Yet during a recent conversation between Optiver and
- How should teams be structured when software can be built in days rather than months?
- Where does competitive advantage come from when models are widely available?
- And how do organizations move faster without becoming more fragmented?
Three quotes from David Meyer, SVP of Product at Databricks, capture the scale of that shift.
"AI transformation is really process transformation."
Like many technology companies, Databricks initially saw engineers become more productive through code generation and assistance tools. But over time, Meyer said, it became clear that the greatest gains were coming not from the tools themselves, but from rethinking the processes surrounding them.
"People had doubled, tripled their productivity with AI," he said. "These were 10x engineers becoming 100x engineers because they had swarms of agents working for them."
What changed, in his view, was not simply the technology. It was the relationship between people and the work. "If you're always hitting the next button," Meyer said, "you should try to figure out what your mind is doing before you're comfortable hitting the next button. Training an agent to do that. Otherwise, you are the blocker."
Databricks has already begun changing the way it hires and evaluates talent in response to this development. Prospective field engineers, for example, are now expected to demonstrate how they use AI to solve business problems in real time.
"The key is figuring out how the processes will change in the rest of the functions in order to figure out who you need to hire," Meyer said.
Lance Braunstein, CTO at Optiver, framed a similar challenge at Optiver. While the firm has invested heavily in tooling, education and access to large language models, Optiver is now focused on the harder question: how work itself changes once those capabilities are widely available.
"The cost of code is going to zero."
Software still needs architecture, testing, governance and maintenance, of course. But Meyer pointed out that AI is rapidly reducing the effort required to produce working code, forcing organizations to reconsider where value is actually created.
He pointed to an example from Databricks' own product development process. Building a new connector once required months of planning and development. Product requirements documents took weeks to write. Development stretched across quarters. Today, he said, "with AI they can basically build a connector in a week."
The implication is not merely faster software development. It’s a fundamental change in organizational economics. "Spending 90 days writing a PRD is the definition of inefficiency," Meyer said. It's "bad process design."
Yet greater speed brings a new set of challenges. If software becomes dramatically easier to create, maintaining coherence across an organization becomes harder. Meyer described instances in which teams, empowered by AI-assisted development, moved so quickly that overlapping or even contradictory capabilities began emerging across the product.
"What we've seen is a more disjointed product at times," he said.
Christophe Godefroy, Optiver's Head of Data Platforms, described this as one of the firm's central technology challenges. After decades of operating with significant local autonomy across teams and regions, Optiver is working to build common platforms that reduce unnecessary complexity.
The risk, in other words, is not that organizations fail to build enough, but that they build too much, too quickly and without sufficient alignment. That is why Meyer believes the importance of platforms is likely to increase rather than diminish in the AI era. Shared abstractions, common standards and carefully designed interfaces provide the structure that allows teams to move quickly without pulling the organization in different directions.
"AGI is already here."
Toward the end of the conversation, Meyer offered what he described as "the provocative statement." Depending on how one defines artificial general intelligence, many researchers would disagree with it outright. Meyer acknowledged as much. Yet his point was less about settling a technical argument than about reframing the discussion.
For many organizations, he suggested, the limiting factor is no longer the intelligence of the models themselves. "Typically these models today, the ones that have been released in the last month or so, if you're honest, are smarter than most of your colleagues at general stuff," he said.
But their weakness, he argued, is precisely what makes human knowledge more valuable. "The problem is they don't have any of the context, the training, the architectural skills that your colleagues have."
Earlier in the conversation, Braunstein described this challenge as capturing "Optiver intelligence" — in other words, the collective knowledge embedded across meetings, documents, code repositories, emails and internal discussions. Creating useful AI systems, he suggested, may depend as much on organizing and connecting that knowledge as on selecting the right model.
If the first phase of the AI era was defined by the race to build ever more capable models, the next phase may be defined by a different challenge: connecting those models to the accumulated knowledge of an organization. Making that information discoverable, useful and secure is proving far more difficult than deploying a model.
"The problem to unlock the potential of your company is really getting the business context, the technical context, the domain context available to the models," Meyer said.
