AI Agents Are on the Horizon. What Should You Buy (and Avoid) Right Now?
A field guide for leaders who want real value from AI now—and want to stay ready for the next wave of agents and automation.
Hey everyone — Before we dive in, I want to say thanks. The response to last week's newsletter blew me away—your comments and the conversations that spun out of it have been amazing. It's clear this community is full of sharp thinkers and authentic operators, and I'm grateful to be building this with you.
Frontier Notes just cracked the top 100 on Substack's technology rising leaderboard (#90!) after only two months. None of this happens without your support, engagement, and word of mouth. 🙏
Are AI Agents Just Expensive Vapor?
You're going to hear a lot about "AI agents" this year.
Vendors will pitch you dreams of cutting 50% of your workload. Peers will claim their businesses have agents running everything but the coffee machine. I've seen these demos. They're mostly fantasy dressed up in a slick interface.
Let's be honest: Everyone's using some form of AI, but most businesses aren’t running truly agentic systems at scale yet in mid-2025.
What people call "agents" often turns out to be basic chatbots wearing fancy hats, copilots that still need your hands on the wheel, or workflow hacks that require more babysitting than my neighbor's toddler. The tech is evolving fast. The definitions are blurry. Most organizations haven't even fully digested the fundamentals.
Brutal.
Here's the reality check you won't hear at conferences: If you haven't integrated the core models and gotten your team comfortable with manual prompting and basic workflows, you're not behind on agents. You're behind on the basics.
The "agent" wave is absolutely coming, and yes, competitive teams will move quickly as the tech matures. But you're not late to the party unless you're ignoring the main models and the foundational adoption curve altogether.
Before we dig into what to implement and what to avoid, let me cut through the fog about where things stand and what "agentic" tech really means.
The Truth About Today's "AI Agents"
Don't believe the hype. Most businesses are not deploying true AI agents today. Not even close.
These tools dominate headlines and investor pitch decks, but practical, wide-release agent tech is still rare. Especially outside the U.S. and for smaller organizations. Even OpenAI's "Operator" remains U.S.-only beta with external access more exclusive than a secret underground club in Tokyo. I had to pull strings just to get a demo.
We're standing on the edge of a wave, not surfing in the middle of it. Competitive businesses will adopt quickly as agentic tech matures, and being prepared now is how you avoid getting left behind later. The real value today isn't in chasing vendor promises. It's in understanding what's possible versus what's still digital vapor.
You need a mental model for this space, so let me break down four key categories of AI tools that everyone confuses:
1. Raw Models (The Engine)
Examples: OpenAI, Anthropic, Microsoft, Google, open-source LLMs.
Think of these as the industrial engines powering everything else. They're usually accessed through APIs, built for technical teams who want full control over every lever and dial. They offer high flexibility but aren't for most SMBs since they require engineering resources and a tolerance for complexity. Tech teams love them; but marketing teams might rather stick forks in their eyes than work with raw APIs.
2. Official Interfaces (The Workhorse)
Examples: ChatGPT, Claude.ai, Gemini, Grok.
These are your Honda Civics of AI. Reliable, user-friendly interfaces designed for non-technical users and teams. This is how most people actually "use AI" today. Great for business experimentation and daily productivity, but bound by the vendor's rules and limitations.
Right now we’re seeing companies try to build entire workflows on these interfaces, which is like trying to run a construction company with only consumer-grade tools from Home Depot. Possible, and practical, but might not be ideal as you scale.
3. Wrappers & "Agents" (The Middle Layer, Often Overhyped)
Examples: Perplexity, Notion AI, browser agents, niche startups.
This is where the marketing gets ahead of reality. Most third-party tools are wrappers: they combine the raw models with their own interface or workflow. Many now brand themselves as "agents." Usually just automated workflows with limited autonomy, like a restaurant hostess who can seat you but can't take your order or bring your food.
True agents take action with minimal prompting, automate tasks, and begin to act independently. They're like having an invisible assistant who joins your meetings, takes notes, distributes action items, and follows up on deadlines. All without you having to say "hey, could you..."
These are still rare in the wild, despite what your LinkedIn feed suggests.
