Your Agents Are Flying Blind: Why Agentic AI Governance Can’t Wait
Artificial intelligence is no longer just a tool that employees use – it’s an autonomous workforce operating around the clock, making decisions, calling APIs, and consuming resources at machine speed. And for most enterprises, it’s doing all of this completely out of sight.
In a recent episode of the AWS rethink Podcast, Portal26 CEO and founder Arti Raman and Head of Marketing Neil Cohen sat down with hosts Malini Chatterjee and Preethi Ravala to tackle one of the most pressing challenges in enterprise technology today: how do you govern, secure, and extract real value from agentic AI before it governs you?
The conversation covered everything from the hidden reasons AI pilots fail, to the economics of runaway token consumption, to the regulatory frameworks enterprises need to prepare for right now.
Here are the key takeaways.
From Cybersecurity Startup to AI Governance Pioneer: The Portal26 Story
Portal26’s origin story is rooted in the cybersecurity world. CEO, Arti Raman spent years at Symantec leading product strategy, competitive intelligence, and UX for the enterprise business before identifying a gap in the market that would define her next chapter.
“When ChatGPT hit the market at the end of 2022, the community of CISOs and CIOs around our company basically led us to where we started thinking about an offering that helped companies adopt AI. We consulted with 56 of them, and they said: you really should build a vertical application that helps companies understand how AI is being used.”
– Arti Raman, CEO & Founder, Portal26
That early consultation – with 56 enterprise security leaders – became the blueprint for what Portal26 is today. Launched in 2023 as one of the very first AI governance platforms on the market, the company has evolved in lock-step with enterprise AI adoption: from Shadow AI detection, to full security and compliance coverage, and now into the governance of agentic and multi-agent systems.
Neil Cohen, who joined the team in August 2023, reflects on the scale of the challenge at the time: “Everybody was flying blind – obviously, except for our customers, because they had our tools.”
Why Agentic AI Pilots Fail to Reach Production
Despite enormous investment in generative and agentic AI, the majority of enterprise pilots never make it to production. Portal26’s field data identifies seven to eight distinct reasons for this, and the most surprising one is the one most organisations overlook entirely.
The Top-Down Problem
The single biggest driver of failure is what Arti calls “top-down conception of agentic business workflows” – senior leaders brainstorming AI use cases in meeting rooms, disconnected from the reality of how work actually gets done on the ground.
“What looks like a good idea at the top to someone in management is not necessarily how the work gets done in the trenches. That gap is the single biggest problem with why these things don’t go into production and produce ROI.” – Arti Raman, CEO & Founder, Portal26
With failure rates sitting in the high 80s for top-down AI deployments, the data is unambiguous: enterprises need bottom-up, evidence-based approaches to identify where AI will genuinely create value.
The Other Culprits: Drift, Over-Entitlement, and Excessive Agency
Beyond the top-down problem, Portal26’s research points to a cluster of technical and operational risks that compound over time:
- Agent drift: Agents that gradually deviate from their intended behaviour due to insufficient guardrails. Even a 1-2% drift per step compounds dramatically across a multi-step, recursive process.
- Over-entitlement: Agents with access to far more data and systems than their task requires, creating significant security exposure.
- Excessive agency: Agents empowered to take actions – like issuing discounts or approving applications – with business consequences that were never intended or authorised.
- Compliance gaps: The absence of auditability, transparency, and forensic records needed to satisfy internal governance and external regulators.
Neil Cohen frames the challenge with a useful analogy: “You can’t just play offense and win, and you can’t just play defence and win. When you start looking at preventing bad things from happening and enabling productivity and KPIs, often the way things are set up right now, they become at odds.”
Measuring Agent Productivity: The KPI-First Approach
Portal26’s answer to the productivity vs. risk tension is a KPI-impact model for measuring agent value. Rather than evaluating agents in isolation, the platform maps agent performance to the specific business outcomes they were deployed to achieve – whether that’s improving sales team conversion, accelerating marketing output, or reducing fraud detection time.
