The Agentic AI Trap

Agentic AI has become the buzzword of the moment. Every major technology vendor is racing to position their products as the gateway to this autonomous, decision-making future. Marketing collateral gleams with promises of AI that doesn't just assist, it acts. It reasons. It orchestrates. It handles complexity without human intervention.

And I get it. The vision is compelling.

But here's what I've learned from spending the last 13/14 months deflecting the agentic AI question: the most mature organisations aren't focused on agentic AI, instead they're focused on automation and integration.

You might ask: what's the difference? But let me explain.

The Four Ways to Build an AI Solution

In my head it's been clear for a while that the unencumbered pursuit of agentic AI has been the wrong way to think about intelligent automation, but I didn't quite know how to frame it correctly. That was until I came across this painfully simple framework (the best frameworks always are).

It maps AI solutions across two dimensions: integration (how deeply the solution connects to your existing systems and data) and automation (how much the solution can do without human intervention).

This creates four distinct categories:

I.e. without automation and integration, there are no agents.

Assistants: The Starting Point

Assistants sit in the bottom-left corner - low integration, low automation.

They're the simplest form of AI deployment: a chatbot that answers questions from a knowledge base, a tool that helps users find the right information without needing to search manually.

Think of a basic FAQ bot on your service portal. It doesn't create tickets. It doesn't update records. It doesn't trigger workflows. It just helps people find answers faster.

Many organisations dismiss this as too basic, or not exciting enough. The point is that this is where users learn to interact with AI. They learn when it works best, and when it doesn't. And, more crucially, it's where your organisation learns what people actually ask about. It's where you discover gaps in your knowledge content that you never knew existed.

This step can't be skipped, it's where you start building the conversational design muscles that every subsequent AI initiative will depend upon.

Copilots: AI as a Collaborative Partner

Move up the integration axis and you arrive at Copilots - high integration, but still relatively low automation.

These are AI capabilities that work alongside human users within existing tools and workflows. In the ServiceNow world, this might look like Now Assist suggesting resolutions to agents based on similar historical incidents, or summarising long ticket threads so someone can get up to speed quickly.

Microsoft's Copilot suggesting email replies in Outlook follows the same pattern.

The AI isn't taking action autonomously. It's surfacing information, making suggestions, and letting the human decide what happens next. The integration is deep because it understands context from your systems, but the automation is deliberately constrained.

This is where many organisations can extract significant value with manageable risk. You get the benefits of AI-augmented decision-making without the complexity of fully autonomous processes.

Autopilots: Automation Without Deep Integration

The bottom-right quadrant - high automation, low integration - is an interesting one. No one seems interested in these anymore...

These are AI solutions that can handle routine tasks independently, but don't require deep hooks into your systems. A simple example might be a chatbot that guides users through a standard process like resetting a password or submitting a request. The steps are well-defined and the AI can orchestrate the conversation without needing to query multiple backend systems in real-time.

ServiceNow Virtual Agent topics that walk users through straightforward request submissions often fall into this category. The automation is high (the user completes their task without human intervention), but the integration complexity is relatively contained.

Agents: The Autonomous Ideal

And then there's the top-right quadrant. Agents. The Holy Grail.

These are AI solutions that combine high integration and high automation. They can understand context from multiple systems, reason about complex situations, and take autonomous action be that creating records, triggering workflows, updating data, orchestrating processes across integrated platforms.

A true agentic solution might handle an order refund end-to-end: verify the customer's identity, check the order history, apply business rules, process the refund, update relevant systems, and communicate the outcome; all without human intervention.

It sounds incredible. And it is, when it works.

But hopefully this illuminates the obvious truth: agents are the hardest place to start.

Why Agents Are the Hardest Place to Start

The allure of agentic AI is understandable. If you're going to invest in AI, why not aim for the maximum impact? Why settle for assistants when you could have autonomous agents?

The answer comes down to compounding complexity.

Agents require two things simultaneously: deep integration across multiple knowledge sources and systems and robust automation that can handle edge cases gracefully. Each of these is challenging on its own. Together, they multiply the difficulty, and the risk.

Pretty much every client conversation I've had started with agentic ambitions and ended up in "Autopilot" or "Copilot" mode. The integration work required for ambitions beyond that can take months, and that assumes the knowledge and system maturity exists in the first place. Edge cases that seemed trivial in planning become black holes that consume development resources.

And the result? The project never sees the light of day. Or it does but it's either so late or been descoped so heavily that stakeholder confidence has evaporated. Or it runs with so many guardrails that it's barely more autonomous than a well-designed form.

The organisations that wanted to lead the AI transformation end up with little to show for their investment.

The Case for Starting Simple

So here's what I've learned works: start in the bottom-left quadrant and work your way up deliberately.

Deploy an assistant. Get it in front of users. Learn what they ask about. Understand where your knowledge gaps are. Build the conversational design capabilities within your team.

Then, as that foundation solidifies, you can start adding integration. Connect your assistant to live data. Let it surface real-time information. You're moving towards copilot territory - AI that works alongside users within existing workflows.

Simultaneously, you can start building automation capabilities. Take those well-understood, frequently-asked topics and add fulfilment. Let the AI not just answer the question but complete the request. You're moving towards autopilot territory - handling routine tasks end-to-end.

And here's the crucial insight: all of this capability stacks.

The conversational flows you design for your assistant become the foundation for your copilot interactions. The automation workflows you build for routine requests become the building blocks for more complex orchestrations. The edge cases you discover and handle at each stage don't need to be rediscovered when you eventually tackle agentic use cases.

You're not delaying your journey to agentic AI by starting simple. You're building the road you'll travel on.

Resisting the Hype

I want to be clear: I'm not anti-agentic AI. I think there are genuinely transformative capabilities emerging in this space and when implemented thoughtfully, can deliver remarkable value. The roadmap towards more autonomous AI-driven service delivery is real and worth pursuing.

But as has always been true with technology-driven transformation, the path matters as much as, if not more than, the destination.

Every major technology vendor right now has a vested interest in selling you the agentic dream. The marketing collateral is polished. The pre-recorded demos are impressive. The vision is compelling.

What the pre-recorded demos don't show is that this was attempt 20. Or the months of integration work, the data quality issues that surface only in production, the edge cases that turn elegant solutions into fragile systems. Or the organisational change management required to actually adopt these capabilities.

My advice? Take a step back. Ask yourself what's realistic given where your organisation is today. Look honestly at your data quality, your integration maturity, your team's conversational design capabilities, your users' readiness to engage with AI.

Then start where you are, not where the marketing says you should be.

The Dividend of Patience

The organisations I've seen succeed with AI are the ones that aren't calling it AI.

They understand that a simple assistant deployed and adopted today is worth more than a sophisticated agent stuck in development indefinitely. They recognise that every improvement compounds, that the work done building out automation flows for a Virtual Agent lays the groundwork for the agentic capabilities they'll build later.

They resist the pressure to skip steps.

And paradoxically, they often reach agentic maturity faster than organisations that aimed there directly. Because they built the foundations. Because they learned from real user interactions. Because they solved the integration challenges incrementally rather than all at once.

Start simple. Build foundations. Let the capability stack.

The future will still be there when you're ready for it.