Customer Identity and Access Management (CIAM): from Homegrown Login Pages to AI agents

Every web team built its own login. A users table, an email column, a password column, a reset link sent by email, and later a hash function someone hoped was strong enough. The sign-up form often asked for everything at once because marketing wanted the data.

Federation for consumers arrived as an experiment. OpenID 2.0 was declared final in December 2007, letting a person prove control of an identifier without the site holding a password. Facebook Connect opened to every website in December 2008 and turned the social network into a login button.

What broke was everything around the password. People reused the same one across dozens of sites, so one breach unlocked many. Reset flows were the softest path into an account, and nobody measured how many customers abandoned a form they could not finish.

Customer identity became a product you bought. Janrain, Gigya, LoginRadius, Auth0, ForgeRock and Ping sold hosted login, social sign-in, user profiles and APIs, priced per monthly active user. Cloud platforms followed with developer services: Amazon Cognito launched in July 2014, and Microsoft offered Azure AD B2C until it stopped selling it to new customers in 2025.

This is the era I built in. I founded and created LoginRadius, a CIAM platform I scaled to over a billion users. The lessons are in building customer identity at scale: conversion, uptime at peak, consent, and data residency decide a customer login far more than they decide an employee one.

Regulation turned identity data into a liability as well as an asset. GDPR applied from May 2018 and CCPA took effect in January 2020, so consent records, preference centres and data export became core CIAM features. The big platforms consolidated: SAP agreed to buy Gigya in 2017 and Akamai acquired Janrain in January 2019.

Machine learning moved into the login path as a risk score. Adaptive authentication weighed device, network, location and velocity, and stepped a sign-in up to a second factor only when something looked wrong. Bot management sat in front of sign-up and login to filter credential stuffing and fake account creation.

The credential problem stayed. In 2023, 23andMe disclosed that attackers had logged into about 0.1% of accounts using passwords reused from other breaches, then reached far more profiles through a relatives feature. A login let reused passwords in, and a sharing model multiplied the damage.

Passkeys gave consumers a phishing-resistant way out when Apple, Google and Microsoft committed to them in May 2022. The market consolidated again: Okta closed its $6.5 billion purchase of Auth0 in 2021, and Thoma Bravo folded ForgeRock into Ping Identity in 2023. I track who survived and who did not in the CIAM graveyard.

The new customer at the login page is an AI agent. A person asks an assistant to find, compare and buy, and the agent needs to sign in to the store, read the cart and pay. The first agents did this by driving a browser with the customer's own password or session, which looks exactly like the account takeover CIAM was built to stop. I cover where this is heading in agentic commerce.

The payment networks moved first. Visa announced Intelligent Commerce and Mastercard announced Agent Pay with agentic tokens in April 2025. In September 2025 Google published the Agent Payments Protocol, which records user authorization as signed mandates, and Stripe and OpenAI released the Agentic Commerce Protocol with Instant Checkout in ChatGPT. All of them answer the same question: did a real customer authorise this specific purchase?

CIAM vendors are adding the identity half. Auth0 for AI Agents went GA in November 2025 with a Token Vault for third-party tokens and asynchronous approval, so a customer confirms a consequential action on their phone. Stytch, Descope and WorkOS ship OAuth 2.1 authorization for MCP servers so a customer's agent gets its own scoped token. Ping's Identity for AI includes detection that separates personal AI agents from other bots.

The other agent in the room works for the company. Support bots and agents now sit in the account recovery path, which is where social engineering already wins. In May 2025 Coinbase disclosed that bribed overseas support staff had copied customer data used to impersonate the company. My view, set out in account recovery design, is that recovery sets the real security ceiling, and an agent with recovery powers raises the stakes.

My bet is that by 2031 the most important customer on many storefronts will not be a person typing into a form. It will be that person's agent, holding its own short-lived token, scoped to a budget and a merchant, with a signed record of what the customer approved. I expect a store that cannot authenticate agents to lose that sale.

I expect passkeys to become the default human credential, with passwords as the fallback and then the exception. Recovery gets redesigned around strong proofs instead of support calls, because both attackers and helpful agents will push on it. I made the longer case in the future of CIAM, where delegated access matters more than the login screen.

Verifiable credentials and wallets change what sign-up means. The W3C published Verifiable Credentials 2.0 as a Recommendation in May 2025, and the EU requires every member state to offer a digital identity wallet by the end of 2026. I expect sign-up to shift from typing data in to presenting a verified claim, such as age or residency, without handing over the whole profile.

What has to be true: merchants, payment networks and CIAM vendors must agree on how an agent proves delegated authority, bot defence must learn to admit good agents, and consumer protection rules must say who pays when an agent buys the wrong thing.