A Framework for Dialing In Agentic AI

The gap between agentic AI’s promise and reality became stark in the fall of 2025.

Deloitte submitted a $440,000 government report filled with fabricated citations and invented research.
Salesforce disclosed a critical vulnerability (CVSS 9.4) allowing data exfiltration through prompt injection.
More than 700 organizations were compromised in the Salesloft Drift breach.
A Replit agent deleted a production database despite explicit warnings not to touch anything.

These aren’t edge cases. They’re production failures at named organizations, and they reveal a pattern: the same AI-based systems promising efficiency and speed create new accountability gaps when deployed without adequate controls.

These controls, including ethical guardrails, aren’t in tension with strategic flexibility. In reality, they’re interdependent. Organizations that maintain clear controls can move faster precisely because they’ve defined boundaries. Those chasing speed without structure will see their projects end up in the 40% of projects Gartner predicts will be canceled by 2027.

In this post, you’ll learn practical strategies for harnessing agentic AI safely, setting boundaries, keeping humans in control, and turning these AI guardrails into competitive advantage.

The Intelligence Hierarchy That Enables Control

Before deploying agentic systems, establish clear responsibility and control through a hierarchy such as the following:

Collective Intelligence sits at the top of this hierarchy. This is your organization’s accumulated wisdom, values, and judgment about what matters. It isn’t one person’s strategy; it’s the organizational capacity to discern purpose and navigate complexity together.

Human Intelligence provides the judgment, creativity, and relational capacity that only people offer. This is where meaning gets made, where context matters, and where wisdom lives.

Artificial Intelligence executes, analyzes, and scales where humans can’t. It is excellent at patterns, but incapable of purpose. It’s powerful and fast, but unable to bear responsibility.

Organizations experiencing major incidents this year inverted this hierarchy. They let algorithms dictate strategy, treating AI recommendations as authoritative rather than advisory, and automating decisions without maintaining human accountability. Deloitte’s consultants used AI to fill “gaps” but appeared to never verify the output (ranging from misquoted judge comments to fabricated case law). The AI performed as designed. The governance and human oversight failed.

Agency Dials

You can’t control AI’s pace of advancement, but you can control how you deploy it through the following six dials. They are adjustable based on task risk, organizational readiness, and strategic priorities.

Ownership ↔ Recommendation

The risk: Your team implements an AI recommendation because “the AI said so.” Six months later, nobody can explain the rationale.

The dial: Make human ownership explicit. Require decision documentation: Who reviewed this? What alternatives were considered? What judgment broke the tie? Implement conviction thresholds. For example, routine decisions can execute autonomously, but significant decisions must include reasoning and wait for human confirmation.

The insight: AI can analyze endlessly. Only humans and our organizations can be accountable.

Speed ↔ Insight

The risk: AI processes customer in an hour, synthesizes them, and automatically posts the summary on your website. Your team knows what the report says but understands nothing about the humans behind it. The patterns are clear and the sentiment analysis is complete, but the actual voice of your customers—their frustrations, their praise, their specific contexts—disappears into aggregate data.

The dial: Create workflows where velocity serves insight. Before synthesizing 10,000 reviews with AI, require manual reading of 50. After AI delivers analysis, schedule time to challenge and interrogate it. Build contemplation into delivery timelines. Establish review checkpoints before any AI-generated content goes public.

The insight: Organizations metabolize truth slowly—through conversation, reflection, integration. Speed without understanding creates the illusion of productivity while eroding institutional wisdom.

Augmentation ↔ Replacement

The risk: Your analysts cannot analyze at scale without AI. So when the system encounters a problem outside its training and errors out, the team can’t act.

The dial: Require core skill mastery before AI assistance comes into play. Use AI to expose teams to 10x more examples but require humans to do pattern recognition. Regularly rotate “AI-free” work as capability maintenance. Train the team in fundamental AI competencies before exposing them to higher-order system work.

The insight: Humans remain fundamentally in charge (remember the top of the hierarchy). When the Replit agent said deleted data was “gone forever,” the user recovered because he possessed foundational knowledge the AI lacked.

Execution ↔ Development

The risk: Productivity climbs, but your employees aren’t learning. Your organization is completing tasks while failing to develop talent.

The dial: Experiment and establish ideal ratios of AI and human work for your work. For example, guarantee 2 hours of human-led skill development for every 10 hours of AI-augmented work. Design mentorship into workflows. Consider how your system requires domain expertise to not only operate but also improve, maintain, and assume responsibility for the system.

