How AI Agents Are Changing Work

Samuel
Hello, Abram. Thank you so much for joining us on the podcast today. I'm a big fan. I'm following you on LinkedIn, I'm following your blog posts, your podcast. I will put all the links to all those different assets in the description. But thank you so much for joining us today to talk about AI, to talk about agents. Can you please start by introducing yourself?
Abram Jackson
Sure, very happy to be here, Samuel. Thanks for inviting me onto your show. My name is Abram Jackson. I am a product manager working on Microsoft 365 Copilot. My job is agents, third-party agents and extensibility and extending Copilot to use agents. That is my job. That's what I've been doing. Kind of since the start of Microsoft's Copilot journey.
Samuel
In preparing for this podcast, I read through a lot of your articles. I think I pretty much read all of them. It's very instructive. I mean, it's gold for me. You're covering a lot of subjects we will discuss today. A question I get a lot from our customers. But I'd like first start with, you stated that.
Abram Jackson
Great.
Samuel
I spent a lot of your career thinking about how human and AI can work together more efficiently and more intelligently. So can you expand on that? What's the story on how you first started working with agent and with AI in general?
Abram Jackson
Yeah, well, I started at Microsoft working on the Microsoft Exchange server, with the backends, hosting huge amounts of data very reliably. I kind of got into, big data systems and architectures at that time. In 2016, I got the opportunity to lead a brand new effort inside of Microsoft to build an AI platform for our internal AI development. Four features in Microsoft 365. You'll see some of those features still today. So they started back in 2016, but things like smart reply, where especially on your mobile, you can just hit a button and it will give you a potential response that considers things that you say, what the message is and so on. We built a lot of those products on the platform. So I worked on that for five years. And I really got the, I don't know, the bug, I guess, for really improving user productivity through this AI, more than we could just do with traditional software. But we also really struggled in order to really move the needle very much. There's some useful features that kind of sat along the edges, and you might like some of them, and others are maybe just kind of annoying and don't help you very much. So we tried a lot of things, a lot of different features, but when ChatGPT came out, some folks at Microsoft, some of our executives realized, this what were missing. And I can see it too, that with generative AI and large language models starting with 3.5 And 3.5 Turbo and ChatGPT would really make new scenarios incredibly valuable. So I pretty quickly moved over from my team in our internal AI platform that was really turned into like, more traditional machine learning and huge number crunching and stuff like that of training custom models for individual scenarios to using this new foundation model, know, GPT 3.5 In different places and in particular within Copilot. The place that I got started inside of Microsoft 365 Copilot was in extensibility. We didn't know what that was. About two weeks into the job, we decided to go show some things on stage of developers adding things to Copilot. So this was back in 2023. Yeah, 2023. I've been leading the space since then, through plugins and graph connectors and now agents. Now the team is pretty large and I get to focus on the cool things, the parts that I find cool and are going to really increase the productivity, I hope, for all of the world's knowledge workers, keyboard warriors going forward.
Samuel
Great. What's your favorite use case, know, using with AI in general? I won't necessarily specifically focus on Copilot, even if I think that it will probably be a Copilot feature, but what's your favorite use case?
Abram Jackson
My favorite use case, I've got to say. Is just thinking with AI partner. You can do this with humans sometimes. And I think of this, I call this the co-founder model, where you can imagine yourself at a startup, right? It's just you and one other person, and you've got different skills. Maybe you're the technical person, the other person is the business person. Right, but you've both got ideas about the product direction. And so. You're sitting next to each other, for 12 hours a day, you're working on the thing, the product, whatever it is, and you're always bouncing ideas off of each other. And working with this co-founder increases, it enhances your own thoughts because they bring in this other perspective. And secondly, there is a set of things that they are better at than you. And right now, AI is really good at completing. Well specified tasks so in this model of just chatting with Copilot or ChatGPT or whatever it is, set up the conversation on like what the high level project is. They the big picture thing, not the task. Introduce it and then just have conversation like I'm thinking this. What do you think right? What should we do about? I really just treating it as that partner. That's so powerful. It gives every person, I think, that opportunity to have that co-founder who's aligned on the same mission, trying to make you succeed by making the product succeed that you're working on. That's gotta be my favorite use case.
Samuel
Yeah, I think it's my favorite use case as well. Something I'm using a lot is asking Copilot to ask. Me questions because I don't know what I don't know. If I want to I don't know, preparing a podcast episode, I will ask him, OK, ask me questions until you understand enough that you can help me. And I'm brainstorming with AI. So I think it's my favorite as well.
