We’re building a new Dell to put the old Dell out of business. The model is the cheap part.

Several years ago Michael Dell set the company a target that reads, on first pass, like a provocation.  

“I stood up and told the company that 5 years from now we will have a new competitor and that new competitor is going to be in every business that we are in and they’re going to be faster, more efficient, and more capable. And they’re going to put us out of business, and the only way that we’re going to prevent that is by becoming that company. It’s gut-wrenching stuff to reinvent and reimagine your business, but if you don’t do it, you go out of business,” said Michael Dell 

We’re building a new Dell that would put the old Dell out of business. Not a competitor’s Dell. Ours. 

The logic holds once you accept how competition behaves when it’s powered by artificial intelligence (AI) across every part of a business. Somebody in your industry and your vertical is going to move faster than you and take the market, and what I’ve watched over the last three years is that when it happens, it doesn’t happen gently. It’s better that the company doing it to you is a version of you. 

Here’s the part that took longer to understand, and it’s the part most marketing organizations are still getting wrong. Whatever lets the new version beat the old one was never going to be the model. Every company buys from the same handful of providers. Your competitor can have an identical one running by Friday afternoon. Whatever separates the two Dells must be something nobody can subscribe to, and there’s really only one candidate on the table. Your own data, and the voice you build out of it. 

A subscription is not a strategy. 

When a board asks whether the company is AI enabled, the reflex is to buy access to a frontier model and check the box. The honest answer in a lot of those rooms is that you’re not AI enabled at all. You have a subscription. 

Think about what actually got bought. A model trained on the same corpus as your competitor’s, carrying no opinions, no history, and no point of view. Train a model on company data and it can take on a personality: a vocabulary, a set of positions, the way people there actually talk. A frontier model can’t do that, because nobody who subscribes gets to train it. It arrived finished, and it arrived finished for everyone else too. 

Which makes the interesting question something other than which model to buy. In our case it was ‘Who is Dell? Who are we?’ Most organizations have never answered that anywhere a machine could reach it. 

We had to point this somewhere first, and the choice mattered. There were thousands of AI projects running across the company at the time. Everybody was experimenting, and almost nobody had real key performance indicators attached to any of it. We asked which capability genuinely differentiates Dell in the marketplace, and the answer came back as our go to market strategy and customer relationships: it was about sales and marketing working better together. Fastest time to value, and frankly where the money was already going. For a product company, most of the spend sits there. 

The moat turned out to be in our own file system. 

The Dell AI Factory has been our primary message to the market for four years. Every region, every product line, every campaign runs through it, and if you’d asked me to state it, I’d have done it without pausing. 

Then we went looking for where that message actually lived. It was in 14,000 SharePoints, holding more versions of itself than anyone could count. Not 14,000 documents. 14,000 separate places, most carrying their own edit, some carrying several. 

Nobody had done anything wrong. That’s what happens to a message inside a large company over four years. But you cannot hand that to a model and expect a distinctive sentence to come out the other end. An answer that exists only in a brand deck isn’t an answer a system can use. It has to sit in the actual data corpus. 

Collapsing those 14,000 SharePoints into what we now call the Dell AI Data Platform is what fixed it. Only current copy goes in. Heritage products that don’t exist anymore stay in archive, because nothing good comes from a model reaching for them. The storage savings were real and beside the point. What mattered was what became possible afterward. When the message changes now, the old version comes out and the new one goes in, and from that moment everyone in the company is working from the same story. Messaging from five years ago stops leaking quietly back into the market, because it’s no longer there to find. 

And what does that get you, once every team finally says the same thing? Consistency. Which is not remotely the same as being worth reading. 

Data gets you consistency. Voice gets you distinctiveness. 

Voice is a separate build, and it sits on top of the data rather than inside it. 

On top of the Dell AI Data Platform sit the knowledge assistants that make up the Dell AI Factory, one for each role, each with an engine behind it. For marketing, that engine does two jobs that look unrelated but aren’t. 

The first is generating net new copy. Raw material arrives from the product team and comes back as something that can go out to the market, to channel partners, to a segment, to customers. The experience is deliberately familiar. It feels like any AI tool a marketer already uses on a weeknight, except everything underneath it belongs to the company. 

