’s price.
2. Write its pass mark before the first run.
3. Decide where its facts land: the CI run, a webhook, Slack, a CRM or Clay.
4. Start with 5 to 10 companies and 1 channel, then widen it once the pass marks hold.

Questions about what to build with AI agents