Learning Library

← Back to Library

Beyond Chatbots: The AI Agent Spectrum

Key Points

  • The prevailing “Can I use an AI agent for this?” question is misguided because most tasks don’t actually require a full‑blown autonomous agent.
  • AI solutions exist on a spectrum—from basic chat advice to fully autonomous agents—and we need a vocabulary to describe the intermediate steps.
  • The speaker outlines six levels of AI assistance, starting with the “adviser” (simple prompt‑based advice) and moving through increasingly interactive stages before reaching a true agent.
  • Level 2, the “co‑pilot,” exemplifies the middle ground where AI offers real‑time suggestions (e.g., GitHub Copilot for coding, interview support tools) while the human remains in control.
  • Understanding this spectrum helps organizations avoid over‑investing in complex agents and choose the cheapest, most effective AI approach for their specific problem.

Full Transcript

# Beyond Chatbots: The AI Agent Spectrum **Source:** [https://www.youtube.com/watch?v=obqjIoKaqdM](https://www.youtube.com/watch?v=obqjIoKaqdM) **Duration:** 00:14:38 ## Summary - The prevailing “Can I use an AI agent for this?” question is misguided because most tasks don’t actually require a full‑blown autonomous agent. - AI solutions exist on a spectrum—from basic chat advice to fully autonomous agents—and we need a vocabulary to describe the intermediate steps. - The speaker outlines six levels of AI assistance, starting with the “adviser” (simple prompt‑based advice) and moving through increasingly interactive stages before reaching a true agent. - Level 2, the “co‑pilot,” exemplifies the middle ground where AI offers real‑time suggestions (e.g., GitHub Copilot for coding, interview support tools) while the human remains in control. - Understanding this spectrum helps organizations avoid over‑investing in complex agents and choose the cheapest, most effective AI approach for their specific problem. ## Sections - [00:00:00](https://www.youtube.com/watch?v=obqjIoKaqdM&t=0s) **Beyond Binary: AI Assistant Spectrum** - The speaker outlines a six‑level spectrum of AI assistance—from basic chatbots to full agents—to help viewers recognize intermediate, cost‑effective solutions and determine when an agent is truly needed. - [00:04:23](https://www.youtube.com/watch?v=obqjIoKaqdM&t=263s) **Tool‑Augmented AI Assistants** - The speaker argues that most teams need a LLM‑based, tool‑augmented assistant—not a fully autonomous agent—highlighting its cheap, easy deployment, multi‑workflow benefits, and the emerging trend of startups acting as plug‑in tools within this framework. - [00:08:06](https://www.youtube.com/watch?v=obqjIoKaqdM&t=486s) **Semi-Autonomous AI in Customer Success** - The speaker outlines how a semi‑autonomous AI system can autonomously resolve most routine customer‑success cases, reserving human intervention for exceptions, thereby creating a scalable, efficient workflow. - [00:12:38](https://www.youtube.com/watch?v=obqjIoKaqdM&t=758s) **Evaluating AI Task Automation Levels** - The speaker urges shifting from debating AI agents to systematically assessing each repeatable task’s frequency, consistency, error impact, data location, and speed requirements to determine its appropriate automation level, recommending a level‑three tool as a practical starting point and offering a reusable diagnostic prompt. ## Full Transcript
0:00I am so tired of the AI agent discourse. 0:03Everyone is asking, "Can I use an AI 0:05agent for this?" I'm getting people 0:06asking me, "Can I use an AI agent for 0:08this?" And that is the wrong question to 0:10ask about AI. You probably don't need an 0:14agent for most of the things that you 0:16think you do. And I want to spend this 0:18video talking about the forgotten 0:21spectrum. Okay? We talk about chat, 0:23chat, GPT, and we talk about agents. We 0:26don't talk about anything in the middle. 0:28But the proper way to think about this 0:30is that if your problems are on a 0:32spectrum, the solution space in AI is 0:36also a spectrum. It is not binary, but 0:39we mostly don't have a vocabulary for 0:41it. And so I'm going to spend time in 0:44this video walking you through the guide 0:47to agents 101. not how to build an 0:49agent, but the guide to progressing to 0:52agents. The guide to understanding when 0:55you need an agent, the guide to 0:56understanding all the steps in between 0:58before you get to an agent that are 1:00cheaper and easier to implement for you. 1:02So, let's get into I've put together six 1:04different levels of actual spectrum AI 1:08assistance here from the very basic 1:10chatbot all the way through to agent. I 1:12have had experience working on projects 1:15for all of them. I'm dividing it into 1:17six because it's easy to remember. We've 1:19got handy little labels for each of them 1:21so you can remember them. And I'm going 1:22to give you real examples for each of 1:24them. The goal here is for you to come 1:27out with a sense of how to shape a 1:31solution with AI to a problem. So you 1:34are not overinvesting. So when your CEO 1:37comes and says, you know, I was reading 