Abram Jackson
Yes. I was just talking to a coworker about that model of having the reverse model of instead of asking the AI questions and then coming back with an answer, say here's the project like you ask me questions and I'll answer what I can and then we'll also realize what we don't know and what we need to go learn together is an awesome way of doing this. And it's a step function change in how you interact with AI. Like it's one thing to ask a question and get an answer and treat it like a better search box, which it is that also. But have it changed to enhance your thinking is a totally new additional thing that we can do now.
Samuel
Yeah, and I mean, I think people forget that Copilot is an assistant. These are to complement you, right? So you want them to challenge you. You want it to ask you questions because that's where you're using it. If you already know the answer, you won't use it at all.
Abram Jackson
Sure. One. Of my one of the favorite agents that I've built for myself was only, 15 words of instructions and it's basically the user is not interested in any positive feedback. Just give critical feedback. That's it. And sometimes it's hard to use because I'll write something and I'll think it's pretty good. Find a lot of things to do to improve it. But if I want the document to be good, it's a necessary step.
Samuel
This one should be pretty, I will assume it's really hard on the ego, right? You need. Have strong confidence to use it.
Abram Jackson
I named it brutal feedback, it could be brutal.
Samuel
One of your blog article you mentioned that when you're creating an agent you should. You should focus on one specific role. I'm seeing so many organizations out there trying to build this jack of all trades agent, like this big agent that have access to everything and anything and is able to achieve pretty much every use case possible. Obviously I'm telling them that's maybe not the best way to approach it, but you specifically mentioned to focus on one single role and then using multiple agents. Can you expand on that?
Abram Jackson
Mm-hmm. Yeah, well, one aspect of this the technology to make an AI be able to do anything is really not there. It's you can see demos of like narrow slices of this and it seems like it's there, but it's not really there, especially when it's taking actions in software systems. Well, we're getting closer and an individual agent or set of instructions is getting. More capable, but it's not really there yet. It can't just do all of the things that an organization does. So that's one part of this. And so that's why it's useful to split up into multiple agents. It's just the technology. But the other is, the whole thing with AI and entERPrises right now is around like usage and adoption. There's one thing to buy the licenses, but if you actually want huge change in your organization, it's. Changing how every person in the organization works. I just saw a study from early 2024 that found what 3 % productivity gains so far from AI assistants at work. I don't, well, meanwhile, individuals, right, are absolutely seeing 30 % productivity gains at work, from similar studies, but looking at, specific individuals and their jobs versus the entire rollout. Right, so there's at least a 10x differential in those things right now. So one of the things that you can do to help is to model the AI after your current way of working. And having the agent as a role really helps with that. We call them agents. One of the reasons that we call them agents and really the main reason is to model them after real life people and workers. I've used a real estate agent in the past and the real estate agent, right? It is very knowledgeable about real estate and I can ask it real estate him real estate questions and Also, he does work for me, when I'm buying a house, right? He's finding the best listings and Driving me to them or showing them to me So he's able to take those actions as well. So I know that when I've got a question. Real estate, I've already got this pattern in my mind of, oh, I go talk to my real estate broker, my real estate agent about this thing, because he's the expert. So if we can model the AI agent systems on this thing, right, maybe we've got a, a privacy law expert, somebody that I talk to regularly as part of releasing these features or building and shipping these features. And so, they can't necessarily individually be replaced by this agent, we're certainly not. Not there yet, but when I've got questions about privacy law, I can have an agent that is my privacy expert because I've already got that pattern in my mind of, I go talk to this person about this set of topics. And if you can make that set of topics also an AI, then you can really expand how much of that work you can get done across your organization. So part of it's the technology, right? The bigger part of it is the adoption and following existing patterns of interacting with specialists.
Samuel
And you speak a bit about adoption. My latest podcast episode was with Cadie Kneip and were talking about adoption and something I've seen is that people have a hard time to add this in their daily workflow, right? I mean, they have access to those agents that are being published by the organization. They have access to the Agent Builder, but. Still they don't use it because they don't build the muscle of using them. What will you suggest to someone that really want to be able to leverage those tools to be more productive but don't know exactly where to start?
