The CEO of Nvidia, Jensen Huang, has called 2025 the year of AI agents.

The CEO of Anthropic said that by 2026 or 2027, AI systems will be better than almost all humans at almost all things.

This year told us a different story. 40% of Agentic AI projects will be cancelled by 2027. 90% of Agentic deployments fail within 30 days.

But at the same time, something remarkable is happening. Agent AI companies hit 400% YI growth.

McDonald's is cutting on boarding time by 65% without hiring a single new trainer.

Walmart's agents save hundreds of thousands of dollars by reducing food spoilage.

Make it make sense. The math ain't mathing. Or is it?

Today, I'm digging into the business models and economics of AI agents. The iceberg of invisible costs. the business models that seem to be working and opportunities hiding inside this paradox. Let's dive in.

So, I started my research by following where the money in a gentic business comes from. And there are two emerging patterns.

Business model number one, agent as a standalone product or agent as a service.

Agent as a service is a product dominant, a productled business model, meaning that the agent is a product.

And business model number two, agent marketplaces.

Marketplaces aren't really about agents. It's a marketplace business. Just like Uber or Fiber or Etsy or Airbnb, it's a business of distribution that follows the rules and principles of marketplace economics.

So, let's look at what's working.

McDonald's has introduced a voice activated AI training simulator that guides new employees through tasks as they do them.

Employees get instructions on how to make burgers, take and assemble orders.

The system reacts to variables that happen around an employee in real time and modifies or adjusts instructions based on what they're doing.

Did it change the business? Yes.

McDonald's is reporting about 65% reduction on onboarding time without hiring a single new trainer. There is a 20% increase in the number of candidates completing the hiring process.

Moving on, Walmart.

Walmart's got a creative one. A selfhealing inventory system.

The self-healing part maintains the balance for the supply chain operations.

They detect demand surges, adjust replenishment schedules, reroute products between distribution centers and also do all kinds of optimization activities.

In Mexico City, it redirects products from overstocked warehouses to facilities with shortages.

In Costa Rica, it reduces spoilage of food because it maps optimal delivery routes.

It even analyzes social media and sales data and adjusts supply depending on the products that are trending.

And lastly, Mercedes-Benz and their MBUX virtual assistant, which is a conversational agent that they placed into select cars.

A driver can talk to the agent in plain natural language, and the agent provides highly contextual, highly personalized recommendations. things like guide me to a fine dining restaurant.

I cannot wait for Google Maps to release the same.

If you look at what's happening under the hood of a system like MBUX, it's not just one model doing everything. It's a system of agents built on top of Google Cloud and Gemini.

And that's a fantastic example of an agentic implementation that works. It takes a manual time-consuming process away from the driver and makes their life easier and in many ways safer.

And Google is doing the same thing but for businesses.

Google has been quietly cooking up a beast in the background called Google Gemini, a space for enterprises.

And they've done it in the Google way. They don't scream about it. They just drop it and let everybody else catch up.

The beauty of it is that it provides exactly what agents are good at. assist, empower, and support, not replace.

It's designed to help every person in every organization remove repetitive, boring tasks and reallocate that time to high impact work that actually moves the business.

It allows companies use Google's AI models and readytouse agents across multiple teams, sales, HR, engineering, marketing, product.

So, there is a solution for all teams across the company.

And here is the part that is often invisible. Bringing and deploying a system like Google Gemini into your company takes a lot of infrastructure and change management.

And that's exactly where Promeo comes in. Promeo is an official Google Cloud partner that deploys entire agentic systems like Google Gemini and they deploy them into enterprises.

Promeo is an official Google Cloud partner that deploys entire agentic systems like Google Gemini and they deploy them into enterprises.

They take care of everything from licensing, setup, training, governance, and team optimization.

Promeo helps companies turn AI pilots into real production grade systems.

And by doing that, they overcome the stages where the vast majority of agent pilots fail, security, compliance, and deployment.

Remember that 90% failure figure that I said at the very beginning.

So, if you're looking at it and thinking, okay, but who's in the 10% that get it right?

That's exactly the space Primo operates in.

If you're leading an AI initiative or overseeing an entire enterprise transformation to AI, you can learn more about it right here at primo.com/ai.

But here is where the story gets really interesting.

The CEO of NVIDIA, Jensen Huang, has called 2025 the year of AI agents.