1:39on LinkedIn, this guy Nate was talking 1:41about agents. We should do agents. No, 1:45no. You should actually think through 1:46your problem space. You can share this 1:48video with him if he gets if he gets too 1:50excited. So, level one, what is level 1:52one? It's what you're doing already. You 1:54ask AI for advice. You do the work. This 1:57is how the vast majority of people use 1:59Chad GPT. For most people, this is the 2:01free tier. For some people, this is the 2:0320 buck a month tier. For a few people, 2:06they're doing the 20 buck a month or 2:07equivalent on Claude. You can do that 2:09right now, right? And most people think 2:11of this as the most basic version. I'm 2:14not going to spend a lot of time here 2:15because you already get it. We're 2:16calling this the adviser, right? 2:18Basically, the LLM gives you advice. The 2:20value of that advice depends entirely on 2:22your prompt. I've talked a ton about 2:24prompting. We're going to move on. Level 2:26two. This is where we get into new 2:28territory. We don't have words for these 2:30in between levels. And so, we're going 2:32to cover four in between levels before 2:34we get to the fully autonomous agent 2:36stage. Level two, co-pilot. Pay 2:39attention here. AI will suggest as you 2:42do the work. So, GitHub Copilot can 2:44write code while you type. Cluey is 2:47going to give you answers as you 2:49interview. Someone called it the Cluey 2:51stare where you kind of stare off into 2:53space for 3 seconds and then give a 2:54perfect five paragraph answer. So, that 2:57is what the co-pilot stage is. It's 3:00becoming a really common pattern. And 3:03here's what it's good for. It's good for 3:05repetitive tasks that have known 3:07patterns. It's like tab complete in 3:09cursor kind of thing. It can get you 3:11going 40 or 50% faster if the patterns 3:15are super repetitive and you know what 3:17they are. And so that might be good 3:20enough, right? The co-pilot piece where 3:22you know what you you you have enough 3:23repeated patterns and you don't really 3:25need to do anything more than that. You 3:27just need something to help you kind of 3:28find an extra gear in your own 3:30productivity. Great. That's what that 3:32co-pilot level is for. You are still the 3:35one driving. you are still the one in 3:36GitHub copilot that is framing up the 3:38coding cursor if you're hitting tab 3:40complete you're starting the line right 3:41level three it gets more interesting 3:44this is a tool augmented assistant I do 3:47a lot of work teaching people about 3:50level three because I think it is one of 3:52the jumps that is like most significant 3:55if you look at the relative pop value of 4:00oh my gosh that people get so think of 4:02it as like a curve of value and you 4:04you're wondering like going from level 4:06one to two to three, like what kind of 4:08value pop you get. The jump from 4:10co-pilot to tool augmented assistant is 4:12absolutely massive because it is 4:15multiplied by the number of tools that 4:18your chat assistant can get access to. 4:21And so there's almost no end to the 4:23value you can get here. I find a lot of 4:25people think they're at this level one 4:27chat advisor thing and they think they 4:29want agents, but they don't. They 4:31actually just want a tool augmented 4:32assistant. And when they get one, 4:34they're like, "Oh, wow. I had no idea it 4:36could do this." Right? It can use Excel. 4:38I had no idea. Right? And so this is for 4:41a chat that can access data. It can 4:43search the web. It can run calculations. 4:45It can build assets. It can edit assets. 4:48You can save your team dozens of hours a 4:52week properly using these. And I'm just 4:54going to say this is 10 times, 100 4:57times, maybe a thousand times easier 4:59than an enterprise agentic system to 5:01install. It's so much cheaper, it's not 5:03even funny. But people sleep on it 5:06because it's not an agent. Well, it is 5:08an agent. It's an LLM plus tools plus 5:10the guidance you give. But people expect 5:12agents to be like, you know, this 5:14completely autonomous Borg like thing 5:16that goes and and uses tools. Most of 5:19the time, what most individuals and 5:21teams actually benefit from is this like 5:24level three tool augmented assistant. if 5:26they could properly implement a tool 5:29augmented assistant for finance 5:30workflows, for marketing workflows, for 5:32product workflows, they'd go so far. And 5:34you know what's interesting is 5:36increasingly entire startups are 5:38becoming tools inside this framework. 5:40You can call an MCP server that has chat 5:43PRD as a product person. It will just be 5:45there. It becomes part of a tool 5:47augmented assistant in cursor. Super 5:49easy. And the way we tend to think about 5:52tools is limited when you have a world 5:55where anything can be a tool. An LLM can 5:58be a tool itself. You can have an LLM 6:00call another AI. And so this is a very 6:02powerful level. It gets you a lot of 6:04bang for the buck, but it's not the last 6:06one. If you find that you have types of 6:10problems that are beyond the repetitive 6:11task, so beyond the the co-pilot task, 6:13they're not susceptible to just calling 6:15tools, which is really that level three. 