Actualyze AI targets the cost and security risks of enterprise AI sprawl

The current AI wave is running into a type of sticker shock reminiscent of what happened when cloud computing went mainstream in the late 2000s and early 2010s. The pitch then was essentially: stop worrying about buying servers and running a data center for capacity you might not need. But it turns out that the elasticity of the cloud, its ability to accommodate however much demand you need frequently led to cost overruns. And for many organizations, their data center never actually went away, so cloud spend landed on top of the existing infrastructure.

The recent rise, and soon disappearance, of tokenmaxxing in AI, which aimed to essentially make a game out of using as much genAI as possible, led organizations ranging from Uber to Microsoft to spend more on AI than they had bargained for.

When the bill arrives before the budget

In many such AI deployments, “the bill arrives before the budget,” quipped Rafi Khardalian, CEO of Actualyze AI, which launched from stealth on August 3 at AI4 in Las Vegas. The company, which has $7 million in seed funding, has developed a platform that sits between an organization’s employees, applications and agents on one side and every model they call on the other.

Khardalian and fellow co-founder CTO Sean Lynch began by asking what they would need to run AI inference inside the kinds of large organizations where they had spent their careers. “What would we need to make it safe? What would we need to make it cost-effective, secure, and have the best model for the job at any given time?” he said. “That’s what we set out to build: the platform that we would want as a customer.”

Both Khardalian and Lynch are engineers. And as such, they began to reflect on the state of the AI market. “The two of us worked together to determine what the market needed, and the idea evolved,” Lynch said.

“There was a great deal of brainstorming and thinking ahead about how the market would evolve,” Khardalian said. “We have talked for a long time about [AI] subscriptions and heavily subsidized inference, and about those subsidies eventually going away. Much of what we anticipated has materialized, so we timed the product and its capabilities around that progression.”

Khardalian said Actualyze had tried to estimate how much inference a flat-rate AI subscription can actually buy. “It is easy to do the math,” he said. “In an analysis we ran a few months ago, someone fully using all of the subscription windows could consume approximately $12,000 worth of inference. That model is unsustainable.”

A central part of the push is to help enterprises optimize AI spending, avoiding the use of an expensive frontier model to draft a routine email or handing a complex coding task to a weaker model. “We look at the per-token cost. We look at all of your teams, and we allow you to create budgets and hierarchical budgets,” Lynch said. “For a project, you can define the inference spend and cap it. If a rogue agent goes completely off the rails, it isn’t going to spend $10,000 or $20,000. It’s going to hit a budgetary cap and stop.”

Actualyze also separates its budget structure from a company’s formal reporting structure, allowing spending limits to follow projects, applications or teams that may cross organizational boundaries. “Teams and budgets are messy in enterprises,” Lynch said. “They need to be completely independent hierarchies that you can map individually. We built that into the product because we came from that world and know how messy it is.”

Their read on subsidized inference was informed by having watched a similar arc from the vendor side. Khardalian and Lynch previously ran Metacloud, a Pasadena company selling managed OpenStack private clouds that Cisco acquired in 2014 for $149 million, according to Cisco’s fiscal 2015 annual report.

Now, Actualyze AI is in a phase of getting early feedback from organizations across a range of industries, including highly regulated ones. “Once we enter a large organization, we want to understand how its needs are changing,” Khardalian said. “ We have several companies lined up at different stages of AI adoption, generally large organizations. The product can also help smaller companies, although the problem is more pronounced at larger ones right now.”

From ‘please don’t’ to enforcement

One of the challenges with a technology that is common in settings ranging from labs and enterprise settings to personal smartphones is the risk of shadow AI. A survey from the lab software vendor Sapio Sciences found that 97% of lab researchers used some form of AI in their work while 45% of those workers were using public AI through personal accounts. Only 32% use company-managed logins.

Lab professionals are part of a much broader pattern. Verizon’s 2026 Data Breach Investigations Report found that regular AI use on corporate devices had climbed from 15% to 45% in one year.

The consequences can add up. IBM’s 2026 Cost of a Data Breach Report found that security incidents involving shadow AI, meaning employee use of unapproved tools, more than doubled to 43% from 20% a year earlier. Breaches in that category averaged $5.39 million, roughly $400,000 above the global average of $4.99 million, and about one in five triggered a regulatory fine. Among organizations that suffered any AI-related breach, 92% lacked basic access controls on their AI systems. Intellectual property, the category most relevant to research organizations, was compromised in 32% of all breaches studied and carried the highest per-record cost of any data type at $196.

Regulated organizations have generally responded with policy. Some bar commercial models outright. Others permit approved tools for non-confidential work and use something of an honor system to ask employees to keep sensitive data out of the prompt. “That’s based on trust, right? ‘Please don’t do that.’ But what happens if somebody forgets?” Khardalian said.

Actualyze applies content and routing policies while each request moves through the platform. “Because all the inference is going through us, we have a purpose-built model that looks for shapes of information that are problematic,” Lynch said. “We see data that is problematic in a HIPAA context. We see PII, PHI, API keys, passwords, personal names, addresses, phone numbers and Social Security numbers.”

The platform can then record the detection or remove sensitive content before sending the request to a model provider. “We can log it, or redact it completely. All the model on the other side sees is the redacted version,” Lynch said. “Because we’re in the middle of the inference path, we can do that.”

Actualyze is entering a crowded category. Palo Alto Networks acquired Portkey, a Bengaluru AI gateway startup that gives enterprises a unified API to thousands of models along with an agent registry, semantic routing and caching, in a deal announced April 30 and closed May 29, folding it into the Prisma AIRS security platform. Cisco, the company that bought Metacloud, has expanded its AI Defense product since its January 2025 launch into AI supply chain governance, controls on agentic tool use and MCP traffic, multi-turn red teaming and runtime guardrails built with NVIDIA NeMo. Kong and Cloudflare now offer AI gateways for routing and governing model traffic, with overlapping features including PII sanitization, token-based rate limits, semantic caching, provider failover and cost analytics. LiteLLM gives teams a prominent open-source, self-hosted option for much of the same work, putting a single OpenAI-compatible endpoint in front of 100-plus providers with virtual keys and spend tracking and budget caps set per key, user or team. Microsoft, Amazon and Google also provide access, security and governance controls within their respective AI platforms.

Actualyze’s believes the AI control layer should not belong to any model provider, and that enterprises will operate across a widening mix of frontier, specialized and open-weight systems. “The greatest benefit for companies around the world comes when inference becomes heavily commoditized,” Khardalian said. “The market should include as many capable model providers as possible, until inference starts to resemble bandwidth.” 

The potential of U.S. restrictions on Chinese open-weight models would add another source of fragmentation. Lynch said reduced access to models such as Kimi K3 and Qwen3.8-Max could accelerate domestic alternatives.

Whatever mix of models enterprises end up running, Khardalian expects the volume of inference flowing through it to keep climbing. “AI touches everything. The key word is intelligence. Demand for intelligence and knowledge has no clear ceiling. We have barely scratched the surface of the ways it will change how people live and work,” Khardalian said.