The project isn't the AI.

Start with the numbers, because they do the work

Cox Communications is the largest private broadband provider in the US: 15,000 employees and six million customers. Its business arm was stuck. One-size-fits-all campaigns had stopped producing growth, and salespeople were researching prospects by hand from lead lists that were 18% accurate.

They worked with Accenture and with Claude through Amazon Bedrock. These are the reported results:

• 7xreturn on the first year of AI investment
• 86%lower cost to validate a lead
• From 18% to 97%lead list accuracy
• 55% fasterspeed to market in marketing

Prep time for a sales call fell from hours to minutes, and about 2,500 people use Claude Code, more than half of them outside technical roles.

At this point every manager reading it wonders the same thing: which model did they use, and what did it cost. That's the wrong question.

The quote nobody will quote

Deep inside the case is a line from Katherine Mohrig, who leads the engagement at Accenture:

“Only about 25% of our time is spent on actual development.”

The other 75% went into data, validation and security.

Think about what that means. In the project that returned 7x and took accuracy from 18% to 97%, a quarter of the work was building AI. Three quarters was the work nobody films for a launch video.

And most organisations budget exactly the other way round.

Why it happens, and why it will keep happening

This isn't a local planning failure. There's a structural reason.

Model quality has become a commodity. You, Cox and your competitors have access to exactly the same models, at roughly the same price, from the same vendors. Nobody wins today because they have a better model.

So the competitive advantage has moved. It used to sit in access to the model, and it has moved to two other places:

• What you feed itYour data, in a format the model can work with, clean enough to trust.
• How you validate itKnowing the output is right before it reaches a customer.

Neither can be bought off the shelf. Both are specific to your organisation, and they are exactly the 75%. A model is a commodity. Your data isn't.

What Cox actually built

Look at the list, and notice how little of it is “AI”:

• The CubeA six-layer data model. This is the foundation: before the model saw anything, someone decided what a customer is, what counts as an interaction and what an account history looks like.
• A lead validation pipelineA layer that checks and enriches every lead before it enters the system, in under a minute per lead. That's what moved accuracy from 18% to 97%. Not the model, but the decision to build a whole validation stage before trusting anything.
• CAMIThe marketing assistant, producing campaigns for 120 sub-industries and eight customer segments.
• SAMIThe sales assistant: prep documents, call scripts, emails, and Salesforce records enriched from voice notes.
• Skill StudioA system that captures tacit business knowledge and turns it into reusable agent skills.
• Governance through BedrockSo everything sits inside the existing security and monitoring principles.
• A model per taskHaiku for retrieval, Sonnet for mid-weight work, Opus for heavy reasoning.

Most of the list is infrastructure, data or governance. Two items are applications, and one is cost optimisation. The model itself isn't on the list at all. It's a given.

The number that matters most

Of all the numbers, one stood out more than the rest. A Cox executive described the new pace this way: not a hundred A/B tests a year, but a hundred or more a day.

The one thing AI didn't bring down to zero is the number of rounds. Knowledge got cheap, judgment stayed expensive, and judgment is only built through a loop of trying, failing and correcting. Cox shows what it looks like when an organisation industrialises that loop.

Notice the connection: they didn't reach a hundred tests a day because they bought a fast model. They got there because they built the data layer, the validation pipeline and the governance, and that's what let them run experiments quickly without breaking things.

The 75% isn't overhead. It's what buys the speed of the rounds. An organisation without an orderly data layer can't run a hundred experiments a day, because it has no way of knowing which ones worked. It runs three, argues about the results and stops.

One detail worth stopping on

Skill Studio, the system that captures tacit business knowledge and turns it into reusable skills.

The philosopher Michael Polanyi argued that tacit knowledge, a veteran salesperson's sense that a deal is slipping, can't be written down, and so can't be loaded into a model. That still holds, and Skill Studio sharpens it rather than contradicting it: you can't load judgment, but once you've been through the loop you can encode what it produced.

It's the difference between handing someone a map and handing them a map drawn by someone who walked the ground. The map doesn't replace the walk, but it turns one person's walk into an asset of the organisation. It also explains why you can't buy it: the knowledge captured there belongs to Cox, and no vendor can sell it to you.

“But we're not Cox”

That's the expected response, and we hear it in almost every meeting. 15,000 employees, Accenture, AWS and a dedicated AI centre of excellence. An Israeli manufacturer with 80 employees reads it and concludes it's a story from another world.

The scale is different. The sequence isn't. What shrinks with scale is the number of systems, the size of the team, the budget and the number of processes. What doesn't shrink:

• Data before modelsThere's no shortcut. An organisation that starts from the model builds something impressive that falls apart at its third meeting with reality.
• Validation before automationIf you have no way of knowing the output is right, automation only raises the rate of mistakes.
• One process before everythingCox started with lead validation: one item, measurable, with a clear number before and after. Not with “AI transformation”.

At the size of a mid-sized Israeli company, the 75% isn't six data layers and a centre of excellence. Usually it's one table someone finally cleans up, and one check someone finally defines. It sounds small. It's what decides the outcome.

What we see on the ground

The clients we work with aren't technology companies. A manufacturer, an importer, a distributor. They have no AI centre of excellence and no Accenture. They have an operations manager who knows everything, and a system from 2011.

And there the 75% has a problem beyond the effort: it doesn't feel like progress. A manager told “first we'll sort out the data” hears six months with nothing to show, and he isn't wrong to resist. In an organisation that isn't a technology company, a project with no visible result loses its sponsor by the third month.

So our job isn't only to do the 75%, it's to break it into pieces that each produce a win. This week we clean up the customer table, and right away the salespeople get a list they can trust. Then we add the validation step, and right away the number of wasted calls drops. Exactly the same 75%, except now every stage pays for itself and buys the next one.

The second part is harder to describe, but it's the real bottleneck. In a non-technical organisation, the hard part isn't the model or even the data. It's that nobody ever wrote down how the process really works. The operations manager knows; it's in his head. In a mid-sized Israeli company, the Skill Studio is a person, and the work is getting what's in his head into a form a system can work with.

What to do on Monday morning

Four steps, in this order:

• 1. Pick one process with a numberNot “improve marketing”, but something you can measure before and after. Cox picked lead accuracy and started at 18%. What's your number, and do you even know it today?
• 2. Ask where the data for that process livesThe answer is usually embarrassing: three systems, two spreadsheets and one person who remembers. That's the work, not a delay before the work.
• 3. Build the validation step before the automationHow will you know the output is right? If the answer is “we'll look at it”, you're not ready to automate yet.
• 4. Measure roundsHow many experiments do you run on that process each month? If the answer is one, your problem isn't the model.

The bottom line

The Cox story reads like a story about artificial intelligence. It isn't. It's a story about an organisation that understood the model is the available, cheap part everyone shares, and put three quarters of the effort into things nobody gets excited about: data, validation and governance.

And if you budget 75% for building and 25% for data, you're budgeting exactly the opposite of the company that got 7x.

Related

• Process automationOrder entry, CRM updates, quotes and reports that someone copies by hand. We take one process that costs you hours every week and rebuild it inside the systems you already run.
• Analytics and reportingReports arrive from every market in a different format and at a different pace, and by the time the picture comes together the decision has passed. We connect the data from the ERP, the CRM and the markets into one view, with a layer that explains what changed and why.
• Organisational enablementMost AI projects don't fail because of the technology. They fail because nobody owns the process, the data isn't ready and nobody was trained to carry on. Organisational enablement is the work that makes it hold after we leave.