Managing enterprise AI costs: Q&A with PointFive co-founder and CEO Alon Arvatz

Key Highlights

• Enterprises adopting AI face a growing, hard-to-track consumption layer of AI models. Startups like PointFive are helping these companies control related costs.
• Citi Ventures is investing in PointFive for its strong team and platform addressing both cloud and AI costs.
• PointFive finds savings in system architecture, not just rate discounts.

Cloud has become a foundational layer of the modern enterprise, and as cloud adoption has scaled, organizations have developed a need for dedicated tools and practices to understand and optimize their cloud spend. We are seeing a similar dynamic emerge with AI today.

As enterprises accelerate their adoption of AI, a new layer of infrastructure and consumption is growing alongside it. The result is a new and compounding cost challenge: organizations are spending more on infrastructure, while much of that spend is difficult to identify, attribute and ultimately eliminate. Traditional FinOps tools can surface where costs are growing but often stop short of connecting those insights to the engineers who can address the underlying inefficiencies and automate solutions. That is the opportunity we see in PointFive.

PointFive is a cloud and AI efficiency company helping organizations automatically identify and solve hidden infrastructure costs. The company began by tackling cloud cost optimization, a large and established market that continues to grow. PointFive is now extending its expertise into the rapidly emerging world of AI cost optimization. Its end-to-end platform moves beyond reporting to autonomously detect inefficiencies across cloud and AI infrastructure, remediate them, and verify the impact.

We are excited to invest in PointFive as it helps enterprises build the infrastructure and operating discipline needed to scale both cloud and AI efficiently. I recently chatted with PointFive co-Founder and CEO Alon Arvatz about our investment and other topics on the horizon related to AI and cost.

The Opportunity

Ziv: In one sentence, what is the core opportunity in helping enterprises understand and control what their infrastructure and AI actually cost?

Alon: Every enterprise can tell you what it spent on AI last month, but not many can tell you what that money actually bought.

Cloud spending grows at roughly 20% annually. In our customer environments, AI spending can multiply five and ten times year-over-year. That is why the conversations has shifted almost overnight from "How do we adopt this" to "Where is the money actually going?" The opportunity has two parts. First, eliminate the spend that yields zero value, which in enterprises AI environments runs close to 40%. Then, tie the remaining spend directly to business outcomes, enabling leadership to measure what a given dollar of AI investment actually produced. Most of the market is still working on the first half. The enterprise value rests in mastering both. .

The Surprising Result: Real Architecture vs. Rate Discounts

Ziv: What is the most surprising savings outcome a customer has realized after deploying PointFive?

Alon: The speed-to-value. Nubank saved over $15 million on a single database service, achieved entirely without rate discounting.

What surprised the industry was not the headline number, but the mechanism behind it. There was no price negotiation, no reserved instance or savings plan restructuring, and none of the conventional levers FinOps teams traditionally reached out.

The optimization was deeply architectural, uncovered inside a system with an exceptionally sophisticated engineering organization already considered highly optimized, all while scaling from 99 million to 119 million customers.

However, the metric I find even more compelling is velocity: Nubank reduced their average time to resolve a waste finding from 210 days down to 16. Dollars are always open to debate, but collapsing a resolution backlog from seven months to just over two weeks represents a fundamentally different operational posture. That is the kind of efficiency that compounds over time.

From Cloud to Agents

Ziv: PointFive started in cloud infrastructure, expanded to data platforms, and is now addressing AI and agents. How has customer demand evolved, and what visibility does PointFive have at the token layer that legacy tools miss?

Alon: A finding like Nubank's is only possible when you examine how a system is built, rather than merely how it is billed. The fundamental problem hasn't changed, only the surface on which it operates.

Cloud infrastructure taught us that waste is an architecture and configuration challenge, not a reporting problem. You do not fix waste in a BI report, you fix it in the system architecture. Data platforms presented the same problem one layer higher in the stack. AI presents that problem once again, except with far less visibility than anything before it. In cloud environments, you can at least see the resource; with an AI, you get an aggregated invoice line item labeled "Tokens."

What shifted is where the spending happens. It no longer originates solely in centralized cloud accounts. It happens locally on thousands of developers' laptops, triggered inside coding agents multiple times every minute.

To solve this, we built TokenShift to run directly on the machine the agent executes. That positioning is decisive: provider invoices only show token consumption after the fact. A gateway sees requests fully assembled. Only the endpoint sees the request being constructed, where system prompts and tool definitions get attached and re-sent on every single turn.

