The market is exploding with AI-powered automation tools and bots, known as “AI agents.” These agents streamline business tasks and reduce manual effort. However, the overwhelming volume of options—hundreds of agents targeting various tasks—creates a new challenge: navigating this crowded ecosystem to deploy AI effectively and avoid creating new problems that confuse employees and add to technical debt.
The Proliferation of AI Agents
AI agents now dominate nearly every business function, from customer service to coding and operations. The rapid growth of these agents (via aiagentsdirectory.com) is evident:
- September 2024: 258 agents
- October 2024: 321 agents
- November 2024: 444 agents
- December 2024: 650 agents
- January 2025: 804 agents (at the time of this recording)
This staggering increase demonstrates the proliferation of AI agents but also foreshadows the challenges enterprises will face. As the number of agents grows, IT and procurement teams will struggle to evaluate and integrate them effectively. Deploying these agents to automate tasks without a cohesive strategy risks repeating the mistakes enterprises made with standalone apps, leading to disconnection, inefficiency, and increased operational silos. Many will find themselves managing too many isolated agents that fail to communicate with each other, reinforcing silos and complicating workflows.
Most AI agents excel at specific tasks but fall short of streamlining broader processes that span teams and systems. Enterprises will eventually recognize that a platform approach, which prioritizes interoperability and orchestration, offers a far superior long-term solution to achieving seamless automation and avoiding the pitfalls of fragmentation.
The Challenge of Deployment
The real hurdle isn’t finding capable AI agents; it’s ensuring they integrate smoothly into existing workflows. Take customer support as an example: separate agents handle documentation, troubleshooting, and escalations, often requiring employees to switch between multiple tools to complete a single task. This fragmentation leads to delays, miscommunication, and inefficiency, as employees must manually bridge gaps between disconnected agents. Instead of simplifying operations, these disconnected tools force employees to coordinate work manually, creating bottlenecks and extending resolution times.
The issue extends beyond agent coordination. True operational efficiency demands seamless integration with existing systems and data flows. Disparate data structures and integration gaps across agents create technical debt that stifles digital transformation.
Integration and Interoperability: The Missing Pieces
For AI agents to deliver on their promises, integration and interoperability must take center stage. Businesses need agents that access data across systems, trigger actions, and orchestrate complex workflows. Proprietary ecosystems and incompatible platforms exacerbate the problem, locking organizations into rigid, siloed systems.
A successful approach bridges these gaps. Enterprises need tools that foster collaboration across teams and technology stacks, breaking down silos and enabling seamless process execution.
Why a Platform Approach is Critical
A piecemeal approach to AI agent deployment undermines transformation goals. True platform solutions like Krista go beyond bundled tools—they provide core orchestration capabilities, predictive analytics, and openness to third-party integrations. These platforms enable businesses to adopt new technologies incrementally while preserving agility and avoiding vendor lock-in.
Consider customer support again: a platform approach automates the entire lifecycle, from triage to resolution, across all support tiers. This orchestration reduces manual intervention, accelerates resolution times, and delivers measurable ROI.
The Future of AI Agents in Business
Unchecked proliferation of task-specific agents exacerbates operational silos and negates their benefits. Businesses that embrace a platform approach will achieve true transformation by prioritizing integration, scalability, and cross-functional collaboration.
As AI capabilities evolve, flexible platforms that support quick change, communication, and interoperability drive innovation. Organizations that approach AI and technology adoption as a core competency will gain a competitive edge in an increasingly AI-driven business landscape.
Krista: Agentic Platform
Krista’s agentic platform eliminates inefficiencies caused by disconnected AI tools. By integrating and automating complex workflows, Krista ensures that AI agents deliver measurable outcomes without creating silos. Designed for adaptability, the platform scales effortlessly to meet evolving enterprise needs. Krista empowers businesses to unify their AI strategy, optimize operations, and future-proof their technology investments.
Links and Resources
Speakers
Transcript
Scott King
Alright.
