Ozgur Akaoglu Says the Hardest Part of Enterprise AI Starts After the Model Works

A working model can make an AI project look further along than it really is. It can predict churn, forecast demand, or surface a customer signal with impressive accuracy. Inside a slide deck, that may feel like progress. To Ozgur Akaoglu, co-founder of Hilbert, it is only the beginning.

“The model is not the product,” Akaoglu says. “A model that works in a notebook is not the same thing as a system that changes what a business does.”

That distinction sits at the center of Hilbert, the AI-powered growth intelligence company Akaoglu co-founded for retailers and e-commerce brands. The company builds systems that read customer data, identify moments that matter commercially, and use AI agents to take action before those moments pass.

The promise sounds clean. The work behind it is not.

Enterprise AI has entered a moment where nearly every company wants agents, automation, and faster decisions. The word “agentic” has become part of the sales pitch across the industry. Akaoglu believes much of the market is still stuck in the performance stage.

“Most companies are not running agentic AI,” he says. “They are showing agentic AI. There is a big difference between an agent that can complete a demo and an agent system that can operate inside a real enterprise.”

The gap matters because live companies are not controlled environments. Their data arrives late. Their customer definitions vary by team. Their compliance requirements can reshape a system. Their schemas are messy. Their edge cases surface after deployment, not before it.

That is where many AI projects lose momentum. A model may identify something useful, but the surrounding system cannot translate that output into a business response. The result may sit in a dashboard, wait for a team meeting, or get buried under other priorities.“

The last mile is where a lot of AI projects die quietly,” Akaoglu says. “The insight exists, but it does not change the business fast enough.”

Hilbert was built for that last mile. Rather than handing clients a model and leaving them to handle the hard part, the company deploys end-to-end AI agent systems that connect directly with a brand’s existing data environment. Its work sits inside the operational layer where prediction, decision-making, and execution have to meet.

Akaoglu’s view of production AI comes from building under conditions where failure is not abstract. He studied computer engineering, worked as an early engineer at Getir, and later spent two years with an SF-based startup focused on LLM-based product development. Across those experiences, the lesson was consistent: AI only matters when it can handle real operating conditions.

“At scale, the unglamorous problems decide everything,” he says. “Configuration layers, schema normalization, incomplete data, slow data delivery. Those are not side issues. They determine whether the system survives contact with the business.”

That belief shaped Hilbert’s technical posture. The company does not treat deployment as a final step after the exciting work is finished. Deployment is the challenge. A model can perform well in isolation and still fail once it meets enterprise constraints.

“If your AI cannot get through messy data schemas, delayed inputs, and a compliance review, it is still a demo,” Akaoglu says. “Production is not a detail you add later.”

The current wave of enterprise AI has created a strange tension. Leaders want autonomous systems that can do more than generate text or recommend next steps. They want agents that can decompose tasks, recover from failure, verify trust, coordinate across domains, and operate on live data. Yet the infrastructure needed to make that reliable is often underestimated.

Akaoglu sees this every time the conversation shifts from possibility to execution.“

Real multi-agent orchestration is hard because the business is hard,” he says. “You are not just solving a technical problem. You are matching the system to how the company actually works.”

For retailers and e-commerce brands, that challenge becomes especially concrete. Growth teams do not just need to know that a customer is drifting away or that another customer may be ready to buy again. They need to know what action to take, when to take it, and whether that action is worth taking at all.

Too much communication can exhaust a customer. Too little can let the relationship fade. A broad campaign may reach a segment, but miss the individual timing that actually matters. Hilbert’s work is built around that narrow decision point, where customer behavior, predicted value, timing, and channel choice all have to come together.

“AI that explains what happened is useful,” Akaoglu says. “But the larger opportunity is AI that tells you what to do next and is accountable for the result.”

That accountability is what separates a production system from an experiment. Enterprise clients do not evaluate AI in theory. They evaluate it against constraints, revenue goals, internal workflows, security requirements, and the problems that appear only after integration begins.

The first major client, Akaoglu says, teaches more than planning can.“You meet the edge cases you did not anticipate,” he says.

“You meet the constraints you did not know existed. That is where the real product gets built.”

Hilbert has raised a $28M Series A led by a16z, with participation from Asylum Ventures, ScaleX Ventures, SV Angel, Valkyrie VC, TIBAS Ventures, and Gunderson Ventures. The company works with clients across the US and Europe, and its team spans Istanbul, San Francisco, Barcelona, and London. Its client base includes Walmart, which places the technology in the kind of enterprise environment where production claims have to be earned.

Akaoglu does not frame that as a reason to speak more loudly about AI. He frames it as a reason to be more precise.

“Clients do not care how elegant your architecture is if the system does not work,” he says. “They care whether it can handle their data, their requirements, and the surprises that come with their business.”

For Akaoglu, the AI industry is entering a useful correction. The next phase will belong less to companies with polished demonstrations and more to teams that can connect models, data, agents, governance, and execution into systems that hold up under pressure.That is the work Hilbert is trying to own. For more information on Ozgur Akaoglu, visit the Hilbert website.

That is the work Hilbert is trying to own.“There is a lot of attention on what AI can say,” Akaoglu says. “We are focused on what AI can run.”