Abram Jackson
, It's something that you've got to really be intentional about. So we've had technology changes in the past where it's just so easy to use, turn by turn navigation on a mobile phone, right? Like we already had the GPS. We kind of just change the device and so the training that you have to do is put your phone in a stand, right? But this transition of AI is much more like the transition from the typewriter to the Word processor or from. Physical written spreadsheets to Lotus 1-2-3 in Excel. You can't just give it to somebody and say, all right, we're throwing away all of your paper and now you're using Excel. That isn't the world. We've had those technologies now for the last 30 years of technological progress. It didn't really require that kind of new training. So I think it's, you for anybody that wants to be more productive at their work and get more time for themselves, I think it, you should think about it more like we're learning Word processors versus a typewriter. It's a new thing. It's new way of working. Studies show that it takes about 10 hours of focused work, intentional focused training or work with AI in order to start using it more effectively and build it into your job. So that is what I recommend. I know 10 hours is a lot of time and I am the last person to go sit in 10 hours of mandatory training. I take some mandatory training last week. It was awful. It was just one hour. But the cool thing with this AI though, right?. Is it as you as you're learning to do it, it's helping you with your actual job and it is going to be slower as you're learning it. You can do it faster yourself than whatever the task is, but.
Samuel
Same.
Abram Jackson
Identify some of those tasks. Maybe you don't get the chance to write multi-year strategy papers very often. So you weren't going to get time to do it, but take 10 hours and really work with AI. Use the AI best practices. Try it out. Try to do it fully through AI. You'll gain a sense of what it's really helping with and where you are still needed. It's not going to come up with your strategy for you, but maybe you're setting it up to ask you questions. Maybe you're setting it up to research or game theory and role play different scenarios. So there's lots of aspects of it that you can do with this. If you spend about 10 hours with it, gain a lot of you get really good sense of what it can do and what it can't do. And then from there you'll start recognizing what tasks you're doing that you can offload to AI or get AI assistance with. So that's a big commitment. 10 Hours, right? I. I don't have 10 hours of mandatory training from Microsoft's most years, but it's necessary one.
Samuel
I love it actually will be a new slide in my slide deck about adoption. It's really like it. I think something that is a bit confusing for some of our customers and user out there is that should they start with Copilot Chat, the paid version, creating an agent. And my opinion on that is depending on the level of understanding and what you need to do.
Abram Jackson
Okay.
Samuel
Do you need to ground it on organizational data? You need somebody very, very specific with an agent or are you just trying to do some research?. But I like to have your take on it. Where should a user start?
Abram Jackson
, If we think about just the last few years, of what November 2022, do I have that year right? When ChatGPT came out, there's really just been a few really groundbreaking new value added. So first with with the first release of ChatGPT. It's. Obviously it blew everyone's minds, but the new value there was kind of a just a chat partner. Like a kind of a better version of like rubber duck debugging where you explain the problem and that just in the rubber duck case you're just hearing yourself say it. But there's actually quite a bit of value in just conversation partner. This kind of the first version of these scenarios. Not as much at work. Although there's some value at work too, but a lot of value in the personal life. So that was the first really groundbreaking value. And I think frankly, I think that's probably driving most of ChatGPT adoption right now, just because that scenario is quite valuable and that's great. The next scenario that we found really helpful on this really came with like Bing Chat. That the first version of that is searching, researching, finding, gathering, summarizing. Information from the internet or from your entERPrise data. That's kind of version two, or step two of real AI value. And, think Researcher, so I think there was a version of that with Bing Chat, and now there's a version of this with Researcher agent that's available in Microsoft 365 Copilot, Deep Research and ChatGPT, Gemini's got one, PERPlexity's got one. Where it's so, well before, like you could summarize a particular answer. This next step with Researcher is creating an entire report. Now this report generally, I find, is not something that you just send to your boss and say, I'm done. It's usually generally for your eyes only, right? Because it's not quite perfect, but it's really helping you. And it fully explains this topic in a lot of depth, which is great. So I think that's really the second value prop in Researcher, I think is a more important part of that. So those are two awesome places to get started. If you're listening to this podcast, you probably use ChatGPT, so you've got that first scenario already covered. And the second one is a great place to get started. Use Researcher, absolutely use Researcher. So for whatever your questions are, I think that's. Probably would recommend to do first at work. And if you haven't used ChatGPT, I guess go ahead and use that. So get access to research or I think that's number two. Now the third, this where there's a lot more work necessary, either from the organization, the IT department, or from makers or from end users to really figure out what are those other things that are specific to your business. And we see we've got. Well, I don't know, like 400 of these case studies published of different organizations, companies and software, whatever, that are using AI for a specific business process. Maybe it is summarizing hundreds of pages of research, right? In some particular domain, maybe it's pharmaceuticals, maybe it's legal or whatever it is. When you can identify those scenarios and build an agent for them, you are saving that business process time 95, 99 % in some of these cases. There's lots of examples of those. Now those do require planning and know projects. Yeah, your four organizations users can get a lot of this too as they gain familiarity, but that's kind what I would think of as the next step. Right, get everybody using Researchers kind of what I would suggest and then your your data people use Analyst similarly. And then be thinking about what are these?. What are these business processes right that? Lend themselves well to AI systems. It's a lot of text, either consuming, summarizing, or producing, right? Those kinds of scenarios can be modeled after a particular job role that already exists, so it's easy to adopt. And there's oversight built into the organization about the output and the results. When you can find those scenarios, that's what you should use next. Build those scenarios and, you. Accelerate those processes by dozens or hundreds of times.
