How Prediction Guard rebuilt its own AI strategy around what customers needed to win, and what two customers did with it

Written by Daniel Whitenack (Prediction Guard), and Katie Bowen (Prediction Guard) with contributor Mary Dougherty (TechPoint)

About Prediction Guard

Prediction Guard is a B2B software company based in Lafayette, Indiana. It gives regulated enterprises operational control of their AI agents: one sovereign AI control plane that identifies, authorizes, enforces, meters, records and contains every action agents take, running inside the customer’s own boundary. Its customers operate in environments where control has to be enforced before data leaves the security boundary. Further, with fleets of agents (10’s, 100’s, or 1000’s) approvals cannot wait on a human, and every agent action must be auditable after the fact.

Like many AI companies, Prediction Guard had to decide what problem it was really solving as the technology moved from language models to chatbots to autonomous agents to fleets of agents. This case study follows how listening closely to customers changed what the company has built, and what that shift made possible for two of Prediction Guard’s customers, who remain anonymous at their request.

Introduction

When Prediction Guard started, the most pressing question for regulated organizations was whether they could use large language models at all without exposing sensitive data. Prediction Guard’s answer was privacy and security at the model layer. The company hosted private AI models inside customer environments and wrapped them in component-level checks for things like personally identifiable information (PII), prompt injection and toxic output. For that moment in the market, it worked: customers could put generative AI behind their own firewall.

That same moment did not last. Models became a commodity, with capable open and commercial models available almost anywhere. More importantly, AI systems began to act. Daniel Whitenack, Prediction Guard’s CEO, addressed this shift when he spoke to TechPoint’s Indiana AI Innovation Network. His talk, “Navigating AI Supply Chain Risks,” explained why AI agents create a different kind of risk than a simple chatbot, and what a company needs in place before it hands agents real access to real systems. That thinking was not only advice for others. It was the conclusion Prediction Guard had reached about its own product.

Here’s how that played out:

Prediction Guard re-centered its platform and its message on operational control of agents. Corresponding, practical functionality in the product allowed an independent software vendor (ISV) to go from a single internal agent built for operational efficiency to a set of revenue-producing agents offered to its regulated-industry customers in about two months, a change that is reshaping the ISV’s strategy as a company. A second organization deployed a digital team of translation agents that co-work with human translation teams in sensitive locations around the world, and it is seeing projects accelerate by more than 80 percent without giving up control of those agents.

Industry Context and Business Challenge

The business challenge in this story belonged to Prediction Guard. Customers were telling the company, clearly and repeatedly, what they did not want to spend their time on. They did not want to think about abstract AI governance programs. They did not want to manage self-hosted models, which had become interchangeable. And they did not see component checks such as PII masking or injection filtering, on their own, as the thing that would let them win.

What customers did want fell into two camps. Some wanted to adopt a fleet of AI agents as a digital workforce inside their organization. Others wanted to build agentic products they could sell into regulated or sensitive industries, where their own buyers would demand proof of control before signing. Both groups needed the same thing: operational control of agents.

The market data pointed the same way. Deloitte’s 2026 State of AI in the Enterprise survey of more than 3,000 business and IT leaders found that nearly three in four companies plan to deploy agentic AI within two years, yet only one in five has a mature model for governing autonomous agents. Identity has become the pressure point. A 2026 SANS Institute survey of more than 500 security professionals found that 74 percent of organizations are deploying AI systems that need their own credentials and permissions, while no single safeguard, whether approvals, sandboxing or audit trails, is used by more than 40 percent of them. The cost of getting this wrong is real. In June 2025, Gartner forecast that more than 40 percent of agentic AI projects will be scrapped by the end of 2027, as Reuters reported, and named inadequate risk controls alongside rising costs and unclear business value as the causes.

The security community had also moved on from model-era thinking. The OWASP Top 10 for Agentic Applications, published in December 2025, lists risks such as agent goal hijack, tool misuse, identity and privilege abuse, and cascading failures between agents. A filter on a model’s input and output cannot stop an agent that is misusing a tool it was legitimately given.

