Why WeBuyCars Refused the ERP and Built Its Own

In February 2018, WeBuyCars was seventeen years old, moved around 2,000 vehicles a month, and ran on a Google Sheet.

One tab. They called it the recon sheet. Inventory, debtors, creditors, the lot. If Faan du Toit decided on a Tuesday that he wanted to track the colour of a vehicle, someone right clicked and added a column, and from that day forward the colour got captured. Everything before that day stayed blank.

So by 2018 the company had seventeen years of car pricing data, and almost none of it was worth anything. "On a Google Sheet, you don't really have anything, because it is unstructured," Wynand Beukes told me. "So that data was actually worthless. We couldn't do anything with the data."

He joined that February to fix it, leaving BCX to do it. Seven years later WeBuyCars is listed, runs 22 branches, employs just under 4,000 people, and describes itself, without hedging, as a data company.

The bus test

The real problem in 2018 was not the spreadsheet. It was that the spreadsheet was a symptom.

"The old saying that if a bus drove over Faan in 2018, WeBuyCars would have been in trouble," Wynand said. "Because most of the knowledge was in his head. And that's most of the time's problems with founder led businesses. If you can't transfer the tacit knowledge of the business."

That is the actual brief he took. Not digital transformation, whatever anyone meant by it in 2018. Take what one man knows about pricing a used car and turn it into a repeatable process that thousands of people can run.

Worth pausing on what he walked away from to do it. A senior role at a well known corporate, for a business that, as he puts it himself, bought and sold second hand cars. What made it work was not the salary. It was that the CEO handed him the problem and then got out of the way. "That was a huge aspect in a successful transformation journey, is your support from the board and the CEO."

The decision a CFO would have talked him out of

The first fork was obvious and the answer was not.

They needed off the spreadsheet. The safe route was an ERP. Proven product, known vendor, implementation partner, board comfortable, risk transferred. "If you talk to a CFO, the less riskier decision is to go with your traditional ERP system, quickly get that in, a well known settled product, and you think you de-risk the business."

They went the other way. They wrote their own.

Supported by

Silicon Overdrive

Silicon Overdrive Founders Series

An AWS Premier Partner running cloud infrastructure for South African businesses: architecting, migration, security and compliance, monitoring and cost optimisation. They also invoice in rand and can arrange proof-of-concept credits.

See their AWS services→

Understand what that meant in 2018. No Claude, no Copilot, no vibe coding. Senior developers were hard to find and harder to keep. And Wynand had to stand in front of a board and tell them this was not a turnkey purchase. "I had to convince the board at that stage, listen, this is not an overnight turnkey solution. It's going to take a year or two to get the foundations right."

A year or two, on a bet, at a car dealership, against a category of software that every competitor in the world already had off the shelf.

Own or control every touchpoint

The reasoning underneath it is the part worth stealing.

"One of our North Star guiding principles in 2018 was we wanted to own or control every touch point with a customer. And the way you do that, or the only real way you can do that, is to build your own software."

Same logic drove them fully into cloud in 2018, while the on prem versus cloud argument was still a live debate at most South African companies. Full managed services, cloud managed networking. "One of our goals was, we don't want to own anything."

Notice the pair. Own nothing physical. Control everything customer facing. Most companies get this exactly backwards: they own the racks and rent the customer experience from a vendor.

The duck

Here is where he breaks from type. Wynand's master's dissertation was on user experience, and it shows.

The first thing he did in 2018, while the business was still running on the Google Sheet, was hire a team of UX designers to redesign the website. Not fix the back end. The website.

"Like a duck," he said. "His feet are struggling or swimming underwater, and on top, they just have to portray this image to the outside that you've got an easy usable user experience."

His argument for it is the strongest thing in the hour. WeBuyCars spends millions building a warehouse, brands it, and maybe 100,000 people drive past it. The website takes 4 million unique visitors a month and 8 to 10 million sessions. "Your website is your digital billboard. If that thing doesn't work, and it's not effective, it's not efficient, and it's complex, then you lose a lot of customers."

He is blunt about the industry failing here. "I truly believe that companies, South Africa and worldwide, are not spending enough on the user experience of your digital channels, and spending too much time on the back end, building these Ferrari engines." An ugly body on a perfect engine, as he puts it, and nobody wants the car.

There is a second half to it that most companies skip. Developers at WeBuyCars use the software as customers. They are given money to go and buy a vehicle, and they sell it themselves. "There's no better satisfaction for the developers if they really see their software in action. And then they come back with suggestions."

And when I asked whether the hard part had been the technology, he said no, immediately. It was the people. "People are scared of digital transformation, that they're going to lose their jobs."

The first hire was a data scientist

He was employee number one in IT. The first person he hired into the business line was not a developer. It was a data scientist.

In 2018. When the discipline barely existed here as a job title.

"We knew that we had to create clean data pipelines in 2018. And that led up to where we are today."

That is the whole story compressed into one decision. The pipelines got built alongside the platform rather than bolted on afterwards, which is why WeBuyCars can now put AI on top of clean data while most companies are still trying to work out what they actually have. "The most difficult thing that businesses sometimes get wrong is that you hear a lot of comments that we have a lot of data, but the challenge is how do you operationalise your data and how do you make decisions with that data?"

The dream, and the 5 percent

In 2018, he and that first data scientist, who is still there, set themselves a target neither of them could have delivered at the time.

Buy and sell a vehicle with no human involved.

