AI Agents in the Enterprise: Hype or Substance?

Will Companies Really Benefit at Scale from AI Agents?

👋 Hello Technocrats!

They told us Siri would be this magical copilot that would get all our work done for us. Alexa was kind of the same. Both AI Agents failed to make a big impact. Will LLMs change that in the Enterprise?

Cheers & let’s dive in! 🦈

Bobby

AI Agents and the term “agentic” is everywhere these days. It’s so intense that you can’t escape it — talk of agents in the Enterprise has reached monumental levels.

Here’s the Google Trend report for the word “agentic.” You can see it took off around February or March of 2024 and so far shows no sign of slowing down.

Vendors drumming up AI Agent hype tell us that enterprise workflows will be transformed with agents.

In fact, they promise automations so powerful as to completely revolutionize how businesses operate.

Maybe, but somehow I doubt that’s how things will go in the next 2 years.

AI Agents are Not New

Remember when United Airlines ✈️ came out with their automated customer service agent in the 2010’s? It could understand complex spoken requests, perform tasks and direct callers.

And that was more than a decade ago.

But guess what you’re STILL using to book tickets and arrange flights? Either a website like Expedia or a service rep from the airline.

So human behavior (in a lot of ways that are important to the airline business model) didn’t really change with United’s advanced agent.

In fact, perhaps agents saved United money, but both Delta ($61B) and American Airlines ($53B) are neck & neck with United ($57B) it in terms of revenue.

Clearly, the “agentic” revolution did not happen as the United Airlines product managers & technologists envisioned it.

Siri & Alexa

Let’s take another look at the so-called agents, Siri & Alexa.

This is a consumer use-case but it’s still insightful in terms of human psychology.

Theoretically Siri and Alexa have a ton of capability, however the reality is that less than 1% of their functionality is actually used by consumers.

These 2 agents may be powerful and even have their own identities that semi-humanizes them, but most of us simply don’t care.

Alexa was envisioned as an all-encompassing household assistant, yet for most users, it’s been reduced to a glorified weather checker. We have a $400 Alexa speaker sitting in our bedroom that only ever tells us the time and temp!

Siri was supposed to redefine mobile interfaces, yet most users still tap and type on their phones instead of using voice commands. We have a $1,000 smart phone with Siri just to tell us “OK, I’ll call Mom…” every Sunday.

Why Haven’t Agents “Clicked”

Siri, Alexa and a bunch of other AI agents never really clicked the way our imaginations assumed they would.

But why?

It’s a simple answer when you think about it.

It because it takes too much effort to manage AI Agents to do the things we want them to do & their data sources are also stinking piles. 💩

In fact managing AI Agents is more difficult than managing people — at least people understand each other fairly well.

With an AI Agent you have to speak its unique “language,” understand its nuances, deal with its LLM-driven unpredictability, remind it of your preferences & idiosyncrasies, and so forth.

Managing AI Agents in their current state takes too much work. 😛

7 Challenges to AI Agents in the Enterprise

While AI agents have historically failed in the consumer markets, with the current LLM advances they do have potential in the Enterprise context.

This potential however is mitigated by several significant hurdles.

Here are 7 key challenges that organizations must navigate when adopting AI Agents.

1. Agent Autonomy is a Misguided Notion in the Enterprise

AI Agents are marketed as “autonomous”, suggesting that these systems will handle both themselves & entire workflow optimizations without human involvement.

But in reality, true autonomy is extremely difficult to achieve. AI will struggle in enterprise environments where workflows are highly complex & there are many, many dependencies.

Another driver for lack of achieving autonomy is regulatory and compliance constraints. Many industries, from finance to healthcare, have strict, complex and shifting regulations.

This makes autonomy in the way we imagine for AI Agents, where they complete tasks without any supervision, an unrealistic goal in the near term.

2. Poor Enterprise Data will Ruin AI Agents

Enterprise data is very messy.

Ask any DBA that has had to set up an ETL process from one system to another. It’s never clean & it’s never easy (no matter what the sales people at Snowflake promise.) 😃

If AI Agents thrive on the knowledge they glean from data and the data is a confusing mess how effective will these agents really be?

If the opposite of agents is just your run-of-the-mill enterprise platform (like Hubspot) then it will be easier to manage the data through that than an AI agent.