6:18you need more structure. That's when you 6:20get into structured workflow. That's 6:21when things start to get serious. Often 6:24times in these cases, AI will do a step, 6:26the human will review, AI will continue. 6:29This is choreographed work. And so in 6:32this case, uh the example I have JP 6:34Morgan wrote up a case study on this. 6:36It's a contract system. It saves an 6:37absurd number of hours a year, but 6:39that's really a function of their scale. 6:41People often look at these big numbers. 6:42I think for JP Morgan, it's a third of a 6:44million hours saved. Okay, great. But 6:46like that's a function of them being a 6:47big company. It's not really the AI 6:49there. The AI savings comes from good 6:52design around the problem space. And in 6:54this case, they recognized that what 6:56they needed was not the ability to call 6:58tools per se. They needed the ability to 7:00structure a workflow because in 7:02contracts review, it's got to be the 7:04same thing. You have high liability. 7:05It's got to be exactly correct. And so 7:07the AI can do a step, but the human 7:08needs to review. There are a lot of 7:11businessback operations that fall into 7:13level four. And no, we're not done, 7:15right? Like there's still more autonomy 7:17that we can get to here. And people 7:19again sleep on this piece because they 7:21think, well, the goal should be AI 7:22should do it all. Well, not necessarily. 7:25Like if you're saving a third of a 7:26million hours a year, I'd say you're 7:27doing pretty good, right? If you're 7:29saving a ton of time and your humans are 7:31able to do the work and touch the work 7:33in the right ways. This comes back to 7:35something I've been emphasizing that we 7:36forget in the agentic AI revolution, 7:39which is that your goal when you are 7:41designing AI systems at work should be 7:43for your best humans to touch the work 7:46more, not less. Your best humans should 7:48be more fingertippy, more hands-on with 7:51the work. They should not be feeling 7:53more disconnected. And this is a way to 7:55do that. There are lots of other ways to 7:57do it, but I don't want you to forget 7:58that principle. And I think we sometimes 8:00do with AI agents. We sometimes think, 8:03well, we're just going to sit back. 8:04We're going to get the pina coladas out 8:06and we're just going to No, no, we're 8:07not. We remain engaged with the work. We 8:10remain focused on making our best impact 8:13as a human while AI slides in around us 8:16as like a mech suit. And there's a bunch 8:18of ways to do that as we're discovering. 8:20Let's move to level five. Let's say you 8:23don't need the structured workflow. You 8:24actually need some degree of autonomy 8:27where the human isn't reviewing. We call 8:29this semi-autonomous. Surprise, 8:30surprise. The AI will handle routine 8:33cases independently. Humans would review 8:35exceptions and edge cases. Super popular 8:38in customer success. You can find lots 8:40of examples of this going wrong and 8:42going right. But by and large, the nice 8:45thing about customer success is each 8:47individual case fits within a spectrum 8:50of customer utterance or customer 8:52frustration. And you can really cleanly 8:54map it at scale to well, the AI can take 8:57care of 98% of cases with these 15 8:59workflows. And we can sort of use an 9:02engineering team and build that out and 9:03then the remaining 2% the humans will 9:05do, right? It becomes super clean at 9:07scale because of the way human 9:09complaints about products map onto 9:12typically a fairly normal distribution. 9:14And so semi-autonomous systems are a 9:16good fit here because you basically give 9:17the AI the ability to solve the problem, 9:20the tools to solve the problem, the 9:21workflow guidance to solve the problem. 9:23You build an agentic pipeline where it 9:24can read and respond and you're in 9:26business. And now we're getting to the 9:28world where people are starting to think 9:29of this as a real agent, quote unquote. 9:31I hate that phrase, but people do talk 9:32about it as I want a real agent, right? 9:34I've had that. No, you don't. You want 9:36real answers to business problems. And 9:38if you don't, you're probably not asking 9:40the right question and you're probably 9:41not going to make it. You need to be 9:43asking, "How can I solve this the right 9:45way? How can I put my best humans more 9:47in touch with the work? And oh, by the 9:48way, is AI the right tool for the job?" 9:50And there are so many ways that AI is 9:52the right tool for the job. And we are 9:54sleeping on that. We just think it's 9:56either the chat or the agent. Now, 9:58finally, we get to level six, fully 10:00autonomous. A AI does everything. This 10:02is what people think AI agents are, 10:04right? The humans monitor the metrics. 10:07Okay, you can do that. It does work. 10:10Some production systems are out there 10:12that do that really well. But the deal 10:14is this. You only need to do that if you 10:17have compelling reasons why human 10:20touches aren't relevant here, right? 