Cox Communications drives a 7x return on AI across its B2B funnel with Claude and Accenture

A growth engine that had run out of room

Adopting AI across a company of this size was a challenge in itself: standardizing on tools teams could trust, getting people who had never written code building safely, and deciding what was worth building at all. Leadership set three priorities: revenue growth, customer experience, and cost. Growth was the top priority, and the hardest to deliver.

In the B2B segment of the business, the growth had slowed and was ready for a new playbook. For years, Cox had pushed the same broad campaigns to everyone. "The one-to-many marketing approach was no longer delivering the growth we needed," said Sarah Kim, Head of Marketing for Cox Business. "To meet our revenue ambitions, we had to shift to a more targeted, personalized approach and orchestrate engagement seamlessly from awareness through close. AI became essential to operationalize and scale this new go-to-market model across marketing and sales."

Field sales ran on fairly manual effort. A seller heading to a street of small businesses would open Salesforce, research each prospect by hand, and assemble a prep sheet, sometimes printing it to carry to sales meetings. The data underneath was the real problem: lead lists ran only about 13% accurate and were expensive to build, which sent sellers and marketers chasing prospects that went nowhere.

Eric Pace, who leads Cox's AI Center of Excellence, had been fielding the same request from leadership for five years. "We want actual marketing automation," he said, "not automated steps in the flow, not more SaaS applications that make it easier for humans to do what they're already doing." Until recently, the building blocks for that kind of automation did not exist.

 A trusted partner, a default model, and a stack they could monitor and control

To rebuild its B2B marketing and sales, Cox partnered with Accenture, a strategic transformation partner already embedded across the company. Before any product, the two settled the operating model: how execution would run between the traditional IT teams and the new AI Center of Excellence. "We had to make the operating model concrete, down to almost a task level, across the AI product development life cycle,” said Katherine Mohrig, who leads the Cox engagement at Accenture. “vs. high level, directional hub-and-spoke.” 

Accenture's approach was to expand the foundation Cox already had rather than impose a top-down program, seeding champions across the company's different technology stacks and training teams to train others. From there, Accenture built the products: the marketing and sales agents, and the harnesses with context and data systems. Cox steered the work to its highest-value priorities.

Claude Code changed how fast that building went. "Only about 25% of our time is spent on actual development," Mohrig said. "The majority of our effort is data enrichment, validation and cleansing, security gateways, and partnering with the business for testing and buy-in."

A default model, accessed with control

Cox runs primarily on Amazon Web Services (AWS), and Claude was the natural default. The team's side-by-side comparisons on its own use cases confirmed it. "If you're going to work in AWS, use Anthropic by default," Pace said. "The outputs are more aligned to the business problems we're trying to solve, and they align to our responsible AI principles." For a 128-year-old family business, that responsible-AI focus mattered as much as performance. "We do good, we don't do harm," said Mark Greatrex, the company's president.

Cox accesses Claude through Amazon Bedrock, which Matthew Shorts, who runs the infrastructure organization at Cox, chose because it "already extended our existing footprint, and it fit very well into our security and observability principles that we already had across the organization." Amazon Bedrock let Cox apply the same tooling and governance to Claude that it used for its other models, including the open-weight models it runs and fully controls on-prem. "We have a strong runtime security requirement, so we can view all calls in and out and potentially jump in and stop if it starts to drift from standards," Shorts said, whether that meant straying off-brand or against the company's data-protection rules. Building on Amazon Bedrock kept Claude inside the observability and guardrails Cox applied everywhere else.

Marketing: the CAMI agents

With the model and its controls settled, the work turned to the products themselves. The first went to marketing and sales, a family of agents the team named "ami," French for friend. Cox's B2B unit markets across 120 micro-industries and eight customer segments, a level of personalization the legacy one-to-many campaigns could not reach. Working from a shared knowledge base, the marketing member of the suite, CAMI (Content and Audience Mapping Intelligent assistant), generates the campaign brief, plans the work, and produces content tailored to each micro-industry and segment, factoring in how many products a customer already has, how they use them, and what competitive intelligence suggests about their area. 

A single campaign can produce roughly 32 pieces spanning marketing content and sales collateral, all of it flowing into Salesforce, for 50-100 campaigns annually. The B2B marketing team runs its campaigns through CAMI, and the personalized content began lifting conversion where the broad approach had stalled and getting campaigns to market faster.

