The Hard Part Was Never the AI

A price change at a national retailer touches more systems than most people would guess. The pricing engine holds the rule. The promotions engine holds the exception to the rule. The loyalty platform holds the exception to the exception, because a customer’s tier changes what they actually pay at the register. The point of sale has to resolve all three in the time it takes a cashier to scan a candy bar, across terminal hardware spanning fifteen different experiences. Then the whole thing reconciles to the penny that night, because the money is real and somebody’s franchise agreement depends on the number being right.

I have spent more than twenty years inside systems like that. I built a retail analytics company and sold it to Verifone, which meant years spent living inside the payment terminal itself. Since then my team has worked on point of sale, loyalty, pricing, and reconciliation for some of the largest retailers in the world. More than twelve billion customer interactions have run through systems we have touched.

So when people tell me AI is about to transform retail, I believe them. I just don’t think it happens the way most of the conversation assumes.

The demo works. The estate doesn’t.

Point a coding agent at a greenfield project and it is genuinely remarkable. Point the same agent at a thirty-year-old retail estate and something different happens.

It writes confident code. The code compiles. The tests pass. And then it breaks something in reconciliation three weeks later, because a rule in the pricing engine that looks arbitrary is not arbitrary at all. It exists because of a tax situation in one state, or a vendor contract from 2014, or a workaround for a terminal model that is still running in four hundred stores.

The agent had no way to know any of that. Neither did the last three contractors. The code is the only documentation, and code tells you what the system does. It almost never tells you why.

That is the actual bottleneck, and it is not a model capability problem. An agent amplifies whoever already understands the system. When nobody understands it, you do not get better answers. You get wrong answers faster, and at a scale that used to take a team of ten a full quarter to produce.

What has not changed in thirty years

Retail systems are never replaced. They are layered.

Every retailer of consequence is sitting on decades of point of sale, ERP, loyalty, pricing, and store infrastructure that works well enough to be irreplaceable and badly enough to be a ceiling. The systems that made the company successful are now the reason it cannot move.

I have watched a lot of executives try to solve that with a replatform. Most of those projects are announced with a date, and most of those dates move twice before somebody quietly redefines the scope. It is not incompetence. It is that the old system is doing a thousand things nobody catalogued, and you find out what they were one outage at a time.

The AI era has not changed this. If anything it has raised the stakes, because the organizations that can safely change their systems are about to pull away from the ones that cannot, and the gap will compound quarterly rather than annually.

Three things we believe, and none of them are common

We meet the as-is, not the to-be. Most firms overthink the target state and under-appreciate the one that is actually running. It is far easier to write code than to rewrite it, and that is more true with agents, not less. Harmonizing with what exists is slower to start and dramatically faster to finish.

We optimize for obviousness, not elegance. Code outlives the people who wrote it. The thing that matters across generations of engineers is whether the next person can tell what is wrong, where to look, and how to add the next thing. Elegance is a compliment an architect pays himself. Obviousness is a gift to whoever inherits the system, and in an era where a meaningful share of that inheritance goes to agents rather than people, obviousness is worth more than it has ever been.

We operate what we deliver. Which puts our developers on the hook for whether their code actually works at two in the morning on the busiest weekend of the year. It changes what people build. It changes it immediately.

Where I think this goes

The retailers who win the next five years will not be the ones with the best AI strategy deck. They will be the ones whose systems are legible enough for AI to work inside safely.

That is a different investment than most of the industry is currently making. It is less exciting than a replatform and considerably cheaper. It means treating thirty years of accumulated decisions as an asset to be written down rather than a liability to be escaped. It means the institutional knowledge sitting in three people’s heads, one of whom is eighteen months from retirement, gets captured while those people are still in the building.

Do that, and agents become genuinely useful on the systems that actually run your business, not just on the new services around the edges. Skip it, and you will spend the next five years generating confident code against a system nobody understands, which is a worse position than moving slowly.

I am not neutral here. This is what we do, and we built a platform of our own because our engineers kept solving the same problem and we wanted the organizations we work in to keep the knowledge after we leave. But the argument holds whether you hire us or not. The retailers I would bet on are the ones treating their legacy estate as something to be understood rather than something to be survived.

The unglamorous version

Everything above can be said more plainly, and one of my colleagues put it better than I have:

People can’t do the hard stuff fast.

That is the whole competitive argument in seven words. The hard stuff in retail is not novel. It is the register, the reconciliation, the loyalty rules, the pricing exceptions, the thousand small integrations that nobody wants to own. It has never been glamorous and it has never been optional, and it is where the AI era will actually be decided, because that is where the money moves.

Got a hard problem in your retail business? Need to go fast?

We should talk.

Wells Burke is the co-founder and CEO of Rocket Partners, an embedded engineering firm working on retail’s hardest systems, and of CodeVine.