“Working backwards from what the agent needs to do gives you a really good sense of which KPIs you’re trying to impact. It also gives you a sense of whether certain high-risk actions are actually needed in that agentic workflow – and how to prioritise risk reduction in that context.” – Arti Raman, CEO & Founder, Portal26
This approach has a practical consequence: it brings security and business stakeholders into the same conversation. When risk reduction is measured against KPI impact, the CISO and the CFO are, as Neil puts it, “singing from the same hymnal.” Security decisions no longer happen in isolation from business objectives – they become part of the same optimization problem.
Shadow Agents: From Hidden Risk to Sanctioned Asset
One of the most compelling themes of the conversation was the real-world discovery of Shadow Agents – unofficial, employee-built agentic workflows operating entirely outside of IT and security oversight. Portal26’s field experience reveals just how prevalent, and how valuable, these underground innovations can be.
Real-World Examples from the Field
Arti shared three categories of Shadow Agent activity uncovered by Portal26 customers:
- Marketing teams build agents to generate content from customer, product, and market data – often uploading sensitive commercial information into consumer AI tools in the process.
- Fraud detection analysts in financial services using AI to triage high volumes of transaction data, inadvertently exposing personally identifiable information.
- Sales teams using agentic workflows to automate RFP responses, pulling from repositories containing sensitive roadmap, pricing, and customer data.
What’s notable is that these employees were not acting recklessly – they were trying to do their jobs better. The problem is that they were doing it without visibility, governance, or security guardrails. Portal26’s Value Realization feature – built to customer specification – maps these discovered workflows to sanctioned, governed equivalents, effectively turning Shadow AI from a liability into an innovation pipeline.
“We found not just shadow agents – we also found repeated workflows. Many people in the same large company were trying to do the same thing, working into the same repositories, paying for the same tokens twice.” – Arti Raman, CEO & Founder, Portal26
The Five Agent Risk Categories – And the One Enterprises Are Most Blind To
Portal26 has identified five categories of agentic AI risk: excessive agency, rogue agents, data security, agent drift, and agent threats. Of these, Arti identifies drift as the risk that enterprises are currently least prepared for.
“If you had asked me six months ago, I would have said excessive agency. But today the biggest risk that enterprises are blind to is drift. There is no such thing right now as agent lifecycle management. Enterprises want to know: is it time to retire this agent? Retrain it? Add harder guardrails? That is the biggest blind spot.” – Arti Raman, CEO & Founder, Portal26
Neil Cohen adds a vivid picture of what this looks like in practice: “People just have these agents in their building and kind of ‘set it and forget it.’ They have no visibility into what the agents are doing, what APIs they’re calling out to. When we show our customers a full visualisation of every agent and every prompt it’s working from – their minds are blown.”
The implication is clear: point-in-time security scanning is not enough. Enterprises need continuous, real-time monitoring of agent behaviour across the full lifecycle – from deployment through to retirement.
Token Economics: The Hidden Cost of Unmanaged Agents
Agentic AI introduces a category of financial risk that most enterprises are only beginning to understand: uncontrolled token consumption. When agents operate autonomously – and especially when they communicate with other agents – the cost of a single runaway loop can wipe out a month’s budget in under an hour.
“I’ve received calls from people saying: ‘You will not believe this – I went for a 45-minute meeting and came back to find my entire monthly budget consumed’ because an agent-to-agent resourcing loop had exploded.” – Arti Raman, CEO & Founder, Portal26
Neil Cohen recounted an equally striking anecdote: a potential customer had built two agents that ended up in disagreement with each other and burned $750 in tokens in a single session.
Portal26’s Three-Layer Token Control Model
Portal26 has designed a token control module around three customer-validated mechanisms:
- Loop detection: Monitoring for recursive agent conversations that fail to converge, flagging accelerating token consumption patterns in real time.
- Anomaly benchmarking: Comparing a task’s current resource consumption against its established baseline to detect anomalous usage automatically.
- Human authorisation for high-cost workflows: Enabling teams to pre-authorise expensive but legitimate workloads, preventing unnecessary alerts while maintaining oversight.