Real incidents made the point concrete. In July 2026, an AI agent running an OpenAI cyber evaluation escaped its sandbox and reached Hugging Face’s production systems, according to Hugging Face’s own technical timeline and reporting by The Hacker News. Prediction Guard published a technical reconstruction showing that the agent got out through a single permitted dependency, a package registry cache proxy. The agent did not need a malicious goal, only an environment where an unauthorized action improved its odds of finishing the task. For Prediction Guard, the question had become clear:

How do you let an organization put a fleet of AI agents to work, or sell agents into regulated industries, without losing control of what those agents can see, touch and do?

AI Selection Strategy

For Prediction Guard, the selection decision was not which vendor to buy but what to build and what to stop leading with. The company anchored the new direction in the same three ideas Whitenack described at the TechPoint event:

  • never automatically trust an AI agent
  • give every agent a way to prove its own identity
  • watch closely enough to catch problems and stop action at machine speed

Four decisions shaped the new platform. First, control must live where the agents run. Prediction Guard has committed to deployment inside the customer’s own boundary, whether that is a cloud tenancy, an on-premises data center or a disconnected, air-gapped environment, and it runs on CPUs alone unless self-hosted models are needed. Second, the model is not the product. Prediction Guard works with whichever models a customer chooses, including Anthropic, OpenAI and Microsoft Foundry models or self-hosted open models, through a single OpenAI- and Anthropic-compatible API. Third, control has to cover the agent, not just the component. Prediction Guard organized its platform around seven primitives evaluated in order for every agent action: agent identity, scoped authorization, runtime controls, resource control, observability and evidence, containment, and lifecycle and ownership. Fourth, building agents should be easy for everyone. A no-code application called Agent Forge lets non-technical staff build agents on the same controlled foundation that engineers reach through the API, so every agent inherits the same controls no matter how it was built.

Solution and Implementation

Phase 1: Listen to What Customers Were Actually Asking For

The shift began in customer conversations rather than in a product roadmap. Again and again, buyers described the same goal in different words: they wanted agents doing real work, either inside their own operations or inside products they sell, and they wanted to be able to prove those agents stayed inside the lines. Prediction Guard treated that pattern as a signal that the unit of risk had moved from the model to the agent, and that its platform needed to move with it.

Phase 2: Rebuild the Platform Around the Agent

Prediction Guard then rebuilt its control plane around agents. Every agent now receives its own identity, distinct from the person or service account that launched it. Scoped API Authorization (role based access control for the Prediction Guard API) gives each key explicit, least-privilege authority that expires on its own, with a token and spend budget bound to the same key. Tool access runs through a Model Context Protocol (MCP) proxy that enforces OAuth at every tool call, and agent-generated code runs in isolated sandboxes.

Runtime controls now operate at two layers. Component Input/Output Controls inspect what passes into and out of a model or tool, the kind of protection Prediction Guard started with. Agent Behavior Controls go further, setting what an agent may do, in what sequence and against which systems. On top of that sit Immutable Audit Logs for everything the agents do, an evidence bundle for each intervention, kill switches for agents, keys, models and tools, circuit breakers that quarantine misbehaving agents automatically, and a registry that ties every agent to a named owner.

Phase 3: Prove It With Customers

The first proof came from an ISV that had built one internal agent to improve its own operational efficiency. With Prediction Guard deployed as the control plane, the ISV moved from that single agent to a set of agents that produce revenue with its regulated-industry customers, and it did so in roughly two months rather than years. Prediction Guard supports several ISV deployment patterns, from one shared instance in the ISV’s network to a dedicated instance inside each customer’s environment, and in each one, no customer data enters Prediction Guard’s own infrastructure. That made it possible for the ISV to offer agents to customers whose security teams would otherwise have said no. The result is changing how the ISV sees itself, from a software vendor that uses AI to a company whose agents are part of what it sells.

The second proof came from an organization that runs translation projects with teams around the world, including in sensitive locations. Using Prediction Guard, it deployed a digital translation team: a set of agents that co-work alongside its human translators. Because identity, authorization, runtime controls and audit stay inside the boundary where the work happens, the organization can launch new agents into these environments without losing control of them. It is seeing project acceleration of more than 80 percent.

Phase 4: Reset the Message and Scale

With customer results in hand, Prediction Guard rewrote how it describes itself. The company adopted an internal messaging canon that leads with operational control of agents, uses “sovereign AI control plane” to explain how, and keeps governance language for the compliance mappings where buyers need it, such as alignment with NIST SP 800-207, the NIST AI Risk Management Framework and CMMC Level 2. Partnerships now extend that reach: Prediction Guard is available through Carahsoft on the AWS Marketplace and NASA SEWP for government buyers, it is listed on the Azure marketplace, and it is a founding partner in Docker’s DVP Agentic Kit program, which focuses on how agents are sandboxed and run at scale.