4. Enterprise Solutions (The Heavyweights)
Examples: Salesforce Einstein, Microsoft Copilot, vertical-specific platforms.
The corporate battleships. Slow to turn but hard to sink. These offer deep integration, security, and scale, but they're expensive and risk lock-in. Best suited to operations-driven teams needing stability and compliance. For most SMBs, these are overkill or simply financially out of reach, like buying a combine harvester for your backyard garden.
Where most people get confused:
Many tools blur these lines faster than a watercolor painting. Wrappers claim to be agents, enterprise solutions rebrand as "AI platforms," and the raw model leaderboard changes weekly like a game of musical chairs. One week Google's Gemini leads, the next week OpenAI drops a new model, then Claude leapfrogs them both. Don't get distracted by the horse race. It's like following sports standings when you should be practicing your own game.
What's actually agentic right now? Practical examples include meeting bots that autonomously join calls, transcribe, summarize, and distribute action points through email to participants without constant human prompting.
Consumer agents like OpenAI’s Operator can handle basic tasks like booking and scheduling, but they're not fully autonomous or widely accessible yet. Plus, most people are still afraid to ask agents to book a haircut without it trying to order a pizza at the same location.
True agentic behavior means tools that act, decide, and report back without any prompting or micromanagement, and they're still unicorns in today's market [in mid-2025].
The takeaway for SMBs and mid-sized teams: Stick to the official interfaces or well-vetted wrappers for now. Pick a platform that fits your workflows, and don't get FOMO over "agents" that don't yet deliver.
Start monitoring agentic developments with an eye for what will actually move the needle for your operations, and be ready to move when the tech gets real. The revolution is coming, but it isn't quite here yet.
Smart Selection: Tools That Work Now and Won't Box You In Later
Stick with the Big Platforms—But Keep Your Options Open
ChatGPT, Claude, Gemini, and Grok are the safest, most capable user interfaces today. I've kicked the tires on all of them, and they each have their strengths. No one model or interface "wins" for long in this field. It's like trying to crown a permanent champion in a sport where the rules change every quarter.
Pick an ecosystem that fits your workflow and budget, knowing each will push agentic features as soon as they're viable. Trust me, you’ll know when these agents are available—they want your money.
If you have the resources, experiment with more than one platform, but avoid scattering your data and processes across too many systems like I did earlier this year. I ended up with notes in five different AI interfaces and couldn't find anything when I needed it.
Lesson learned.
Don't Buy Hype—Buy Fit for Your Real Problem
Identify the most repetitive, high-friction tasks in your team's day. For a finance team, it might be processing receipts and expense reports. A mind-numbing time sink that everyone hates. For sales, it might be meeting notes and follow-ups. For support, it might be categorizing tickets.
Look for tools or wrappers that automate those specific pain points. Don't chase "agents" for their own sake. That's like buying a Ferrari to drive to the grocery store. Most "agent" startups are immature or unproven. Use with caution and demand evidence before investing. You don’t want to get burned by three different "AI agent" startups that promise the world and deliver a paper map.
Spotting the Red Flags in Today's 'Agent' Products
Before you adopt any new tools or systems, let’s talk about where things go wrong—and how to dodge the most expensive mistakes.
🔴 Marketing Hype ≠ True Agency
Most "AI agents" marketed today are just chatbots or copilots with a new label. True agents act autonomously and execute multi-step workflows without constant hand-holding.
Ask yourself: Can this tool take actions, adapt, and achieve goals with minimal human nudging, or is it just a fancy prompt engine wearing a bow tie? Most production "agents" in mid 2025 are still far from the real thing, despite the marketing fairy tales.
🔴 Security Risks and Custom-Build Pitfalls
Security and Compliance: Agents with access to core business systems are a security risk if not handled properly. It's like giving an unknown contractor keys to your office, server room, and financial records. If security isn't front and center in their pitch, walk away. Gartner predicts agent abuse will drive a significant chunk of data breaches by 2028.
Custom Builds and Complexity: Building your own agents is a money and time pit for most SMBs. Romantic in theory, painful in practice. Most early-stage agent failures we’ll see in 2025 will come from biting off too much. Teams underestimate the complexity, the edge cases multiply like rabbits, and suddenly your "six-week project" is in month eight with no end in sight.
