As AI innovation continues seemingly unabated, organizations are moving beyond LLM experimentation and embracing the future of intelligence systems: autonomous agents that can access enterprise data, interact with applications and execute complex workflows.
That autonomy raises a simple question: can enterprises trust agents, protect sensitive information and remain accountable while they work inside business-critical systems all within a single organizational framework?
In a recent PingCast webinar, Securing the Agentic Enterprise: From Pilot to Production, Loren Russon, VP of Product Management at Ping Identity, was joined by Matt Girdharry, Head of Security and Observability Partnerships at AWS, and Pankaj Jain, Managing Director and Digital Identity Lead at Accenture. Together, they explored the trust challenges around autonomous AI.
They covered why promising pilots stall before production, how authorization models need to evolve and how organizations can build governance into the foundations of agentic AI.
Pankaj Jain (Accenture): We've observed that about 70% to 75% of pilots with AI are actually quite successful. But when it comes to deploying them in production, only about 13% to 20% actually get deployed.
We do see challenges in going from pilot to production. I would put these into a couple of main categories.
The first category is that when we build an AI agent, we use a curated data set. But once we go into production, the data is so diverse that we stop seeing the results we want, and we run into those challenges. This is typical with AI adoption: the agent has to be continuously retrained, and the data is also dynamic. And with the amount of money we are spending on legacy platforms, the data is very siloed, and we can't collaborate or work through all the data we have.
The second, and biggest, challenge we face when going into production is something we've seen firsthand in the pilots I have been doing.
Anytime we deploy, we run into challenges and issues around trust. We've got to go through all the security clearances; we've got to go to ethics and everything else, asking: Where is my company data? Which AI LLM are you going to use? How is data getting shared? What are the privacy concerns? All those challenges have to be confronted.
Matt Girdharry (AWS): I saw some data that S&P Global found that about 46% to 50% of AI projects were scrapped between proof of concept and production. I think this was a 2025 statistic from some studies they did with companies. That was up significantly compared with previous years.
I think what we're finding is that while there were a lot of issues companies were encountering around finding the right model, how much it's going to cost, whether they're getting the right outputs based on their inputs and looking at performance, very quickly the questions became less about technology.
Our Bedrock service here at Amazon offers lots of model choices across cost and performance. But when it comes to the guardrails, governance, identity, data protection and all of the IP-related things that could cause significant harm to these companies, they had to quickly say, "our traditional security measures that we have in place to protect our infrastructure and our applications are not going to work in this new world of non-deterministic agent activity that can access lots of things through APIs and the network to get the job done."
We call it the intern problem. Instead of a human intern that you're telling to go out and do something, you now have an agent or a set of agents that's out there that can do significantly more damage and present significantly more risk than your human interns do, where you're saying yes, no, yes, no.
MG: Going along with the theme of trust here, from our perspective, trust is identity plus authorization plus observability plus the ability to instantly revoke things that need to be revoked. All four must execute in real time, especially in an agentic AI model.
You can think about this in four layers.
Agent identity. Each agent gets its own IAM execution role, short-lived STS tokens (15 minutes maximum), and IAM Roles Anywhere extends this to agents outside of AWS.
Dynamic authorization. Permissions are computed at runtime and are not static. This is where evaluating fine-grained policies at the point of action, not just session start, comes in. Which is why Ping and AWS go together like chocolate and peanut butter. I hope that's not an outdated metaphor, by the way; I've been using it for a very long time.
Observability. Capabilities like CloudTrail log every API call, while Bedrock Guardrails enforce content policies, PII redaction, toxicity filtering and topic restrictions uniformly at the platform layer.
Human oversight and the kill switch. The ability to revoke. Every agent has a named human owner, instant credential revocation and the ability to shut it down in seconds if something goes wrong.
PJ: So the way I would kind of answer this is that typically, any kind of AI tooling or agentic AI architecture kind of tooling these days is not doing a single task. It's a multi-layered workflow that's done. And typically, the guidelines that have been given are that you've got to have a human in the loop, okay? When I talk to my customers, I say it's not human in the loop; it's human in the lead. You need to have a human in the lead overseeing this entire process. Unless that is in place, you're taking a risk. So I would say 100%, a human has to be in the lead for this.
Every agent running on your platform in your organization has to be associated with a human identity. Not just associated with a human identity; you need guardrails around it. And the guardrails are saying: who is running this, under what identity, what permissions, what authorizations, what is it doing, and monitoring of it. And that is at runtime.
And this isn't new–we've been running non-human identities, service accounts, and application accounts for a long, long time. Ever since the genesis of identity as a separate field in cyber, we've been doing that. We just have to adopt it for agents. We have to have defined processes to say, when you're building an agent out, these are the things you have to do with the identity team as well. The bottom line is there has to be a clear line of ownership to say this agent is owned by this person in your organization.
PJ: Typically, an identity platform has two parts.
With agents coming into play, you need to define all these aspects and publish processes so that when teams create a new agent, they know exactly what to do rather than building isolated security features.
MG: Picking up on what Pankaj said: a lot of companies go headlong, based on their risk profile, into "We just want to execute this, figure out how it works, and then design a governance and security practice based on mistakes we made". That's an onerous and painful process. On the other hand, I'm working with a publicly traded tech company that literally spent years with consultants like Accenture designing proper policies, guardrails, and compliance regimes before letting employees touch AI/ML for business operations.
Corporate cultures vary widely. Conservative, low-risk-tolerant companies want to secure 99.9% of everything that can go wrong before launch—though they never truly will given the non-deterministic nature of agents. Others run very fast, accepting risk to learn as they go. Counterintuitively, sometimes the most innovative companies are the most conservative, while heavily regulated companies choose to move extremely fast.
As autonomous AI works its way deeper into enterprise operations, identity management must go beyond logging in human users and protecting traditional applications. Agents need unique identities, narrowly scoped permissions and access controls that respond to the context of each task. Just as important, organizations need governance that establishes ownership, keeps agent activity visible and makes it possible to intervene when something goes wrong.
Beyond the necessity of unique identity management, the takeaway for enterprises heading toward production-scale agentic AI is clear: security and governance can’t be afterthoughts tacked on once a pilot succeeds. Organizations should evaluate whether their existing identity infrastructure, security architecture and operational resources can support autonomous systems at scale. That may require a strategy more focused on redirecting investment from maintaining legacy technologies toward modern authorization capabilities, observability and governance.
For organizations that want the benefits of agentic AI, building this foundation of identity and trust may be the single most important step on the road from experimentation to enterprise-wide adoption.
Continue the conversation. Explore more perspectives on securing agentic applications and preparing enterprise security for autonomous AI.
Watch Securing the Agentic Enterprise from Pilot to Production→
ABOUT PARTNER INSIGHTS
Partner Insights is a TechChannels-sponsored content offering developed in collaboration with TechStudio, bringing technology leaders and industry experts together to share perspectives on the challenges, trends, and innovations shaping today’s technology landscape.