TBW - Rand Hindi (Zama): "The agentic boom is the product manager's revenge on the developer"

The Big Whale: You run a tech unicorn. What has AI concretely changed in the way you manage Zama?

Rand Hindi: The most visible decision is how we manage headcount. We've stopped hiring until every team has integrated AI into their day-to-day work. When someone tells me they're short on engineers, my answer now is: show me first what you've automated. Take your workflow, try to replicate it with AI, and let's see what's left. There cannot be a single team - internally or externally - that skips that exercise. Once it's done, we can talk about hiring.

Which roles do you still hire for, no matter what?

Two. Sales and cybersecurity.

Sales, because closing enterprise deals is fundamentally about human contact and relationships. There's a relational dimension to the commercial process that will never be automated. AI can manage a SaaS marketing funnel. It will never close your enterprise deals, it will never build your strategic partnerships.

Cybersecurity, because we are entering an arms race. We're back in the early 1990s of computer viruses, except the attackers now have AI - and so do you. Defending against AI-powered attackers requires a security team that itself knows how to use AI. A good security engineer is hard to find, and the workload is brutal: monitoring 50 services in parallel, all the signals, everything happening across the network. AI allows you to multiply that capacity. With the right architecture, a startup can today have the security coverage of a major bank. And that holds across every sector.

Where do you draw the line between what AI produces and what a human signs off on?

We're still figuring it out. The rule I've set for now: not a single line of code goes to production without going through a human. The deploy button is pressed by a person. That's not necessarily a line we'll hold forever, but it's where we stand - and it makes sense.

Brian Armstrong announced Coinbase is cutting 14% of its workforce and that non-technical staff should be able to push code to production. Do you share that view?

Non-technical people writing code - fine. Pushing it to production after a human has reviewed it - fine too. But you need guardrails. Models are evolving so fast that what you got yesterday isn't what you'll get tomorrow. Determinism in AI outputs is a genuine problem right now. You need a process with filters and checks to make sure the model hasn't introduced a bug, hasn't pasted a credit card number, hasn't done something careless. Until we have a stable framework for putting AI-generated code into production, keeping a human in the loop is non-negotiable.

"The deploy button is pressed by a person. That's not necessarily a line we'll hold forever, but it's where we stand"

Several tech CEOs argue that the AI and agentic wave will hit junior profiles first. Do you agree?

I don't enjoy saying it, but yes. The cost-benefit of training someone versus hiring an experienced profile directly - even a much more expensive one - who already knows how to use AI tools: for me, there's no contest. A senior with ten years of experience who knows how to run a project, leading an army of agents and LLMs, does what an army of juniors used to do. That's true in marketing, in engineering, in product, in business development.

Three years ago I told my teams to stop taking interns. They come for six months, you train them, they leave. No interest.

So what happens to juniors in the labour market?

They become founders. That's my thesis. An entire cohort of graduates won't find jobs, and the consequence is a wave of startups built by 20-year-olds who grew up with these tools and can build companies faster than my generation (Rand was born in the 1980s, Ed.) ever could at that age. Over the past decade, most startups were founded by people in their thirties and forties - the market had grown more saturated, more SaaS-heavy, you needed accumulated scar tissue. That's changing.

My prediction: there's first a trough where juniors struggle to find work. Seniors get paid extraordinarily well because they know how to run an army of agents. And juniors become entrepreneurs. You end up with a new category of startups - faster, cheaper, more efficient than what we have today. Most one-person unicorns will be built by young people, not by experienced ones.

Is this a reshaping that goes beyond the tech sector?

Completely. Think of millions of micro-companies - the equivalent of your parents' hair salon, but in tech. Your small outfit does a few hundred thousand euros a year, you work with a partner or your family, one or two employees. That's going to exist across a huge number of sectors. You'll launch an e-commerce business solo. You'll launch a mobile app solo. You probably won't build a competitor to Anthropic on your own, but a large chunk of what used to require a hundred-person startup, you can now do with a handful of people.

What profile are you looking for at Zama in this context?

It's changed a lot. We used to send candidates a coding test. Today, the process focuses on your ability to run a project and deliver quality work, rather than your ability to crank out code.

Product management, project management - those are suddenly highly valued skills. The agentic boom is the product manager's revenge on the developer, in terms of value creation. The best profiles we've seen are engineers who became managers. That dual skill - manager and engineer - is exactly what you need to orchestrate agents.

"The best profiles are engineers who became managers. That dual skill is exactly what you need to orchestrate agents"

Has AI at least narrowed the gap between an average engineer and an exceptional one?