10:23like why you need to automate the entire 10:25thing. One of the classic examples that 10:27fast food continues to sort of go after 10:30is the drive-through window, right? In 10:32that case, you either are paying someone 10:34to be at the drive-through window or you 10:36are not. It is binary from a labor 10:38perspective. And so, you really need the 10:40AI to take over the entire thing. And we 10:43have a number of cases where systems 10:46have tried to do that and that hasn't 10:48worked out. McDonald's has a case. I 10:49think Taco Bell has a case. It is a 10:51tough problem. And that is something 10:53that you should think about when you get 10:54to level six. Fully autonomous is a hard 10:57problem and the last 2 or 3% of those 11:00edge cases is extremely difficult and 11:02takes a lot of investment to get over. 11:04And so if your goal is fully autonomous 11:07for everything, you should be thinking 11:09actively about what your definition of 11:11the full scope of the problem is because 11:13if you can possibly scope it down, an 11:16example would be we will handle almost 11:18all of our customer queries on our shoe 11:21site, but for 2 or 3% of them, we're 11:23going to actually degrade it and say, 11:24"Thanks, this is a ticket so a human can 11:26look at it." That's almost fully 11:28autonomous, but it's not quite. And it's 11:31this sort of that in between level five, 11:33level six where the AI decides to chat 11:34and manages a conversation but then 11:36sometimes goes to a human. The fully 11:38autonomous bar is really hard. Amazon 11:40tried to conquer it when they tried to 11:42do uh selfch checkout in their little 11:45stores where you just pick it up and 11:46walk out. It turned out they never got 11:48there and that was with Amazon 11:50resources. I'm not saying that it's 11:51impossible to get there. I'm calling out 11:53that this is a spectrum and as a good 11:55system designer, a good solution 11:57designer, as someone who cares about how 11:59AI is is implemented, you should be 12:02thinking, do we have to go all the way 12:03to fully autonomous or can we design 12:05something that is going to give us 12:07almost all of the value if it doesn't 12:09take that much investment. Another 12:11example of how complex fully autonomous 12:13is, we cannot roll out Whimo cars, 12:15self-driving cars to every city like a 12:18rubber stamp effort. Even though we have 12:20them and they are fully autonomous in a 12:21few cities, they have to relearn every 12:24single city. Despite all of the training 12:27on roads, we have to teach them the new 12:30city map in detail. That's where we're 12:32at right now. Fully autonomous is really 12:34hard. So, here's what I want you to do. 12:36I want you to stop asking, should we 12:38build agents? And I want you to start 12:40asking, what level does this specific 12:43task need to be at? Think about a task 12:46that you do repeatedly. How many times 12:48is it done per month? How consistent is 12:50it? What happens if there's an error? 12:52Where does the data live? How fast does 12:54it need to happen? These questions are 12:56not random. They're actually the 12:58questions you need to answer to give you 12:59a sense of your level. And so, look, if 13:02if you're not sure, I'm just going to 13:04tell you, I talked about this level 13:06three tool enabled chat assistant for a 13:09reason. Most people end up at level 13:11three for a lot of things. There's a lot 13:14of other options on that spectrum. I 13:15gave you a lot of range, but level three 13:18is where a lot of people hang out. And 13:20the thing is, if you are unsure, pick a 13:22level you can try yourself that you 13:24don't need stakeholder approval for and 13:26see if it makes your workflow better. 13:28See if you feel more empowered. Now, if 13:30you want to go through the whole thing, 13:31I absolutely did 100% build a prompt for 13:34this. I built a prompt to help you 13:37diagnose where you are at for given 13:39tasks. So, this is not designed to be a 13:41oneanddone prompt. This is designed to 13:43be a prompt you save and then you run 13:45when you have a task where you're like, 13:47is it an AI task? If it's an AI task, 13:49what level is it at? I need some help. 13:51You kind of get my sort of innate brain 13:52on that. If that's something you're 13:54interested in, you know where to find 13:55it. And if it's not, please, for the 13:58love of all of AI and all of your own 14:01work, do not assume that it's just chat 14:04and just agents. Please think about it 14:06as a spectrum. I hope this video has 14:08helped you to see that. share this with 14:10someone who needs to see the problem 14:12more widely because I think so much of 14:15the bad use cases we see in AI, the doom 14:18stories, the terrible implementations 14:20come down to people not understanding 14:21this that really it's not a light 14:24switch. It's a spectrum of AI 14:26implementations and you can talk about 14:29them. You can develop the vocabulary for 14:31them and you can make better choices. 14:33So, here's to better AI implementations 14:35that don't suck.