We’re already use to it.

The LLMs behind AI Agents are also a bit mysterious in how they reason — so adding bad data into their “fuzzy logic” is not a good mix.

3. Managing AI Agents is Hard

Even when AI agents are successfully implemented, they introduce a new layer of management challenges: maintenance, monitoring, and governance.

The assumption that AI will reduce human oversight is fundamentally flawed—managing AI agents requires just as much effort, if not more, than managing people.

They need continuous training and tuning to improve accuracy, monitoring to detect errors, bias, and security risks, and governance oversight to ensure compliance with regulations and company policies.

Rather than eliminating work, AI shifts the burden of humans from direct execution to constant supervision and optimization.

4. Talking is an Inefficient Form of Communication

One of the biggest limitations of AI agents is that they rely heavily on natural language interactions, which, while intuitive, are often inefficient compared to other forms of communication.

In enterprise settings, employees are accustomed to structured interfaces, dashboards, and direct manipulation of data—all of which allow for faster, more precise inputs and outputs.

A well-designed UI can provide instant access to relevant information, while an AI agent might require multiple conversational steps to arrive at the same conclusion.

This inefficiency is why AI agents may work best as embedded features rather than standalone assistants.

5. The Political Effects of Agents in an Enterprise Are Real

AI agents won’t just introduce new technology—they will disrupt workplace dynamics and power structures.

When AI takes over tasks traditionally managed by employees it will create friction, uncertainty, and resistance.

Teams that once held ownership over key functions may feel sidelined, while managers may distrust AI-driven recommendations that challenge their authority.

At the leadership level, AI can create tensions between departments. If an AI agent recommends cost-cutting measures that impact headcount, HR and Finance may clash with engineering, for example.

Similarly, AI-driven performance insights can expose inefficiencies, putting pressure on managers to justify decisions in ways they didn’t have to before.

Boards and CEOs may push for AI adoption to drive efficiency, but middle management—who must implement it—will resist due to fears of oversight or job erosion.

6. Legal & Compliance are Unchartered Territory for Agents

AI agents operate in a legal gray zone, and no one really knows what happens when they screw up.

If an AI leaks sensitive data, makes a bad financial call, or violates compliance rules, who takes the fall—the company, the vendor, or the AI itself? Regulators haven’t figured it out, but lawsuits are inevitable.

For industries like finance, healthcare, and government, AI decisions can trigger audits, fines, or worse.

Compliance teams now have to police AI like they would rogue employees, ensuring it plays by the rules.

And if you think it's bad now, just wait—AI laws are changing fast, and what’s legal today might be a lawsuit tomorrow.

7. Agents Create MORE Work, Not Less

Vendors like to sell AI agents as the ultimate workforce replacement—why pay for people when you can have tireless, intelligent automation?

Sounds great, except for one problem: AI agents don’t actually work that way.

Instead of eliminating jobs, they create a new category of work—babysitting AI that constantly needs retraining, debugging, and explaining when it inevitably screws up.

Rolling out AI agents isn’t cheap either.

Companies sink massive budgets into them only to realize they still need humans to monitor them, correct their mistakes, and make sure they don’t violate compliance laws.

And when AI spits out nonsense or makes an “automated” decision that lands in legal hot water, guess who has to clean it up? Not the AI, that’s for sure.

Closing Thoughts

The cold truth?

Right now at least, a well-designed UI and traditional automation can do 90% of what AI agents promise, without the headaches.

But despite the skepticism and hurdles, AI agents are not a lost cause—far from it. The enterprise world is always evolving, and AI is undeniably part of the future.

While today’s agents may not be the revolutionary, fully autonomous workforce replacements vendors claim, they can still bring immense value when deployed strategically.

History shows that disruptive technologies take time to mature.

The early days of cloud computing were filled with doubt, and yet now, no serious enterprise operates without it.

AI agents today may feel like clunky, unreliable assistants, but as LLMs improve, data integration challenges are addressed, and organizations learn how to effectively manage them, their potential will grow.

Companies that start experimenting now—without buying into the hype—will be best positioned to reap the rewards when the technology reaches its next level of maturity.

AI agents are not a magic bullet, they may not even be a step forward right now now, but they are worth experimenting with.

The key is to remain realistic about what they can and can’t do for your use-cases.