AI Agents vs AI Workflows: Agent Washing in Action.

2025 was supposed to be the year of AI Agents. That's what all the headlines said. That's what every vendor promised. But what we're actually seeing is the year of AI Workflows.

And honestly, it's time we cleared up the confusion.

Gartner shared some research and found that only about 130 of the thousands of agentic AI vendors out there are actually real. Out of thousands. The rest are "agent washing" - they're just slapping the word "agent" on things that aren't actually agents.

If you're scrolling through LinkedIn right now, most of what's being shared as "AI agents" are actually AI workflows or basic automations in disguise. People are posting about their amazing "agents" and if you look under the hood, it's a workflow with some AI steps.

And here's a big part of the problem: when Microsoft, Google, and Salesforce start calling everything an "AI agent", it sets a precedent. It ripples through the entire market. Suddenly, it becomes an industry blanket title. Every vendor follows suit and every consultant adopts the language, which leads to the term losing all meaning.

Some in the industry think this distinction is trivial. I think it's a credibility crisis for AI. What's being touted on keynote stages by executives about what an "agent" can do rarely matches what engineers actually build, or what customers ultimately experience. Which is a gap that has real costs and presents a number of challenges.

Too many companies chase AI agents for the optics. It looks innovative, sounds impressive, and makes for a great slide deck. But behind the scenes, it's often smoke and mirrors - pilots that don't scale, no measurable ROI, and nothing that actually transforms how work gets done.

AI theatre distracts from the real opportunity: using AI to make workflows faster, smarter, and more human-centred.

Nobody can agree on what an AI agent actually is. Marketers call anything with ChatGPT an agent. Engineers talk about autonomy and goal-seeking systems. Vendors use whatever definition helps close deals.

When a term means everything, it ends up meaning nothing. This confusion makes it almost impossible for executives to compare products or benchmark capabilities, and it fuels the cycle of overpromising and underdelivering.

The agent hype has kicked off a gold rush, and not the good kind. Everywhere you look, there are $500 "Become an AI Agent Consultant" courses and no-code bootcamps teaching people to build simple workflows and rebrand them as agents.

The result is a flood of undertrained consultants, poor implementations, and broken client trust. When businesses get burnt, they don't just lose faith in one freelancer - they lose faith in the entire AI category. That damage takes years to undo.

Here's what most keynote slides leave out: fully autonomous agents are not production-ready for most businesses yet.

What's actually working are AI workflows - structured and predictable processes that sit between basic automations and true agents. They work because they are deterministic, controllable, and cost-efficient.

Real agents that can operate independently across complex systems are still experimental. The technology is not mature enough for business-critical use cases, and pretending otherwise only widens the gap between expectation and reality.

Many so-called agents are built with no documentation, no audit trail, and no transparency. When something fails, the business can't explain why because no one actually understands what's happening inside the system.

That is not autonomy, that is chaos. For regulated industries, it is a governance nightmare waiting to happen. You can't scale what you can't see.

Executives are being sold the dream of autonomy but delivered glorified automations that still need human oversight. When the results fall short, vendors pitch the next level agent - the one with memory or context - and the cycle repeats.

The fallout is predictable: disappointment, scepticism, and resistance to future adoption. It's not that leaders don't believe in AI; they just don't trust the people selling it anymore.

There are many different definitions. I won't pretend to have perfect clarity on what is/isn't an AI Agent. That definition is still evolving as the world gets to grips with what the technology allows. But here are a number of definitions I've come across that resonated:

Jacob Bank from Relay.app puts it succinctly: the real difference between a workflow and an agent comes down to one question: Who creates the flowchart, and when? Is the human or the agent defining the steps it takes to move through the process?

That question for me is super clarifying and aligns closely with the standard definition of an AI Agent: give it a goal, and tools + knowledge + instructions, and let the Agent determine the path to completing the job.

A workflow? The human defines the path. If you give a system instruction: "If X happens, then do Y. If Z happens, then do A. Check this field, compare it to this threshold, then take this action" - you're defining every step, every condition, every outcome.

That's a programmatic workflow. Even if AI is executing those steps, you've given it the choreography. You've told it: first do this, then check this condition, then do that.

Think about it: you're giving it the playbook. The AI might be smart about executing the plays, but you wrote the playbook for what to do next.

But if you give that same system a goal: "Research this company and put together a brief on their go-to-market strategy", and you don't tell it "First go here, then check this, then compile that" - that's different.

Now the AI has to figure out its own flowchart. It has to decide: "Alright, I'll start with their website. Then I'll search for news. Then I'll check LinkedIn for key hires. Oh, this press release mentions a partnership, I should dig into that. Now I'll synthesise what I've found."

You gave it the destination, not the map. That's an agent. It's building the flowchart in real-time based on what it discovers along the way.

AI Workflow is like a train running along fixed rails. It's built for consistency, precision, and speed. It follows a set route from start to finish, executing each step predictably and efficiently with little deviation. Everything is pre-defined; once it's in motion, it just goes.

AI Agent, in contrast, is more like a car. You tell it where to go, but not how to get there. It makes its own decisions, choosing routes, adjusting for traffic, navigating around obstacles, and even learning from past journeys to improve the next one. It's flexible, autonomous, and capable of creativity, though that freedom means every trip can unfold a little differently.