The insight: Organizations trading development for efficiency discover too late that they’ve created brittle, dependent, irresponsible workforces. Take charge, dial in for long-term gains and not just short-term replacement.

Transparency ↔ Opacity

The risk: Your agentic system saves millions. Nobody understands it. Then it fails catastrophically, and no one can explain why.

The dial: For significant recommendations, demand to see reasoning chains, data sources, and assumptions. Build review rituals: “What data might the AI have missed? What bias might be embedded?” Train algorithmic literacy across leadership.

The insight: You cannot govern what you cannot see. The OpenAI ChatGPT “ShadowLeak” vulnerability executed malicious instructions automatically, with server-side exfiltration that no traditional tool could detect.

Autonomy ↔ Oversight

The risk: Terrified of errors, you implement exhaustive approvals. Innovation stops. Talented people leave.

The dial: Create risk-tiered autonomy—low-stakes decisions run freely, high-stakes require human judgment and documentation. Replace “never use AI without approval” with “use AI responsibly, document your process, be prepared to explain.”

The insight: The answer isn’t surveillance. It’s building cultures where responsible use is easier than reckless speed. IBM found 20% of breaches involve shadow AI, yet only 37% of organizations have policies to manage usage.

What Actually Works

Despite failure rates of 70-95% on basic business tasks across major benchmarks, organizations are seeing genuine results with AI agents. Adopt these 5 practices to dial in your agentic work:

1. Perfect your scoping

Scope Agentic AI projects appropriately for the organization (for example, thinking carefully about targeted piloting versus “giving everyone ChatGPT”).

2. Orient to Human-AI collaboration over replacement

The best implementations augment human expertise with insight or speed instead of attempting full automation of processes.
Start with tasks that naturally include a (human-in-the-loop) HITL decision or approval to minimize process disruption and pattern good review and auditing practices when using AI.
Take an approach of freeing teams to perform higher-value work and AI performing routine tasks.

3. Ensure integration and compatibility with existing systems and processes

Many failed AI solutions take a traditional, engineering-heavy approach instead of playing to AI’s strengths as a highly integrateable technology. It’s important to choose tools or strategies that are out-of-the-box compatible with existing business systems. This approach allows you to build AI systems that don’t rely on highly technical subsystem development, such as API connectors. For example, if an existing system manually operates via a series of emails, consider embedding AI intelligence into the process organically instead of taking an API-connected approach.

4. Hone culture

Organizational strategy and cultural development are key components of implementation: Training, policy, discussion, implementation performance tracking, and measurement.

5. Establish governance and accountability

Start from a comprehensive, mature set of policies and controls surrounding data privacy, security, decision authority, and liability. This becomes your fabric for AI to be woven into, instead of being considered after the fact.
It’s important to establish clear ownership of responsibility and data, escalation paths, and auditability.

From these five practices, the path to dialing in Agentic AI unfolds:

Build evaluation infrastructure first (treating it as the unit test for agents)
Match agents to existing processes
Focus on specific pain points
Invest in data governance before deployment
Redesign full processes rather than just automating steps

The Strategic Advantage of Ethical AI

Research from Gofast.ai indicates enterprises lose $67.4 billion annually to AI hallucinations. Carnegie Mellon found top LLM-based agents fail over 70% of the time on basic business tasks out of the box. Yet the path forward is straightforward.

The dials are your framework for distinguishing what you can control from what you cannot. They ensure your organization remains deliberately human-led while effectively leveraging autonomous systems.

Organizations that establish clear guardrails can move faster precisely because they’ve defined where judgment lives, where humans remain essential, and where machines serve collective intelligence. Those chasing speed without structure become cautionary tales.

Every agentic system you deploy represents a statement about what you trust, what you value, and where you believe accountability belongs. The winning organizations won’t be those with the fastest AI. Instead, they’ll be those with the clearest controls enabling strategic flexibility.

Interested in Learning More?

Watch Julian’s Agentic AI foundational series to learn how you can apply this in real time.
Hear from our CEO on why our newest video series is designed to bring immediate, applicable value to your organization.
Reach out to Julian with any additional questions!
Read more about our Agentic AI solutions.

Julian Lancaster

As Chief Information Security Officer at Prowess Consulting, Julian Lancaster brings a grounded, refreshing, and practical perspective to the Agentic AI series. A passionate advocate for responsible AI adoption, Julian focuses on building foundational understanding of how large language models (LLMs) work, and how teams, individuals, and organizations can leverage them to increase efficiency, scale capacity, and drive smarter decision-making. With a strong background in cybersecurity and enterprise operations, Julian helps demystify AI technologies so they can be used effectively and securely across the business.