Skip to article

Start With the Problem, Then Clean One Slice of Data

Take ten steps back before picking a tool, and skip the cleanup of all customer data. What Cardinal Health's Alyssa Cox described, and where BNEDai would start.

By Jacqueline V. TwillieNovember 20, 20264 min readGuide
Quick answer

Characterize the problem first, then choose the tool. Alyssa Cox of Cardinal Health described that sequence at an online enterprise AI summit on November 12, 2026, and said her team started with a plain-language chat on the data. BNEDai's addition: start with the one slice of data that answers a question people ask all the time.

A consultant once talked to Alyssa Cox, VP of Transformation at Cardinal Health, about putting caviar on Taco Bell. She told the story at an online enterprise AI summit on November 12, 2026, and said AI may be over-engineering. You have to match the solution to the need. Her remarks on two points follow.

Characterize the problem before picking a tool

Cox said that when someone comes to her team with a thing AI can solve, "we stop, and we take 10 steps back." First they characterize the problem and what the person is looking for, and then they decide on the tool. You cannot start with the tool in mind or you will over-engineer it, she said.

On Cardinal's pricing research tool, they came in with a plan for what AI would do, then found that asking it to think through every permutation of the decisioning was inefficient. So a person makes the decision at the key points. Maybe in the future AI will make that call, Cox said, but not today.

BNEDai's note: write the problem in one plain sentence before looking at any software, with what a good result looks like beside it. A person makes the decision at the key points, as in Cox's pricing tool, and the AI agent prepares.

Skip cleaning all of the customer data

There is no business in cleaning all of your customer data, Cox said. Ask what counts as customer data and every person has a niche of it they want included. You will not get through it all, she said.

Build the simpler ask first

Cox's team started with a plain-language chat on the data, for questions like whether a customer is covered by their group purchasing contracts. She said a simpler ask is often the foundation for a harder one.

BNEDai's note: where we would start is one question your people ask all the time, and only the data that answers it.

How BNEDai sets up a slice of data

BNEDai's own build practice has a version of the same step. A knowledge source is a folder of documents an AI agent pulls from first. Jacqueline V. Twillie described it in a published video using her speaking AI agent's folder, which holds her one sheet, topics and bio, kept in the place she already keeps her documents. In the video she did not connect her own folder, because she wanted to check what was in it first, and the page on what to put in an AI agent's knowledge source lays out the order: prepare the folder, read it through, then connect it.

The read-through is where a team decides what the AI agent can see.

A first test

BNEDai's published test is to run the AI agent once, look at the first result, and give feedback with reasons. Jacqueline V. Twillie said not to say try again, and to say what the AI agent surfaced and why it did not help. See how to test an AI agent and give feedback. The question and the slice of data are what you test against.

Frequently asked questions

What did Cox say about cleaning all customer data?

That there is no business in it. Every person has a niche of data they want included, and you will not get through it all.

What did Cardinal Health start with?

A plain-language chat on the data, for questions like whether a customer is covered by their group purchasing contracts.

Where does a person stay in the decision?

Cox said that on Cardinal's pricing research tool a person makes the decision at the key points.

Source: remarks by Alyssa Cox, VP of Transformation, Cardinal Health, at an online enterprise AI summit, November 12, 2026. The notes marked BNEDai's, and the knowledge source and test practices, are BNEDai's.

Reviewed on November 20, 2026.

More on Data for Non-Technical Managers