This endpoint vantage point unlocked critical insights: about 80% of the bill consists of prompt-cache traffic, the agent repeatedly paying to re-read context it already had before the user's task even began. Conventional context compression tools can barely address 5% of the total bill. Nobody looking at a month-end invoices would ever find that.

The other thing the endpoint gives you is complete cross-ecosystem visibility. Model providers only report to you what you consumed with them, yet few enterprises rely on a single agent. The workstation is the only place where all of that activity converges. Cloud is where we learned the method, but it is not where the hard part is anymore.

Why Citi Ventures?

Ziv: What made Citi Ventures the ideal strategic partner for PointFive's next chapter?

Alon: We wanted an investor whose operating environment mirrors the enterprise scale we serve.

Citi has absorbed every major technology paradigm shift over the past century, which is rarer than people appreciate. From mainframes to distributed systems, the internet, cloud, and now generative AI. Institutions that successfully adopt emerging technologies early gain competitive advantages, while governing tightly enough to ensure institutional resilience. That is exactly the balance our product has to hold. Citi Ventures has been doing this for well over a decade, across hundreds of companies, and they bring the operator's perspective alongside capital.

Practically, what we build has to survive within environments that are massive, heavily regulated, globally distributed and mission critical. An investor living withing those operational constraints every day will tell you what is wrong with your product faster and more precisely than a board deck ever will. That is worth more to us than the capital.

Working with financial institutions

Ziv: How does controlling AI costs inside a large, regulated financial institution differ from a technology company, and what do enterprise environments demand that standard FinOps tools fail to deliver?

Alon: In a tech company, the question is typically "What did this cost?" In a regulated institution the questions are: "Who authorized this spend, against which budget and can you show an audit trail?"

Enterprise customers tell us the same things: First, it has to be read-only and agentless, because nothing gets to sit in the data path. Second, it has to produce a record that survives an audit, not just a chart. And it has to seamlessly integrate with the change controls that already exist rather than around them.

AI is testing the same principle in a new way. Most tools that want visibility into agent spending ask to act as proxies, to sit in front of the model and inspect traffic on the way through. In tier-one financial institutions, security teams tend to reject proxy models, partly because of what it means for security review and partly because you have just put a dependency in front of every developer's work.

PointFive's endpoint analysis inverts this paradigm. The prompt content remains where it was written, while only the derived metadata is transmitted. That is the AI equivalent of a read-only architecture, and it is what gets the conversation past InfoSec instead of ending there.

Moreover, AI adds something new on top. Spend is now initiated by something that is non-human, while every control institution's build assumes a human made the decision. When a coding agent initiates a session in the middle of the night, governance controls must bind dynamically to the agent identity and a pre-approved budget envelope, otherwise, governance breaks down entirely.

The Road Ahead: What CIOs Should Do Tomorrow Morning

Ziv: As spend shifts toward AI inference and token consumption, how do you see the market evolving, and what should a CIO be doing about it tomorrow morning?

Alon: Because most of AI cost are decided before a single token is generated, the market is shifting from compression to control. And the only place to enforce control prior to generation is where the agent runs.

The levers that work are governance ones. Budgets. Choosing the right agent for the task instead of the strongest model for everything. Routing between models. Deciding what personal use of a company AI account looks like. Compression still has a place, but only where it targets the reasoning nobody sees.

This is usually where people brace for a surveillance conversation, but our research showed the exact opposite. We studied 2,000 U.S. workers this year, the most striking finding was not how much of company AI is being used for personal work, though it is a lot. It was that employees want automated guardrails. They recognize that substantial spending is happening directly from their keyboards, and they would rather have a budget and a policy than find out later that they were the line item somebody questioned.

Ultimately, enterprise leadership needs a clear metric: cost per unit of useful work. Calculating this is harder than it sounds. In traditional software the system tells you whether a transaction succeeded, so you divide the bill by the successes and you have your number. An AI agent always returns output. Determining whether that output was any good is a judgment, not a readable signal you get from the system, so pricing a single call is impossible. What you can do is sample thousands of tasks and compare cohorts month-over-month.

Here is the gap I would point at. Today, every large enterprise now maintains cloud cost efficiency leaderboards and budget targets by team. In AI, almost none of that exists yet. That gap will close over the next 12 to 24 months, and the early movers will spend less and get superior output per dollar spend.

For a CIO looking to take immediate action tomorrow morning: pick one high-volume workflow powered by AI agents. Find out what one completed piece of work costs today, and identify who decided it was worth paying for. Many organizations cannot answer either half of that question, and the second half is the one that should worry them.

For more information, email Ziv Idan at [email protected].

To learn more about Citi Ventures and our portfolio, click here.