The CEO of Anthropic said that by 2026 or 2027, AI systems will be better than almost all humans at almost all things.

Well, 2026 is around the corner.

And to all of you who are watching this, how many of you have been replaced by an agent?

Write it in the comments. I'll wait.

Yep, I figured as much.

This week, Dwaresh Patel published a podcast with one of the most influential researchers in AI, Andre Carpathy, a Canadian scientist, by the way, who shared a very sobering assessment of the industry's progress, including AI agents.

He was very vocal about how agents produced very brittle and unpredictable results, and how they lack basic reliability, how they don't possess reasoning, and how they don't learn unless you go and train them by hand.

He was very vocal about how agents produced very brittle and unpredictable results, and how they lack basic reliability, how they don't possess reasoning, and how they don't learn unless you go and train them by hand.

The market of Agentic AI is not just weekly commercialized. We haven't even scratched the surface of where it can go.

The agents we have today are great helpers.

They do automate basic tasks. They can help you do research asynchronously.

They can eliminate a lot of admin tasks or save you plenty of time doing routine work.

I can speak for our channel. For example, those of you who have been watching us know that we treat this channel as a startup and the agents really help us save money on the headcount.

We use Notion Agent, Google Agent, Perplexity Agent. They are incredibly helpful in what they do.

They format documents. They create to-do lists. They assign tasks. They do all kinds of work that we would otherwise be doing by hand. A lot of admin work.

And that is exactly what modern agents are good at. basic repetitive monotonous tasks, but they're nowhere close to autonomous operations.

They don't have cognitive abilities. They don't learn and reflect on past experiences.

Everything they do needs to be checked, validated, and refined.

And what I am covering right now is individual/s small business use case.

The real revenue lies in B2B.

And B2B adoption is miles miles away from where it needs to be.

It is miles away from a point when anyone can say that an agent can realistically and reliably forget the word replace at least cut somebody's workload by 20%.

Agent adoption in B2B is very slow and it is slow because security teams block those deployments and there is a reason for it.

90% of AI agents fail within 30 days of deployment at enterprises because they cannot reliably or even unreliably handle messy and unpredictable business operations.

But as I kept researching failed rollouts, I found something very strange.

The 90% that failed were all trying to save money and the 10% that succeeded weren't trying to save anything at all.

The best example is Harvey.

Harvey hit $100 million in annual recurring revenue in August 2025, which is a $400 Yi growth rate.

They charged $1,200 per attorney per month. That is 10 times more expensive than traditional legal software.

They've got 500 enterprise customers and their customers are doubling seat count within 12 months.

How is it possible that the most expensive agents are the most successful?

Harvey's got a high cost, hightouch model.

Yes, they charge $1,200 per attorney per month and they do it with 12 months 20 seat minimum contracts.

They've got 10% of their team consisting of ex lawyers making sure that the firms which is their customers hit the usage threshold for renewal.

They provide multimodal orchestration.

They position their product around preconfigured agentic workflows.

So they make it very clear what the agent is going to do and how it's going to automate a certain step.

Agents make sense when task predictability exceeds 90%.

Agents make sense when decision logic is simple and when you require zero errors.

However, if each problem, each instance is unique. If each instance requires reasoning across unstructured data, if it involves a lot of natural language interaction and it improves through continuous learning, agents may either not make sense or they need to be deployed with variable costs.

Harvey's example is striking because they're in the minority of Agentic businesses that managed to find a way to make it work.

So, I kept digging into the budgets to understand why Harvey worked.

And here is what's really unusual about agents as a business model. Every software product you have used, Salesforce, Slack, Zoom, Microsoft Suite, they all compete for the IT budget.

And the IT budget is usually around 2% for any company, 2% of their total spending.

But AI agents is the first technology in history that competes for the labor budget. That is 60 to 70% of companies total spendings.

Let me put that in perspective with real numbers.

Think about how a typical law firm spends money.

$45 to $50 goes to labor, lawyers, staff salaries, benefits, stuff like that.

$2 goes to technology. It's all kinds of software that that law firm is using.

Harvey isn't competing for that $2 technology budget. They're competing for the $45 to $50 labor budget.

So, when a law firm looks at Harvey's $1,200 a month price tag, they're not comparing it to the $2 that they allocated towards technology.

They are comparing it to a firstear associate at $13,000 a month. That's $150,000 salary plus benefits.