"That was a dream in 2018." Seven years of what he calls building the puzzle, or playing Tetris, block by block, towards pricing models good enough on both sides of the trade.

It is running. Around 5 percent of what WeBuyCars buys and sells is now fully automated, no human anywhere in the loop, and they have put more than 6,000 vehicles through it in roughly a year.

The next block is computer vision, which sounds easier than it is. Telling a facelift from a non facelift model, confirming a variant from photographs. Solve that, he says, and a large part of the pricing decision automates behind it.

Meanwhile, a detail that says more about scale than any of the headline numbers. The pricing models got so large that South Africa could not run them. They were trained on A100 GPUs, and the only compute they could find was in West Germany.

Why explainability came before accuracy

The interesting constraint was never the maths. It was Faan.

"One of the great things about Faan was he had an inherent feel for a car. And probably if we could have cloned Faan, WeBuyCars would have been easy."

You cannot hand a man with twenty years of instinct a number from a black box and expect him to trade on it. "You can't have just this AI model and it's this black box and people don't trust it and don't understand it. So you have to get this middle ground in this change management phase."

Explainability was not a compliance exercise. It was the price of adoption.

And what the model is optimising for is not what most people assume. WeBuyCars is deliberately high volume and low margin. A quarter of the stock sells on day one, on auction, the morning after it is bought. "If you can sell a quarter of your stock for cash the next day, and you get money and you can buy more, that's how the business scales." Push the margin too hard and you slow the engine.

Because they control both sides of the trade and sell no new vehicles, they are tied to no OEM. Short of parking bays, throttle the buying. Short of buyers, thin the margin and move stock faster. "We can dial the knobs or pull the levers to whatever side we want to."

The 98 percent rule

On agentic AI, he is further along than most and more disciplined about it than almost anyone.

Agents are already live in the flow. If the photographs you send are missing the front lights, an agent asks you for them before a human ever touches the valuation.

But agentic negotiation, an agent haggling on WeBuyCars' behalf, is not switched on. Not because it is hard. "To build an agentic AI layer on top of that is easy. That's not the problem. That's just developers doing that."

It is because the number underneath has to be right first. "That pricing model should be a 98 percent accuracy." Get from 80 to 81 and the next percent costs more than the last. That is where the last two years have gone.

And the reason is data provenance. Ask a general model what a diagnostic code will cost you to fix and you are trusting training data you do not own or control. "We want to use proprietary data to feed those LLMs. The technology or the principle of LLM is there. We can build it, but the data accuracy is the challenge. Otherwise your hallucinations become too high."

Internally they have already built what he calls a small language model per lead manager, one per person, learning which cars that individual actually likes to buy. German or Japanese or Chinese. New or old. Whether they get nervous above 300,000 kilometres. On the dealer side, recommender engines have been running since 2019, so the first 24 vehicles a dealer sees when they log in are the ones they are most likely to bid on.

When the customer sends an agent instead

I ran a live test during the interview. I asked Claude to research cars for me, and it came back with nothing useful from WeBuyCars, because the site is a JavaScript app and returns an empty shell to a fetch.

His answer was not defensive. They have already shipped a WeBuyCars plug-in for ChatGPT. Claude is next.

"I think the principle of websites is probably going to be removed in a couple of years time. My kids don't even know Google. They only use things like Claude and OpenAI."

Which is quite a statement from the man who built his strategy on the website being the digital billboard. Both are true at once, and that is the point: the billboard is moving, and he is moving with it, aggressively.

Buying the farm next door

The Chinese vehicle wave is compressing margins right now. Buyers who used to spend R300,000 to R700,000 on a used premium car are buying a new Chinese one instead.

He thinks everyone is reading it wrong. "We don't see it as a structural shift. It's cyclical." Those cars take three to five years to reach the second hand market, and then they become stock. WeBuyCars has already bought and sold as many Chinese vehicles in the first six months of this financial year as in the whole of the last. The ownership period on them is shorter than on Japanese or German cars, so they turn faster.

So in a deflationary year, while dealerships pulled back, WeBuyCars opened six branches in nine months and added more than 3,500 parking bays, over 30 percent more capacity.

"It's like when you're farming and it's in a drought period, and you buy the farm next to you, and next year you get some rain again, you have the land."

Take the double whammy of branch cost and margin squeeze now. Own the capacity when the wave arrives.

Clarity is kindness

The last stretch was about people, and it was the most direct part of the conversation.

On hiring: they prefer people from outside the motor industry, because they arrive without preconceptions. "You don't need to be a petrol head to join WeBuyCars." Attitude over skills, because skills are teachable and everything is on the internet. Drivers at WeBuyCars have become branch managers and sales managers.

On entitlement, he does not soften it. "It's not my responsibility to grow you. We're not looking for entitled people that come and sit here and think that we must do everything for them."

On communication, he admits it has been a weakness and that fixing it is this year's work. Twenty two branches, and the leadership team travels to all of them. "There's a saying that says clarity is kindness."

On fairness: "One of the things that people really don't like is unfairness." Be fair, be transparent, be consistent, and trust follows.

And on the transition from founder led to listed, which is where most companies quietly become something worse: about five shareholders hold roughly half the business and think in long horizons, which buys him room. The leadership team that started in 2018 is still intact, and he says his job this year was keeping it together for the next five.

"If you walk in here, you won't even know it's a listed environment. We handle it like a startup and we will keep on doing that."

Which brings it back to 2018 and the Google Sheet. Every decision since has been a version of the same one: keep control of the thing that matters, and be willing to look reckless in the short term to get it.