Well, hey everyone, this is Scott and Chris again at the Union Podcast. Thanks for joining us. Chris, how’s it going?
It’s going well. It’s weird winter weather in Texas. It’s sunny today but freezing next week.
Scott King
Hot, cold, hot, cold. A plumber’s nightmare.
I want to talk about AI agents today. Someone asked me if we could create a list of AI agents, comparing them to others. I found this website, aiagentsdirectory.com, and was floored by the number of agents. There are 804 agents across 45 categories. I don’t know what it looked like a month or two ago or what it’ll look like in the coming months, but it will only grow.
How would someone evaluate and choose an agent? For example, the categories include platforms, frameworks, productivity, voice, and customer service. There are 47 customer service agents. How do you pick one?
Chris Kraus
When we predicted two years ago that there would be a proliferation of AI technologies, it happened. That prediction came true. These agents are grouped by tasks or specific parts of an organization.
For example, someone coding is looking for agents integrated into their coding tools. Someone in order management may look for agents focused on sales orders or customer service. These tools are very task-specific. Some agents naturally emerge from platforms, like those assisting with writing code. That makes sense.
But it’s a burden if an employee in a company has to interact with multiple generic agents. They’re expected to know where to look, which is overwhelming. It’s like trying to find the right SharePoint site. You might document it, but now there are more icons or web pages to navigate.
Scott King
Exactly.
Chris Kraus
The human still has to know what to do and find the right one, which often depends on their technical expertise or subject matter knowledge.
It’s impractical to have separate agents for level one, level two, and level three support teams. That’s chaos.
We need to think through how to deploy agent technology effectively. Are these agents handling atomic tasks, or are we optimizing outcomes for the business?
Scott King
People will purchase or subscribe to agents to help them with specific tasks. For instance, ChatGPT helps people write content.
Okay, that’s great, but that’s all it does, right? It helps you create something out of something else, whether it’s your own data, a corpus, or internet data. It’s helpful, but in the grand scheme, I think this reflects how people are used to using their own AI tools.
They don’t take a higher-level view to see the process across different teams. Some procurement guy, or the CFO—the ultimate budget holder—is going to look at all these subscriptions and contracts for these agents and realize, “Wait a minute, what are we doing?” And then he’ll notice that none of these tools work together.
Chris Kraus
Right, for sure.
Scott King
The CIO knows this agent does one thing but doesn’t integrate with another, and that agent doesn’t talk to a third. It’s like John’s prediction—he said we’d see more AI agents layered on top of and between apps. That’s clearly happening. There are 804 agents listed now, and next month there will probably be 1,000. So what?
Chris Kraus
For sure.
Scott King
Each of these agents will have varying levels of success. Some will be great, some will fail, and there will be consolidation. But as a business leader—say I’m a customer support VP running a call center—I could look at this list and find 12 agents I might need.
How do you even start? If you need 12 agents and have to evaluate three options for each, that’s 36 evaluations. You couldn’t do that in a year.
Chris Kraus
Exactly, and that’s not even a great approach because you’re thinking at a task level. Just because there’s a task-level agent for answering warranty questions or gathering customer information doesn’t mean you should use it.
Think about customer support. Everyone’s familiar with the experience of reporting, say, an internet issue. There’s always level one, level two, and level three support. Different subject matter experts handle different levels.
At level one, agents usually document and gather information. That first-level person often just writes things down, and you know you’re going to have to repeat it because they’ll either get it wrong or miss something in the script. Then it gets passed on.
Scott King
Most definitely.
Chris Kraus
Now, could that first-level person be replaced by an AI agent? Probably. They’re just following a script, gathering data, and escalating when needed. The customer knows they’ll end up repeating themselves anyway. So why not make that process smoother with AI?
Chris Kraus
For multi-hour or multi-day service level agreements, it escalates to level two, where they determine if the question is about a product warranty, an order or payment issue, or maybe a technical problem that requires an engineer to run a line test.
