The fifth episode in a series of Silicon & Steel conversations with the founders building what’s next.

A big supplier deal can have twenty things on the table. Price, payment terms, warehouse space, transport, and more. Put that many pieces together and there can be a billion ways to shape the deal. Nobody can hold that in their head, so people pick two or three, usually price first, and let the rest land where they land.

My guest today built a company around the idea that software could do better.

Martin Rand is the co-founder of Pactum AI, which runs supplier negotiations for large enterprises. Its first customer was Walmart. Before Pactum he worked at Skype, then co-founded VitalFields, a farming software company that reached a million hectares across seven countries in Europe before Monsanto bought it. Today he is searching for his next company and backing small teams in India, Nepal and Bangladesh.

We recorded this one in person, in the heart of Silicon Valley.

The full conversation is on YouTube, linked below. What follows is the short version.

Let’s get into it.

Pactum started with a negotiation that went badly.

After Monsanto bought VitalFields, Rand stayed on as a manager, first at Monsanto and then at Bayer after Bayer acquired Monsanto. In one deal in France, the other side asked for custom pricing, rewritten terms of service, a white label product, and the direct relationship with the customer. He said no. They stood up and walked out, and he was left with nothing.

That moment sent him somewhere new. He started asking whether negotiation had a real method behind it, and whether computers might one day do it.

The answer he landed on comes down to scale. Once a deal carries many terms, the number of possible outcomes explodes. People cannot weigh them all. Machines can.

i) He sold Walmart a pitch, then built the product

Pactum met Walmart almost by accident. Walmart came to look at an Estonian startup, and that startup put out an open invitation to come and meet them. Rand says his team was nearly the only one that showed up. Walmart liked the pitch, and he flew to Bentonville, Arkansas to close the deal. There was no product. There was a strong presentation and a network of negotiation scientists already working with him.

That first deal took about two months. He is quick to add that this was unusually fast, and that later enterprise deals took six months to a year and a half.

ii) Do not show the product

This was the most surprising rule in the whole conversation.

He admits it sounds backwards, since the product is the thing the customer wants to see. But the product is also the only leverage you have, and you need the customer’s data to size the problem before you can price it. So you trade the demo for the process. His playbook, in plain steps:

That budget, he points out, is usually already promised to another project.

iii) Press opened the door

How does a small startup get noticed by a Fortune 500 company in the first place? His answer was PR. Social media and advertising are crowded, he says, and few companies do proper press work anymore. One story in a tier one outlet travels a long way inside a company the size of Walmart, which has two million employees. Somebody reads it, it reaches a manager, and people start reaching out. At Pactum, he says, almost half of the Fortune 500 interest came in inbound.

I asked whether a startup should wait until the product exists. He said no. Pactum did PR before the product existed. A strong idea and a clear concept are enough.

iv) Most moats are disappearing

He gave a simple two step test for any founder. First, can Claude do what your product does, out of the box, within six months?

Second, if you succeed, will there be five thousand competitors six months later who built the same thing quickly with AI?

The old list of moats, he says, is thinning out. Complexity, platform, ecosystem, and the strongest one of all, network effects. People stay on a platform like LinkedIn because switching is a hassle, and that hassle is what holds the network together. Agents have no such hassle. For agents to find each other, check each other and talk to each other costs only a few extra tokens.

By data he means original, proprietary, non synthetic data that a model does not already have, ideally from sensors or straight from the business. By regulatory he means a product customers are required to use.

v) Where a human still wins

I asked where he would tell a company to keep a person in charge. His answer was strategic deals between very large companies, because AI does not yet have enough context from inside those businesses. Two executives who fly somewhere to play golf, he says, may find a creative solution that two guardrailed machines would miss. He expects that to change.

Almost everywhere else he thinks machines have the advantage, because they can weigh every term at once. And when both sides run AI agents, he believes the deals get better, not just faster. He pointed to a concept from negotiation science called Pareto optimality, where neither side can gain more without the other side losing. Agents can keep hunting for that point without psychology getting in the way.

There is a second set of experience sitting behind this company. Arkestro’s CHis view on moats reaches well past startups. He thinks the big software platforms face real pressure too. He recently spoke with the CTO of a Fortune 500 company whose plan is to hire software architects to build the scaffolding and the context that AI needs, so the company can build more of its own software in house. In that world, he says, even Salesforce is not safe. An agentic system would need to reach into Salesforce once, take the data, store it however it likes, and carry on from there.

He was also open about what he got wrong. Like almost everyone, he did not see large language models coming. He pointed to an essay called The Bitter Lesson, which argues that decades went into building specialist systems when the effort could have gone into generalist ones. He had just read Ray Kurzweil’s The Singularity Is Nearer, which describes how neural networks were debated in the late seventies and eighties, judged unable to compete with specialist systems, and then set aside for decades.

His favorite book on AI is Superintelligence by Nick Bostrom, which works through why a mind smarter than ours would be so hard to contain. On supply chain he could not recall the title, but it was about the shipping container, and how one standard box changed global trade.

The biggest lesson from his journey was not a tactic. It was about discomfort.

Early on, very smart people told him machines could never negotiate, because negotiation is charisma and feeling and therefore human. One investor had made a lot of money from VitalFields and told Rand he would back whatever came next, whatever it was. Rand pitched him Pactum over dinner. At the end of the meal, the investor said:

“Martin, I really thought you’re gonna do something useful this time.”

He did not invest. Rand’s point is that this kind of feedback, even from people you respect, is not a reason to stop.

Rand says he was lucky enough to catch two waves early, digital tools in farming and agentic workflows before ChatGPT. Now he is looking for the third, and he is working through two ideas.

The first is about experts. Today an engineer, a designer or an inventor often works inside a large company and hands over the intellectual property in exchange for a salary. In an agentic world, he thinks that expert can build a company around themselves, with agents handling capital, legal, sales and finance, and keep most of the IP. Other experts join as co founders, paid in fractional ownership or future revenue. Innovation sits at the center, and these lighter companies start competing with the old ones.

The second is science. Software has become cheap, he says, and hardware is still slow, but both mostly optimize the world we already have. Science can grow it, through new energy sources and new materials. He is watching automated labs that run experiments under AI guidance and feed the results straight back into the models. A lab in a loop.

He is also giving small grants that take no ownership to teams in India, Nepal and Bangladesh, so they can buy a welding machine, a 3D printer, or a computer and an internet connection. He started after traveling.

His advice to anyone starting out now was short, and it came back to the same idea.

Build something that is hard to copy, and keep going when the smart people tell you it cannot work.

Subscribe if Silicon & Steel is your kind of thing. More conversations are on the way, and my thanks to Martin for being on this talk show.

Gaurav Singh Chaudhary
Silicon & Steel
Silicon & Steel, intelligence for chips, batteries & trade

Advertisement

Ads Ends