Sales: SAMI, from prep to order fulfillment

Sales came next. SAMI, the sales side of the suite, picks up where CAMI leaves off, running on Claude through Amazon Bedrock like the rest of the ‘ami’ agents. SAMI can help proactively identify high priority opportunities for a seller, and then help them pursue it. Where a seller once built a prep document, a call script, or a customer email by hand, SAMI assembles it automatically: account history, product usage, prior call-center sentiment, propensity to buy, and a recommended, micro-industry-specific script, alongside competitive intelligence pulled from public signals. 

New capabilities in development let a seller dictate notes between visits and have the system enrich the opportunity and update Salesforce. The suite reaches further down the funnel, too, helping order-fulfillment specialists organize and execute their work. All of it shares the knowledge base and connects to Cox's tools through the Model Context Protocol (MCP).

A cleaner data layer: The Cube and lead validation

The agents were only as good as the data beneath them, so the team rebuilt that layer too. The result is an AI-curated data architecture, six-layer model Cox calls The Cube. The lower layers references the classic machine learning the company already has, and adds new signals and feature stores. The upper layers, Pace said, are "where we inject Claude models behind the scenes to drive insights or intelligence that we couldn't get before." 

On that foundation, Cox built data products to fix specific problems, starting with the unreliable, expensive-to-build lead lists. A validation pipeline now uses Claude to research and verify each lead, turning messy lists into clean, usable records. To keep costs down, Cox matches the model to the task: Haiku for retrieval and summarization, Opus for the heaviest reasoning, and Sonnet in between.

Citizen development with Claude Code

The enterprise products were only part of the build. Cox opened Claude Code to its broader workforce, and adoption ran well past engineering to roughly 2,500 people, more than half of them in non-technical roles. Employees built their own agents with Claude Code, separate from the enterprise systems built with Accenture. The marketing team, for one, used it to fill the gaps those systems left. Cox kept that work governed, with guardrails and open visibility into usage and spend, rather than letting it become shadow IT.

An agentic engine for the whole funnel, now in pilot

The most ambitious system is still in pilot. Cox and Accenture are building an agentic go-to-market capability that runs from a prospect's first signal all the way to contract close. It generates and places ads through Amazon Ads, builds a tailored landing page for each prospect, and moves the opportunity along on its own. Rather than script every step, the team gave the agent room to define its own process. It tunes itself against goals like qualified-opportunity conversion and velocity, and carries lessons forward through persistent memory. Humans don't run the funnel; they validate, oversee, and audit it. 

Guardrails were added around brand, action, profitability, and spend. "You can't spend $10,000 on Amazon ads to acquire this customer who will spend $1,000 total," Mohrig said. “But the agent can choose to place an ad if it believes that will further its goals, and it will generate content for that ad.”

To capture the judgment that used to live only in people's heads, Accenture also built a "Skill Studio" where business users document tacit knowledge as reusable skills the agent calls on when it needs them.One of the skills includes how to interpret competitive data and formulate actions from it. The majority of these skills are optional, meaning the agent can choose to leverage them but is not required to. The agent can also modify skills over time as it optimizes its actions and learns based on its results.  

Cox is running the pilot on a broader segment of small business buyers who purchase a single product digitally, a slice challenging to serve with dedicated human attention. The target is on the order of 1,000 new customers in four to six weeks. Pace frames it as a chance to "push the envelope and test our chops" on what an agentic system can do.

Tangible returns, and a template to scale them

The clearest proof is at the top line. Cox saw a 7x return on its first year of AI investment, a number Greatrex's team tracked deliberately by building business cases and reading back the ROI. That result, and the ones beneath it, kept the company investing even as other discretionary work was trimmed.

Underneath the top-line return are the operational gains. The marketing suite drove 55% faster speed to market and saw an early indication of a 2-point lift in conversion, as well as 40% higher efficiency. Cox redeployed that capacity against new growth rather than cutting. Lead validation got cheaper and more reliable at once: the cost to validate and enrich a lead fell by 86%, accuracy climbed from 18% to 97%, and each lead now processes in under a minute, so sellers spend their time on prospects worth pursuing. On the sales side, SAMI cut the work a seller spends preparing for a call from hours to minutes.

The cost of running these systems is real, and so far the value has outrun it. "Every time we show up, we say, 'Yes, this sounds expensive, but look at the value it's producing,'" Pace said. "And they say, 'Okay, keep going.'"

Iteration sped up too. "You're not doing traditional A/B testing where you do a hundred a year," Shorts said."You're doing a hundred or more a day, because you've been able to make those changes that much quicker and with less risk."

The agentic go-to-market pilot is meant to prove a pattern the company can repeat across every function. "It's a blueprint we can apply everywhere," Pace said, "ultimately driving toward our goal, which is to make the company 10x what it is."