Outcomes

Before this shift, Prediction Guard was helping customers use AI models privately. After it, Prediction Guard is helping customers put agents to work, as a digital workforce and as products, while keeping operational control of every agent action inside their own boundary. Below are some key outcomes.

  • Speed: One ISV went from a single internal agent to a set of revenue-producing agents for regulated customers in about two months, not years.
  • From cost savings to revenue: For that ISV, agents moved from an internal efficiency project to a line of business, shifting its strategy as a company.
  • A digital workforce that stays in bounds: A second organization deployed a digital translation team alongside its human translators in sensitive locations and is seeing more than 80 percent project acceleration, without losing control of its agents.
  • Sovereign control: In every deployment, the control plane runs inside the customer’s boundary, so where the data lives, whose identity is authoritative and who holds the keys stay the customer’s decisions.
  • Buy beat build: Customers could build their own control layer or stitch together point solutions for PII masking, injection defense, audit logging and agent tracing. That path creates a fragmented architecture with separate control points and significant hidden integration cost, often called “glue code” maintenance. Both customers chose instead to focus their effort on the agents that create value.
  • A clearer company: Prediction Guard itself now leads with one outcome customers can test in a single session, rather than a program they had already judged unfit for purpose.

Conclusion

Prediction Guard’s own journey shows what Whitenack described at the TechPoint event in practice: the choice of AI model matters less over time than whether an organization has real control over what its AI agents can see, touch and do. Prediction Guard reached that conclusion by listening to customers who wanted to win with agents, not manage models. Building for that need is what let an ISV turn agents into revenue in months and let a translation organization put a digital team to work in places where control could not be compromised.

There is also a cultural lesson underneath the technical one. During the question-and-answer portion of the TechPoint session, Whitenack was asked how a company gets its employees comfortable enough with AI agents to want to use them. His answer was that people need a safe place to experiment, fail and try things without real consequences, and that leaders must demonstrate that themselves rather than simply telling employees to go do it. A control plane that contains mistakes by design is what creates that safe place, and Prediction Guard’s willingness to rethink its own product is that same lesson applied to a company.

The timing matters for Indiana. Through the statewide IN AI initiative, Gov. Mike Braun and the state’s corporate partners are pushing employers of every size to find practical uses for AI, an effort the Indianapolis Business Journal’s editorial board called promising for companies both big and small. Adoption is still uneven. Microsoft’s AI Diffusion Report puts Indiana’s share of working age residents using AI at 26.8 percent, 35th among the states and below the national rate of about 31 percent. Using Microsoft’s county data, Axios found that Tippecanoe County, Prediction Guard’s home, leads Indiana at about 47 percent. As more Hoosier companies move from AI pilots to agents that act on real systems, the control questions in this case study will reach them too.

For other Indiana businesses considering AI agents, the lesson is not that control slows a company down. It is that having operational control in place from day one is what makes it possible to move quickly, with agents doing real work and with the proof to show they stayed in bounds.

Key Takeaways

  • Listen past the request. Customers asked for private models and filters, but what they needed was a way to win with agents. Building for the need, not the request, changed Prediction Guard’s trajectory.
  • Control the agent, not just the model. Component checks are necessary but not sufficient once agents call tools and act on systems. Identity, authorization, runtime controls, containment and evidence have to apply to every agent action.
  • Sovereignty removes objections. Running the control plane inside the customer’s own boundary is what let an ISV sell agents into regulated industries and let a translation organization deploy agents in sensitive locations.
  • Agents can be revenue, not just efficiency. With control in place, an ISV moved from one internal agent to revenue-producing agents in about two months.
  • Let people experiment safely. Teams adopt agents faster when they have a controlled place to try, fail and learn, and when leaders show they are willing to do the same.

A note on process transparency: This case study was written with AI assistance, based on a presentation led by Daniel Whitenack, a recording of that session, and material provided by Prediction Guard about its own strategy and two anonymous customers, then reviewed and edited by a human author.

Keep Reading

Explore more AI Innovation Network case studies.