What often happens is we need to step back and realize we don’t need 30 agents just so level one, level two, and level three can move tasks faster. Instead, we should think about optimizing the overall process. If you’re the VP or head of a call center, your focus should be on improving the entire outcome.
Maybe the first-level agents use an AI tool to gather and verify all the necessary information. The AI could ensure that secondary information is collected and route the issue appropriately. For example, level two shouldn’t involve a person handling FAQ-type questions or accounts payable queries that need to be passed along. Instead, those should go directly to the right subject matter expert.
If it’s a line test, for instance, that’s usually just an API call to some outdated SCADA software. An agent could perform that step directly. The key is to remove unnecessary steps in the process, glue it together through conversational integration, and eliminate bottlenecks to improve efficiency. By doing so, the process becomes faster while ensuring subject matter experts at different levels can interoperate effectively.
Scott King
Yeah, but are you removing a step, or are you just automating it? It’s no longer a person’s task; it’s handled by an AI.
Chris Kraus
Exactly. We’re taking manual tasks and shifting them to software. The task still happens, but it’s no longer sitting in someone’s queue. By removing these people bottlenecks, the software operates at the speed of computers.
Scott King
I had to log in to AT&T the other day because, when I got back from out of town, the internet was spotty. It’s such a frustrating experience because the accounts are split—some bills are in my name, and others are in my wife’s name.
So I have to remember, okay, it’s her login and her credit card. It’s difficult for me to authenticate. They never seem to understand my questions, like, “Hey, I need a line test.” They’ll run one test, then move to another, and eventually, someone has to come out. That person ends up doing multiple things.
It took forever just for them to identify who I was. It was really frustrating. Honestly, if I could just jump into all the systems they’re logging into, I could probably solve the issue myself. Authenticating me is a nightmare.
As for the agents, you mentioned note-taking for level one. As we move toward systems where customers can log in and identify themselves, authentication becomes easier. But finding the right information is still a nightmare, especially with big companies that have multiple systems.
Imagine a company grown by acquisition, where they have two or three CRMs. You’d need multiple agents integrated into multiple systems. Ideally, one agent could integrate across everything, but most of the time, these agents are standalone. If one system gets acquired, fails, or changes, it creates a mess.
Chris Kraus
Exactly. You’re describing interoperability issues. It’s not just about having an agent give hard-coded or scripted responses. The agent should actually do something—look up your bill, request a line test, and deliver the results.
There are two parts to interoperability. One is physical: ensuring the software integrates across multiple systems within the business. The other is knowledge-based: each system involves a different subject matter expert.
If everything is done manually, you’re handshaking to a new expert for every domain. These experts have specialized knowledge, but if you assign a separate agent or task to each of them, it’s chaos. That lack of interoperability across the business process will bog everything down.
Something needs to orchestrate the bigger picture. Otherwise, you’re just creating queues of work for people or agents, without skipping unnecessary steps or optimizing workflows across subject matter experts. Interoperability is a tough problem to solve.
Scott King
Yeah, and I think a lot of people are realizing this. With many of these agents, if you can purchase the service with a credit card for $20 a month, you can get away with buying your own agent. But you’ll run into problems later with workflows.
Sure, you might automate part of your own job, but it doesn’t really help the business overall. You’re spending money for no real purpose. Eventually, someone is going to ask, “Why do we have all these agents, and nothing operates faster?” It might operate a little faster for one person, but that’s just because they figured out how to automate their job. That’s not a sustainable solution.
So, let’s stick with the VP of Customer Support example. Everyone can relate to having dozens of subscriptions on their iPhone, Apple ID, or Google Play. If you’re in a position like that, taking a higher-level view, what should you do? What’s a practical approach to buying and integrating multiple agents?
Chris Kraus
First, you’ll hear about classic agents built into software, and then there’s all this talk about “agentic platforms.” Honestly, a lot of that is marketing whitewashing because it’s still so new.
With Krista, for example, we purpose-built a platform to communicate with people.
Scott King
By the way, we’re calling it an agentic platform, so let’s not talk bad about agentic platforms.
