Share
Explore
AI Agents, Including China's Manus, Will Not Replace Workers For a While
Yesterday, we woke up to yet another proclamation of a technological revolution. This time, it was an article on Forbes heralding the arrival of Manus AI, an autonomous agent developed by a Chinese company that supposedly "changes everything" for workers worldwide. It sounded ominous, and while it does highlight notable technological advancements, the article overlooked several critical points — so we can dial back on the hype: human jobs are not at risk, at least not from Manus.
The Promise of Manus AI
According to the Forbes article, Manus AI can:
At first glance, this sounds impressive. An AI that can initiate tasks and make decisions? It's no wonder the article claims everything has changed.
The Reality Check
However, as someone who has been closely following AI developments, especially in the area of software for enterprises (as both a vendor and user for nearly 20 years), I can't help but feel a sense of déjà vu. We've seen these grandiose claims before, and they often overlook crucial differences between agents that can function in a silo (i.e. schedule your vacation, look at resumes) versus workflows that involves multiple decision makers — which is pretty much all business workflows.
The Difference Between Doing Work and Making Decisions
Here's where I believe the Forbes article, like many others, makes a critical error: it conflates doing work with making decisions. There's a vast difference between the two, and AI agents are still far from emulating human decision-making in enterprise settings.
Why? Because companies have intricate, intertwined processes that are unique to their operations. It's unlikely that even a dozen neural network models could accurately codify this behavior into an agentic system. Jumping from one step in a workflow to another, in enterprises, are nuanced with variables that can change depending on the input that is important for that specific instance.
Uniqueness Pose Challenges for Generic Agent Models
For example, a seemingly simple task: send a quote to a potential customer. So lets program an AI agent add the prospect to the CRM, ask a few questions about potential buying volume, do a D&B search to qualify them into a price tier, and send them the quote, right? Wrong.
Few businesses work that way. Most new accounts are vetted by at least 2 and often times more people in the company (i.e. sales person, sales manager, accounting manager..and more if the account is potentially a large account). They get input to decide whether to do business with a potential customer based on whether it’s a right fit; decide if this prospect a competitor to existing customers that would risk existing relationships; could this be a fishing expedition—asking for a quote so the prospect can calibrate their internal metrics without any desire to actually get into a business relationship; and many other considerations unique to that business.
The Human Element in AI Decision-Making
What many fail to realize is that the quality standards and metrics used by AI to accomplish tasks are all established in advance by human judgment. Without this, Manus wouldn't know how to proceed through a workflow.
Consider this example: How many years of Java experience does a VP of Engineering want for a particular project? This requirement might change based on circumstances:
Human judgment adapts to changing circumstances. AI, as it stands today, does not.
The Same Business Process May Require Several Agents
Let's take a look at another simple workflow example: putting slow-moving inventory on discount. Even this basic premise can result in several routes and options:
a. Discount item "A" but still try to make some margins because 25% of stock is returnable for credit.
b. Discount item "A" to reduce stock, but after X days readjust pricing if it’s selling too slow or too fast.
c. Discount item "A", but don’t let pricing affect item "B", unless “B” is slowing down too, then bundle them.
For more complicated business processes, the variables multiply exponentially. Add in costs (e.g. should we print new price tags), time constraints (do we have time to reprogram the POS system at every store), and contract obligations (is there a minimum MSRP from the manufacturer), and you begin to see the complexity that AI must navigate even for a seemingly straight-forward objective.
What AI Agents Can't Do (Yet)
Based on what we've seen from Manus and other companies' announcements, AI agents are not yet capable of:
This particularly challenging for companies attempting to provide AI agents in enterprises with thousands of workflows and processes to learn and then create customized models for each process. Adapting ML models to reflect these workflows and optimize best practices for each customer will be a monumental task even for the biggest software vendors today.
The Long Road Ahead
Coding business processes into AI using neural networks will be a long-term task, similar to implementing CRM and ERP systems. Even after 30 years, ERP implementations remain complicated and time-consuming.
Why? Because:
We don't foresee one AI system handling the complexity of even one department or horizontal process, let alone running thousands of workflows across an entire organization
.What to Expect in the Near Future
In the coming years, we're likely to see:
The Bottom Line
While AI agents like Manus represent significant progress for individual workers in specific silos, they're far from replacing human workers in complex decision-making roles of enterprises. The nuances of business processes, the need for adaptive thinking, and the challenges of codifying institutional knowledge mean that human workers will remain indispensable for the foreseeable future
.So, workers of the world, you can breathe easy... at least until next Sunday, when another article will undoubtedly proclaim that another AI agent has changed everything, again.
Want to print your doc?
This is not the way.
This is not the way.
Try clicking the ⋯ next to your doc name or using a keyboard shortcut (
CtrlP
) instead.