The second job is guardrailing what comes out, and that’s the part that protects the voice. Where does a marketing team’s week actually go? A surprising amount of it goes into checking content for trademark, copyright, and patent exposure, staying inside brand vernacular, staying inside the palette. All of that used to get handed to an agency by default. Now it runs against rules that only had to be built once. The multimodal side has improved enough in the last 18 months to generate a lot of our graphics inside those same rules, faster than an outsourcer could turn them around. The guardrails keep getting sharpened. The education never has to happen twice. 

A voice that isn’t maintained decays. 

Which is the trap in treating any of this as a project with a finish line. 

Marketers keep asking how they’ll know when the pilot is over. Michael’s answer, and ours, is that there’s no better time to start than right now, and anyone not yet using this for production work is already behind their competition by years at this point. Across every industry and vertical, the separation between winners and losers over the last three years tracks almost entirely to who invested and who waited. 

Urgency gets you a launch, though, not a capability. Prototype fast, reach inference scale, put it in front of every user, run real concurrency, and then keep going. We used to call that DevOps with traditional SDLC. With a model it’s called fine tuning, and it’s the same discipline wearing different clothes. 

Our marketing assistant gets retrained every week, because every week another pile of work lands in the Dell AI Data Platform. Parameter efficient fine-tuning runs over the weekend, and Monday’s version is smarter than Friday’s. The teams I watch struggle are the ones skipping that cycle. Eggs get broken either way. The only difference is whether anything gets learned from them. 

Agents belong inside the same loop. On our platform they’re enabled for marketing with no hold back, on shared infrastructure that stays governed and secure, with a model hub so anyone can choose which open weight model to run against. That choice matters more than it sounds, because voice shifts depending on which model sits underneath it. One of the most useful things an agent does here is test exactly that. Run this copy through these four approved models, compare them, and show me which one is better and why. 

Fully Autonomous agents are still a long way off. This remains human in the loop territory, and will for a while yet, at least for enterprises. 

What a voice that works can actually do 

Here is the old Dell and the new Dell, in one campaign. 

A couple of years ago, watching what Meta was doing with hyper personalized marketing, we brought in an independent software vendor that had started life as a marketing company and rebuilt itself as an AI company. They were a longstanding marketing outsourcer for Dell, and had built their own engine for hyper personalization. 

The target was a set of accounts that had never bought Dell. The old motion was hard copy and old school email campaigns, running at roughly tens of thousands of dollars per customer with a return on investment that never showed up. We got some results. Nowhere near what we wanted. 

What he built instead was a campaign anchored on a personalized message from Dell leadership. Picture a highly visible Dell executive directly in conversation with you as the CEO of that company, addressing you directly, talking through the AI Factory and where it lands in your industry, drawing on external research about what that company and that vertical are dealing with. One voice, defined once, holding thousands of individual conversations. 

Conversion came back 50 times stronger, at thousands of dollars instead of tens of thousands. No competitor could have bought their way to that, because the asset doing the work wasn’t the model. It was Michael’s voice, and the data underneath it. 

Which changes what an agency is for 

Follow that one step further and the agency relationship looks different than it did. 

If voice is the asset a company owns, maintains, and can prove, then buying voice from outside makes less sense every quarter and getting access to it from inside makes more. The work ahead is figuring out how to safely and securely open Model Context Protocol (MCP) connections into those data sources, so agencies can create on our behalf with the brand already loaded. We’ve trained the model on who we are. What’s missing at that point is only copy fed up into it. 

None of which means agencies get replaced. Everybody braces for job loss, and the likelier outcome is a change in what the job is. The agencies we work with now are AI powered themselves, increasingly running agents that talk to our agents, humans reviewing on both sides, work leaving the funnel faster at lower cost. The hours it takes to produce an asset drop sharply. Does the rate per hour have to drop with them? Not necessarily, because the number of customers a single agency can serve goes up a great deal. 

Forty years in IT means more calls than I can count with a marketing team and a room of SMEs, all of us on the phone working out what we wanted. Very little of that is necessary now. The lift is that much lighter, which frees the relationship for something more valuable than production. 

The old Dell paid agencies to manufacture its voice one asset at a time. The new Dell owns the voice and hands out the keys. Agencies stop being where assets get made and become where more customers get reached, and the ones that see it first won’t be competing for the same work. They’ll be doing different work.