Chris Kraus
AI.
If you look at the original purpose from four years ago, it was the orchestration of people, systems, and AI. Back then, generative AI wasn’t what it is now. AI was mostly categorizers, predictors, anomaly detectors—calculating numbers, yes/no decisions, or finding patterns.
We realized the problem to solve was enabling work to span multiple parts of the organization and multiple subject matter experts. The goal was to automate or remove some of those manual tasks. That’s still valid today, even as new AI technologies emerge.
Just because we have new tools like generative AI doesn’t mean it can’t fit into a larger orchestrated process. You need a platform that supports multiple types of AI. For example, answering questions about internal policies or warranties—what we call Q&A—can benefit from retrieval-augmented generation (RAG), which is hard to implement well.
We leverage different large language models (LLMs) like Claude, Gemini, or OpenAI, because each has unique strengths. Your platform needs to include robust capabilities—workflow interaction, system integrations, and core AI like RAG, predictors, and categorizers—but also be flexible enough to integrate with third-party tools.
LLM capabilities evolve constantly. A platform that locks you into a single AI technology is a problem. If Claude gets better next year, you’ll want to use it. If Gemini becomes cheaper and delivers similar results, you’ll want to switch. The platform must allow for this kind of adaptability, integrating new features as AI evolves.
Scott King
The innovation happens so fast. Can you imagine swapping one agent for another? You use it for a month and then find a better one that uses a different LLM. It would be a nightmare.
Chris Kraus
Exactly. You’re not going to keep switching agents manually. That’s why you need plug-and-play compatibility with different AI technologies, alongside core platform capabilities.
Think about two years ago when we first started discussing generative AI. It could write sentences, but it hallucinated a lot. Back then, the focus was just getting coherent sentences and paragraphs. Now, we’re talking about much more advanced capabilities.
For instance, reasoning is a new frontier. An LLM can now determine something like the total cost of insurance by reasoning that it’s your monthly premium plus your annual deductible. That kind of capability is a huge leap from where we were two years ago.
The big question is: Are these new advancements just interesting, or are they production-ready? Can we apply something like reasoning to solve real problems effectively?
I guarantee that before the end of the year, we’ll see some new enhancement—some breakthrough feature. The challenge will always be figuring out how to apply it to solve meaningful problems.
Scott King
All right. I can think of a lot of different podcasts we’ve covered, and just from listening to you, I’m reminded of those topics. Titles like “Your AI Must Do Something” come to mind. We’ve talked about RAG, integrations, and more.
But if you’re not on top of everything—security, privacy, data—it can all spiral out of control. Chasing agents for every little task will only make your life harder. The platform approach, by default, solves many of these issues. That’s why we advocate for it.
And, you know, John mentioned the other day how the level of compute we’re using now is drastically different from what was available just two or three years ago. Back then, it was cost-prohibitive. Now, with faster chips and lower compute costs, innovation is accelerating. The number of agents will only keep growing as compute gets cheaper.
Chris Kraus
Absolutely. As you said, these agents can’t be disconnected. They can’t just be isolated tasks. They need interoperability.
A platform approach emphasizes security, scalability, and multi-step processes. The real value of a platform is that it allows you to step back and see the bigger picture: a multi-step process with multiple subject matter experts working together to solve problems.
Sometimes that means removing unnecessary steps so AI can handle tasks on your behalf. Other times, it’s about realizing new solutions that didn’t exist before. For instance, instead of searching through a manual with “Ctrl + F,” you can ask a question and get a solid answer immediately. We’re fundamentally changing the way work gets done by applying better technology.
Scott King
Great, Chris. I think we’ve covered this topic thoroughly. If you’re exploring agents, please compare them to a platform approach like Krista. It allows you to build autonomous workflows that make work faster and handle the hard parts for you.
Chris Kraus
Mm-hmm.
Scott King
Go ahead, evaluate agents, and let us know what you learn.
All right, thanks so much for joining us. Until next time.