The opposite. The gap is widening. Engineers who were stars before are even more so with AI. Someone average with AI stays average. Someone exceptional with AI does the work of fifteen people. And the exceptional ones are going to be paid far more than before, because their productivity is frankly absurd.

To use these tools effectively, you still need an engineering mindset, even if you're not writing code day to day. You need to be a good manager, specify exactly what you want, frame problems with precision. Management skills, project management, a sense of systems architecture - all of that makes an enormous difference in what comes out the other end.

Revenues at the major AI labs are growing parabolically. Is the agentic economy actually sustainable, or are we going to end up paying more for agents than for humans within three years?

That's slightly the case right now, because there hasn't yet been a genuine, measured ROI on this wave of investment. The real answer will come in about a year, once companies have truly integrated these processes and can measure the impact on their revenues.

At its core, it's a mathematical equation. In business, you put X euros into the machine and Y euros come out. Whether you spend that on headcount, infrastructure, software vendors, or tokens at Anthropic or OpenAI - it doesn't matter. It's measurable. Companies are not going to spend tens of billions if they don't see a corresponding increase in revenue, or at minimum an improvement in margins. Either it cuts costs, or it generates business. It can't do nothing.

I'm convinced that within a year we'll have the answer. And I'm convinced it will be a positive one.

That's also the bet you're making on your own technology. Walk us through FHE for someone who's never heard of it.

Fully Homomorphic Encryption allows you to compute on encrypted data without ever decrypting it. Think of it as end-to-end encryption - the kind you have on Signal or Telegram - but extended to any service you use online. Today, when you query Claude, Anthropic sees your prompt in order to run the model, then returns a response. The data is encrypted in transit, but not during the computation. With FHE, the data stays encrypted even during the computation. The provider never sees your prompt, and never sees the response.

"Either it cuts costs, or it generates business. It can't do nothing"

The ideal privacy layer for enterprise AI, it would seem. Yet you say that's not actually where your business is.

That's correct. The enterprise AI privacy use case sounds compelling on paper, but in practice nobody really cares. People send their prompts to Anthropic without a second thought. There's an established level of trust in serious vendors. Nobody refuses to use Claude on privacy grounds. It's the same with Google and email.

There's also a contractual reality. When Zama signs an enterprise contract with Anthropic for our staff to use Claude, Anthropic is contractually prohibited from using our data to train their models. That's already settled. The only residual risk is a breach — someone exfiltrating data from Anthropic's servers. FHE solves that. But that's a cybersecurity problem, not a privacy problem. And it's not what keeps executive teams up at night.

So where is the real use case?

Blockchain. On Ethereum or Solana, every transaction is broadcast publicly. You can see someone's balance, what they bought, what they sold, their portfolio positions. You have no privacy at all. With FHE, we can add a privacy layer on top of public blockchains without moving assets to another ledger. Your money stays on Ethereum, you can still transfer it, still trade it, you keep all the advantages of a public chain - but observers can no longer see what you're doing.

In the context of agentic finance, where agents execute transactions autonomously, what does that change?

It changes everything for strategy protection. In on-chain financial transactions, whether the actor is a human or an agent, if you need to protect your strategy, your assets, your positions, you need them encrypted. Whether an agent is making a payment or a human is managing a portfolio, the privacy requirement is the same.

There's a separate question around upstream agentic confidentiality - vis-à-vis the model provider running the agent - and that's technically feasible with FHE, but as I said, nobody really cares about it today. The interesting layer is what the agent does once it leaves Anthropic's environment: payments, trades, communicating with another agent. There you're back in the real world, and that's where Zama and the blockchain stack have a role to play.

Zama is a European company. What share of your business is European?

Almost none. Around 95% of our blockchain clients are in the United States, Asia, and Switzerland. Switzerland is technically in Europe, but not in the EU.

It's not that there's no interest in Europe. It's that Europe as a political entity doesn't promote blockchain adoption. Look at Christine Lagarde saying stablecoins are useless for promoting the euro globally. When the institutional ecosystem doesn't want blockchain to win, that's mechanically where the business isn't.

In AI, there may be a slightly stronger strategic case for Europe, but honestly: I've never seen a company use an inferior product because it's European. The AI race is too important.

If you want to be competitive, you have to use the best tools, and right now those are mostly American. They've invested hundreds of billions of dollars. Europe has invested a few hundred million. The Americans have decades of accumulated expertise in scaling software and technology. Seventy years of entrepreneurial and venture culture are playing out right now in AI. This isn't a privacy question.