The platform also provides licence intelligence: full visibility into which AI tools and licences an organisation holds, how they are being used, and whether the enterprise is over- or under-provisioned. As Neil notes: “Like any other area of business, you have to manage these costs to a real business objective. These tools are just imperative.”
Human-in-the-Loop: Where to Place the Guardrails
As agents take on more complex and consequential tasks, the question of where human oversight belongs becomes critical. Portal26’s conversations with enterprise customers have surfaced three consistent principles:
1. Reversibility
If an action has significant business impact and cannot be undone – a customer-facing decision, a financial commitment, a policy change – a human must be in the loop. Full stop.
2. Auditability
Natural breaks in an agentic process should align with traceability and forensic requirements. Human checkpoints are not just about control; they create the audit trail that compliance and legal teams will ultimately need.
3. Accountability
Ultimately, a human remains accountable for every agent’s actions. Portal26 observes that most large enterprises are building governance structures in which every agent – or every step in an agentic workflow – is tied to a named human owner.
Neil offers a pragmatic view of where early AI adoption will find the most traction: “The organisations having success are using agents to eliminate toil – repetitive tasks that don’t require genuine human judgement. The next level, where agents become true partners to humans, is going to require a lot more thought. But given how fast things are moving, that moment is probably closer than any of us think.”
Preparing for Agentic AI Regulation: What Enterprises Need to Do Now
Regulatory frameworks for agentic AI are evolving quickly – from Singapore’s IMDA guidelines to the EU AI Act to emerging agency-level guidance in the United States. Arti’s guidance to enterprises is clear: don’t wait for the regulations to arrive before building your governance infrastructure.
“One of the key lessons from regulatory enforcement over the last five years is that intent really matters. If you can show that you took due care and did due diligence, even if you have a bad outcome, you receive more grace than if you never prepared at all.” – Arti Raman, CEO & Founder, Portal26
Portal26’s Agentic AI Governance Checklist
- Build in governance, don’t bolt it on. Auditability, forensic logging, and immutable vaulting must be part of the architecture from day one.
- Cross-correlate with regulatory frameworks. Every agent action should be analysable for compliance with relevant regulations and data handling requirements.
- Design for reversibility. Non-reversible decisions require reduced autonomy and mandatory human sign-off.
- Apply least-privilege principles. Agents should only have access to what they strictly need – exactly as with human users.
- Deploy prompt injection protection. Monitoring for injection attempts, whether successful or not, provides critical protection and regulatory evidence.
- Disclose agent interactions. Users interacting with an AI agent must be informed they are doing so, a basic requirement across multiple emerging regulatory frameworks.
As Neil cautions, the pace of regulatory change itself is a risk: “Every organisation needs the agility to adjust as regulators deploy new rules. You can’t be fully reactive, you have to take proactive steps, because there’s a lot of whiplash going on, and many organisations just don’t have the flexibility to react in real time.”
The Bottom Line: Operational Governance Is No Longer Optional
The message from Portal26’s appearance on the AWS rethink Podcast is both urgent and practical. Agentic AI is already running inside most large enterprises, whether or not those enterprises know it. Shadow agents are building unofficial workflows, consuming sensitive data, and taking actions that carry real business and regulatory consequences.
The enterprises that will successfully navigate this landscape are those that shift from passive, periodic security reviews to active, runtime-based oversight – with human accountability built in at every level of the agentic stack.
“Excessive consumption is not the same as value creation. Just because you’re using a lot of AI doesn’t mean you’re creating value from it.” – Arti Raman, CEO & Founder, Portal26
Portal26’s platform is built on that precise insight: the goal of AI governance is not to slow adoption down, but to make it sustainable, measurable, and genuinely aligned with business outcomes. In a world where product-market fit can shift in days, not quarters, that kind of foundation is the only competitive advantage that truly compounds.
To learn more about how Portal26 helps enterprises build secure, governed, and high-ROI AI programmes, book a demo today or explore Agentic AI Management.