When you're competing for labor budgets instead of software budgets, expensive becomes cheap.

This leads me to my final question in this research. Are agents supposed to be saving money.

Agentic business and agents as a service are fundamentally incompatible with traditional SAS economy because in SAS model once infrastructure is deployed adding an extra user costs near zero for the business because software can be replicated infinitely at minimal cost.

Because in SAS model once infrastructure is deployed adding an extra user costs near zero for the business because software can be replicated infinitely at minimal cost.

Going back to the examples that I provided before, at Shopify, every new business that wants to become Shopify's customer doesn't cost Shopify anything. They acquire them at almost zero cost.

In an AI agent model, marginal cost is very far from zero. But in the AI business model or AI agent model, marginal cost is very far from zero.

Every action burns GPU compute and energy and costs do not trend to zero even at scale.

Compute costs in AI is the primary variable in aentic products and they come in two primary ways per interaction basis and per token basis.

Foundation models like GBT4, GBT 5 charge per 1,000 tokens.

But Agentic systems can consume from five to 20 times more tokens than simple chains.

Because agents have loops, retries, multi-step planning.

Every routing decision, every tool selection, every context generation can trigger multiple LLM calls.

And that is what becomes your multiplier.

And going back to our example, here is what McDonald's doesn't publicize. The 65% reduction in onboarding time cost them $12 million to deploy across 200 locations.

That's $60,000 per location and that is more than a new hire will earn in 2 years.

The agent in their case did not replace labor cost. It frontloaded it.

On top of this, there is a layer of invisible costs that don't count as agentic costs.

But if you do the proper math on how much it costs to deploy an agent, it should count.

And that bottom of the iceberg that you don't count is in the pre-eployment data work.

You have to curate and prep training data because the agent is only as good as the data it operates with.

In addition to that, you need rag knowledgebased construction.

It requires embedding generation, chunking strategy design, and semantic indexing.

You have to be really smart with the amount of context.

You have to dduplicate data sets and get rid of irrelevant context because if you don't, your storage can go up close to 25%.

And that's more tokens that you'll be burning through, more unnecessary tokens you'll be burning through and paying for it.

On top of it, add data costs, cloud infra that can cost you between $2 to $10,000 and storage optimization.

This is why agents are not and should not be marketed as automated labor.

So we're approaching the end of 2025, the year of AI agents.

And the reality is there is a huge hype reality gap in enterprise and business agentic AI adoption.

AI agents come with significant often invisible overheads.

From technical debt to integration costs, security is the number one concern because autonomous systems increase possibility of cyber attacks that were not predicted before.

Nevertheless, we are at the very beginning of a gentic era and even now there is already a series of emerging opportunities in Agentic AI.

Listen closely.

Agent ops is slowly forming into a new business function.

It is not mainstream yet, but there has been an analysis of 3,000 plus AI job postings, and there are already signs of job requirements around agents, orchestration, and multi- aent workflows.

There are already at least 17 agent ops tools on the market, and there is evidence of major platforms embedding agent ops functionalities into their core offerings.

Now, one quick thing that isn't databacked. It's just my personal opinion.

Personal opinion based on experience and observations.

I personally think that the agent ops as a function is going to be a lot more spread than DevOps for example.

I don't think it is going to be such a centralized function as DevOps because fixing debugging rerouting agents is going to be a lot more accessible accessible to people on non-technical team and there will be lots of agents across all kinds of departments customer support project management finance even accounting maybe

and there will be lots of agents across all kinds of departments customer support project management finance even accounting maybe

in addition to agent ops there is a space that I would dare to call particularly hungry for startups and ideas and it is infrastructure layers for agent management.

In other words, tools for teams like agent ops.

There is nothing on the market that can match the quality of tools that are available for DevOps.

There are fragmented tools, but there is no clear market leader.

So to all entrepreneurs, if you're looking for ideas, this one is still very green, but definitely worth pursuing because agent ops is definitely going to be a thing.

We are approaching the end of 2025 and it wasn't the year when agents replace people.

It was however the year when we learned what agents actually are. Not cheap labor, but a new category of software that competes for labor budgets instead of it budgets.

It was the year we learned we've got a long way ahead of us when it comes to Agentic AI.

2026 is around the corner. Let's see what it's going to bring us.

These are really interesting times we're building in.

As always, we hope this was helpful.

Let us know what you think in the comments.