How Alpura CTDO Nahmias Unlocks Compounding Payoffs With Foundations-First AI Architecture

For enterprises watching their own proofs of concept stall, the sequence itself is often the problem. The standard AI rollout starts with the agents and works backward to the data they touch. Pick the use case, point the technology at whatever data it can reach, and sort out governance later. A different sequence is proving more durable: govern the data first, centralize the definitions the whole business runs on, retire the platforms fighting each other, and only then bring in the agents. Done in that order, the agents stop hallucinating and start compounding.

Marvin Nahmias is Chief Technology & Data Officer at Alpura, a roughly $2 billion dairy producer with more than 12,000 employees and one of the largest in Latin America. He came to the role after building centers of excellence at Coca-Cola FEMSA, running technology and transformation at Grupo Lala, and co-founding a cloud-native digital bank. His team put the governance-first sequence to work at Alpura, building the data foundation in 120 days before deploying agents at scale, and the experience cemented a core conviction for him.

"My belief is AI does not replace people. It just amplifies them," Nahmias says. Getting to that point meant resisting the pressure to start with the agents at all.

Governance before agents, not after

Nahmias arrived at Alpura to a CEO fielding constant calls to "do AI" and an industry full of consultants reinventing themselves for the buzzword of the moment. "You've got all these consultants that used to sell encyclopedias, then they sold used cars, then they sold SaaS, and now they're selling AI," he shares. He pushed the other way. Rather than chase agents, his team built a governed lakehouse foundation on Databricks, consolidating the platforms and standardizing the business logic underneath in about four months.

The centerpiece was the semantic layer, which he treats as the source of truth for how the business defines things. This is key for solving a commonplace problem. When finance and sales calculate the same metric two different ways, for example, feeding both into an agent produces exactly the mess everyone blames on the model. "If you just feed that to agents to do analysis, you're getting hallucinations because you're not going to the source of truth," Nahmias explains. In his telling, the failure isn't the model inventing things, but a person handing the model conflicting inputs and calling the result a hallucination. The semantic layer resolves this.

Two kinds of AI, one set of guardrails

Alpura splits its AI into two categories. One is tactical productivity tooling for the roughly 90% of employees who aren't power users, built mostly in Google's stack so people can compare job descriptions to candidates or draft analysis without touching anything sensitive. The other is governed agents that run against company data through the platform. More than 550 agents are now in production, and a majority were built by business users rather than engineers. The split lets Nahmias put real controls on cost, a discipline many teams learn the hard way. "I just heard Uber spent their allowance for AI for one year last trimester," he says. He set up token pools and negotiated allowances so spend stays inside the guardrails instead of arriving as a surprise invoice.

Every agent also gets registered, whether it was built by code through GitHub or spun up by a business user through a prompt, then audited in a central catalog. That registration requirement is where Nahmias watches consulting firms lose their footing, because the governance layer they used to sell is already in place.

Governed data compounds

Once the foundation held, the returns showed up across the business. Alpura layered machine learning models to address what he describes as one of the industry's most challenging aspects. "The hardest part for a CPG company is demand planning. Cows give milk every day. Should I convert it to powder? Yogurt? It's expensive to get wrong." With his approach in place, Nahmias reports a 15 to 16% improvement in planning accuracy. On top of a revenue growth management platform, agents now run predictive and prescriptive what-if scenarios his team couldn't see before.

The same governed data reshaped how Alpura works with retail partners. When Walmart's data used to take weeks to arrive, Nahmias connected it in two weeks and found his team seeing more than the retailer's own reports showed, enough to sit with merchandisers and flag where hard discounters were moving in.

Internally, the productivity change reads in how people spend their time. He describes senior staff who no longer disappear into building slide decks and instead argue about the actual business question. "Instead of PowerPoint, now they use Claude or they use Moda, and their productivity has gone sky-high," Nahmias says.

Change management ran in reverse, too

The technology was the cheaper half. Nahmias estimates change management cost twice what the platform did, and he ran it against the usual playbook. Instead of telling skeptics they were in the valley and needed convincing, he handed the wins to the people who earned them.

His team proved the approach with a small group in sales working on a protein product, got results in weeks that had eluded the team for months, then put those employees in front of the crowd to explain what they'd done. "You're going in the town hall meeting to explain what you did," he recalls. "You say it, we'll back you." Recognition, not mandate, did the spreading.

He's also redrawn where technology sits relative to the business. Reporting to the CEO, his function acts as an enabler rather than the owner of outcomes or the department that gets blamed. The mindset he insists on cuts both ways: technology can't wave off a problem as a business issue, and the business can't wave off AI as something for IT to figure out.

That posture is where his second non-negotiable lives. Alpura is closer to B2B than a financial org, but Nahmias keeps strict policies on personal data. "We protect the data of people even though we're not a bank," he says.

The discipline Nahmias continuously returns to is in the sequencing: go slow enough to get the foundation right, then let everything downstream move quickly. Alpura did the whole first phase with 13 people, and he frames the takeaway not as a milk story but as a formula any enterprise can copy. Govern the data, centralize the semantic layer, then let the agents compound. "It's about governance of data first, and then it's exponential what you can do with AI, not the other way around," he says.