🟠 No (or Weak) Integrations
If the tool you want to automate doesn't have open APIs or integration hooks, agents can't do their job. It's like hiring a personal chef but locking your kitchen door. Avoid products built on "walled gardens" or closed platforms; these are dead ends for automation.
Most SMBs should look for solutions built to plug-and-play with their existing stack, not ones that demand you rebuild your entire digital infrastructure just to accommodate them.
🟠 Bad Data = Bad Agents
Agents are only as good as the data you feed them. If your internal systems are messy, siloed, or lack governance, you're setting up for frustration or failure.
Don't expect magic from agents until your data is clean, accessible, and up to date. No AI can turn your digital landfill into gold. Not even the most expensive ones.
🟡 No Clear Goals = No ROI
Don't invest in agent tools just because "AI is hot right now" or your competitor mentioned them on LinkedIn. Focus on defined, repetitive workflows with clear pain points and measurable impact.
Start with a simple goal: reduce the time our support team spends categorizing tickets by 50%. That’s specific, measurable, and meaningful.
Start small, measure obsessively, and implement an agent service only when it can demonstrate real results. The best implementations build momentum through visible wins, not grand visions.
🟡 Lack of Transparency Breeds Mistrust
Prep your team thoroughly too: agents are productivity multipliers, not job killers. Frame them as taking the robotic parts out of human work, not taking human jobs. How you message this internally matters more than most leaders realize.
The agent revolution is a no-brainer. It's coming. But right now, we're in the messy middle, where buzzwords outpace reality and the most impressive demos often hide the most fragile implementations.
Avoid shiny objects and stay focused on tools with genuine autonomy, robust integrations, error handling, and a clear path to value. Clean up your data like you're expecting company, set measurable goals your CFO would respect, and look for vendors who understand SMB realities instead of enterprise fantasies.
This market moves faster than fashion trends. Be skeptical, start small, and keep humans in the loop. You don't need to be first in line for every new platform, but you do need to be ready when the truly transformative ones mature.
And if a vendor tells you their agent system is "fully autonomous" and "never makes mistakes"?
Show them the door.
In the AI world, honesty about limitations is the first sign of competence.
Scout Report
Here are the two signals I’m recommending this week—big picture trends and real talk on where AI and agents are actually headed.
🔗 Visual Timeline: AGI by 2027 → This visual timeline doesn’t pull any punches. It’s a fascinating, sometimes bleak, visual of where agents fit in and how fast things will accelerate. → Read more
🔗 Video: Are Your Job Titles on Borrowed Time? → If you like the Diary of a CEO, this is a great debate. Sure, it’s a little overhyped, but there’s real insight here about agents, automation, and the future of work. → Watch video
Heard on NotebookLM
This week, I'm sharing something straight from my workflow—a custom audio clip generated using NotebookLM.
If NotebookLM isn't on your radar yet, it should be. I'll go deeper on my research tool stack soon, showing how to build your own custom audio summaries using deep research without learning a complicated new workflow. That's coming in next week's newsletter.
For now, I've created an audio summary using all the research behind this week’s newsletter. If you're still fuzzy on what AI agents actually are (and why they're suddenly everywhere), this clip delivers the straight goods.
Forward Signal
I hope this gave you a sharper sense of where agentic tech actually stands. As you've seen, even the definition of "AI agent" is all over the map, and right now, implementation is complicated for most teams.
But here's what I see coming: agent engineering is about to become its own specialty. Over the next year, expect to see a wave of new roles—engineers dedicated to setting up, maintaining, and evolving these agentic systems.
The bottom line: most teams won't implement agents without technical talent, and that's a real opening—AI isn't just about automation, it's also creating new opportunities (and new job titles) for people who can bridge the gap.
That’s it, that’s all! Look for my upcoming piece on research workflows—it's shaping up to be my most actionable yet.
If this edition of Frontier Notes for AI Leaders gave you great insight, please share it with someone who leads through change. And if you haven’t subscribed or upgraded to premium yet, we’re delivering frameworks and systems that make AI practical every week—sign up below.