Inside Cox Automotive’s drive to engineer agentic AI

Most enterprises don’t yet have a playbook for how to architect, let alone deploy, their agentic AI solutions. After all, it’s still early days for this nascent class of autonomous bots, many of which are capable of executing multiple workflow steps unsupervised.

Count Cox Automotive among the outliers.

Under the guidance of Chief Product Officer Marianne Johnson, the $10 billion automotive services and software maker is turbocharging its software development with agentic AI. In early March, 100 of Cox’s 680 agile scrum teams will transition to an agentic development model.

“It’s a massive change of roles, responsibilities and skillsets,” Johnson tells CIO.com ahead of the CIO 100 Leadership Live Atlanta conference March 5, where Johnson will participate in a panel discussion on tech-enabled business transformation.

Cox’s “AI first” initiative comes at a time when most companies are still taking baby steps implementing agents as a part of their automation arsenals. Of the 23% of organizations that are scaling agents somewhere in their business, most are only expanding the deployment and adoption in one or two functions, McKinsey research finds.

Hammering the pedal on AI

Under Johnson, Cox, which operates Autotrader, Kelley Blue Book, Mannheim and other automotive brands, is preparing for a future in which humans “co-work” with agents to improve how software is built, refined and run.

To that end, Cox has built a reference architecture for how it plans to adapt to agentic AI development. It has shifted to specification-driven development, which entails writing a detailed spec document to refer to as the source of truth as code is created.

Although hardly new, spec-driven development is gaining renewed interest because as more organizations use AI to generate code, the spec can serve as a reference point for the LLMs, as well as the humans reviewing their output, to validate against.

Cox works closely with both Amazon Web Services and Anthropic to preview and test new AI tools before they’re broadly available. To build agents, Cox’s software developers are using GitHub Copilot, Anthropic Claude Code and Claude Enterprise, as well as Amazon Web Services’ Bedrock AgentCore development environment.

But Johnson says it’s incumbent on her and her teams to continually analyze the AI ecosystem for emerging tools that may better help Cox achieve its outcomes. “Everybody is trying to be that platform provider and the reality is [companies] are going to anchor in on a couple vendors,” Johnson says.

Johnson expects that 100% of Cox’s development teams will operate in this new agentic engineering model by the end of 2026, ensuring speed to value for code creation.

The fastest cars need the best brakes

Speed to value is a noble goal, but the intense work rides along some critical guardrails Johnson and her leaders began installing as early as 2024.

As Cox learned how to work with agents, its engineers built an evaluation tool to monitor models’ tendencies to “drift,” or lose focus and context—the kind of wayward activities that can lead to hallucinations, wholesale data deletions and other catastrophic issues.

Today Cox employs a risk-based model that determines when agents execute tasks autonomously and when humans will remain in the loop to oversee other jobs. Cox also has agents watching agents and triggers that alert humans to anomalous activities.

Cox isn’t handing the keys to its software pipelines to the machines. The company is still holding humans accountable for code that gets put into production, Johnson says. And despite Cox’s keenness to lean heavily into agentic engineering, that could be the case forever, she acknowledges.

“Who knows if we’re ever going to feel 100% comfortable that we know what has gone in isn’t corrupting existing code and is getting to the outcome we need?” Johnson says.

Product plus IT equals tight alignment

Johnson enjoys the distinct advantage of having a broad purview in shepherding Cox through its AI transformation.

As her title implies, Johnson oversees product strategy and management spanning four portfolios, but her remit also includes IT in its entirety. That is, software engineering and data science, as well as traditional back-office IT functions such as the company’s Workday, ServiceNow and Salesforce implementations. Cybersecurity rounds out the mix.

It’s an unusual arrangement, but Johnson says that linking product and engineering ensures alignment while eliminating gaps that can arise in organizations where the functions are distinct.

“The amount of speed to trust, the amount of alignment that you get from that organizational structure allows you to be a nimble organization and innovate at scale,” Johnson says. “It’s getting to value faster in a more reliable, sustainable way.”

Johnson will discuss her approach to IT leadership as well as Cox’s tech-enabled business transformation at the CIO 100 Leadership Live Atlanta conference March 5.