All that being said, the spectrum gets even more granular when we differentiate between Workflows, AI Workflows, Agentic Workflows, and Agents.

Wade Foster from Zapier talks about the 'AI Automation Spectrum', which is the best visualisation of what's actually happening in the industry right now. It shows the progression from left to right:

Pure Automation sits on the left - everything is deterministic, rule-based execution where you define every single step. This is a pure workflow. It's the classic "if this, then that" automation we've had for years.

AI Workflows come next. This is where you take that deterministic structure and you add AI into specific steps. Maybe you use AI to write a message, or to analyse some data, or to categorise incoming requests. You're still defining the overall structure - first we do this, then we do that - but AI adds intelligence within that structure.

Agentic Workflows follow. This is where multiple AI components work together, making more complex decisions, orchestrating multiple steps. But they're still operating within a framework you've designed. You've given them the playbook. It's sophisticated, it's intelligent, but it's not fully autonomous.

True Agents sit at the far end. These have instructions, knowledge, and tools. They can reason independently. You give them a goal, and they create their own flowchart to achieve it. They decide what steps to take based on what they discover along the way.

Most production systems today sit in the middle - in the AI Workflow or Agentic Workflow space. Not pure agents. And this is appropriate for where we are in the technology cycle. That's where you want to be for most business processes. Because that's where reliability meets capability. You get the intelligence of AI with the reliability of deterministic structure.

The far right (true autonomous agents) - that's still emerging. It's exciting, it's the future, but it's not reliable enough yet for most business-critical processes. About 98% of business tasks today are better served by workflows than by pure agents.

So when someone shows you their "AI agent", ask yourself: where does it actually sit on this spectrum?

Understanding this distinction isn't about gatekeeping - it's about building the right foundation. Here's why it matters:

False expectations waste time and money. Companies buy into the hype, investing in "AI agents" that are really glorified automations. The result is expensive disillusionment.

It protects credibility in the AI industry. Overuse of the term 'agent' dilutes meaning. Clear definitions separate innovation from marketing fluff.

It clarifies the path to maturity. Understanding the progression from automation → workflow → agent helps teams scale responsibly. Most importantly for people thinking about applying AI into their business.

It aligns technology with business value. Automations handle volume, workflows handle complexity, and agents handle uncertainty. Each has a purpose.

It builds literacy and trust among decision-makers. Most executives still can't explain the difference between an automation and an agent, and that's dangerous.

If you're looking to get started with AI in your business, my advice: start with AI workflows. Master them first.

Build workflows that save you time. Workflows that reduce errors. Workflows that make your team more productive. Get really good at identifying where AI adds value within a deterministic structure.

Once you've built 10 or 20 workflows successfully, once your data quality is solid, once your team trusts AI with smaller tasks, once you've debugged them when they break, once you've refined them based on feedback - then you understand what it takes to make AI systems work reliably in production.

And that's when you're ready for the more advanced stuff.

Let me give you a concrete example. Say you want to automate your weekly update to your team. Here's how you might approach it as a workflow:

You've defined every step. You've told the system exactly where to look and what to do. The AI adds intelligence in synthesising and writing, but you've created the structure. That's an AI workflow.

The thing is you're not publishing it automatically. You still review it, you add your personal touches, you make sure nothing sensitive is included, you add context that the AI can't know. But the heavy lifting is done. The research, the recall, the initial drafting - that's automated.

This kind of workflow might save you 30 to 45 minutes every single week. And more importantly, you understand how these tools connect, which builds a muscle and instinct for what's possible. You understand how to prompt AI for consistent output. You understand how to handle the edge cases.

I pulled these recommendations from Wade Foster's thinking on a recent podcast.

1. Look at your calendar right now. What are you doing repeatedly that doesn't require deep creative thinking? What's on your calendar every week or every month that makes you think, "Ugh, I have to do that again?" That's your target.

2. Start with one narrow task. Not "automate all of sales" or "fix all of customer service". One. Narrow. Task. Can you generate a prep doc for interviews? Can you auto-categorise support tickets? Can you summarise meeting notes and extract action items? Pick one specific thing.

3. Build incrementally. Don't try to build the whole thing at once. Get one step working - maybe it's just pulling data from one source. Then add the next step - maybe it's using AI to analyse that data. Then the next - maybe it's formatting it nicely. Each step, you're learning. Each step, you're building on what works.

4. Refine weekly. This isn't set-and-forget. AI workflows need ongoing refinement. What's working? What's not catching the right signals? What rules need adjustment? What prompts need refinement? Make it a habit. Put it on your calendar.

5. Share what you built. Show your team. Walk them through it. This is how you build fluency across your organisation. Not through declarations about AI being important. Through demonstration. Through showing people what's actually possible.

The companies that will succeed with AI aren't the ones chasing the "agent" buzzword. They're not the ones putting "AI agent" on their website and hoping it attracts customers.

They're the ones building systematic, reliable, increasingly intelligent workflows that compound value over time. Week after week. Month after month. They're saving 30 minutes here, an hour there. They're making fewer errors. They're responding faster. They're scaling without adding headcount.

That's the real promise of AI. Not some autonomous agent that magically solves all your problems. Systematic improvement. Compounding returns. Building on what works.

At the end of the day, what matters isn't what you call it. It's whether it solves the problem and makes work more intelligent.