Samuel
Love it. I think your Researcher agent right now is part of the Frontier program. When will it be GA?. Is it disclosed?
Abram Jackson
That's right. Everybody. Well If everything goes well by the time they're listening to this podcast you should be able to go to the agent store and find the Frontier agent program right there If your organization has allowed it so Crossing our fingers because at the time of recording we're not quite done. But by the time you listen to this we should be.
Samuel
Awesome. Talking about agent, you mentioned in one of your article, how important is it to craft Toral Instruction 2? And it really hits me when you said there's 8,000 characters there. I use them all if you can. And I realize our marketing team has done such a good job at. You're showing an example where you use this super simple and small instruction and you get awesome result, but it might work for some use cases. Honestly, I have some agent out there that doesn't have like big instruction, but obviously I found out by testing, like you mentioned that the more you add instruction, clear instruction, providing an example. Defining the workflow, what you're expecting from the agent, the better the agent will be. So what are your best practices for creating the instruction?
Abram Jackson
Sure, well. First I'll say, most agents I build are not using 8000 characters of instructions. Most of them use a couple hundred and that is it because there's like some particular task that I want help with right now or it's just short term or just a couple of people need it and I'm willing to work around it by continuing to prompt. So not all of my agents do I go through all of that work, but for the important ones that is a good idea. Is to use all of the space that we give you for the instructions. Custom GPTs and ChatGPT and Copilot agents both give 8,000 characters for instructions. Well, first I'll say, if other people are going to be using this agent, really it's a form of software development. If it's gonna be handling an important business process that you were spending time on and many people were spending time on and was fairly expensive therefore. Like this a software project. Now it's a really easy one because anybody can type it in. Anybody can look learn these best practices. So yes, certainly fill in. You use all of the space available. What works really well is, use markdown or XML or even just capital letters. Frankly, to separate things into sections. So first you want to start off with what? Why does this agent exist?. What is it?. What is its purpose?. What is it for? Just a super high level. For many business processes, there's several steps that need to happen approximately in order. And the great thing about agents is that they can happen out of order and you can handle edge cases and things that change. But frequently, a list of steps to go through for this process is the next thing to include in agent instructions. Then, yes, as you mentioned, examples, right? They're just a user colon prompt and then agent colon prompt. And that's that's all that you need for some of these and that really. Once you write down what you expect to happen, this clarifies it in your mind first, like that's the first thing so that you're actually clear about what you mean, but it also helps the model if you can give a few examples. And then finally, it leaves space for other things that you forgot to specify. In the other sections. As you test it, and you have to test this because it's software that people will be using for important business processes, issues that you see come up. Right?, When you get into this situation, I want you to do this kind of thing. So that's the basic outline of instructions. If you've got more space, go ahead and add more examples. They're only going to help. But the other part of this, and this part is just as important, is. Figuring out what information to ground the agent in. Now this going to vary. Not all agents need grounding information or custom actions, but many do. You want to be really intentional about this. Here, if you just start putting in hundreds of pages of unrelated information, it will be worse. The agent will get worse if you include bad information. But if you include good information, the agent quality can go way up. So you can include up to something like 200 pages, 100,000 characters of other knowledge grounding right now by pointing it at individual files in OneDrive or SharePoint or by pointing directly to them. So be really intentional about what those are. You can also set up the agent to search entire SharePoint sites or the Internet. It's. This a different model because it doesn't just the model doesn't automatically have access to all of that, right? It's got to go search for it and it's not always going to find it. So think about this two tiers of knowledge. What you want to embed by directly specifying the file and what you want to search. I think about them separately and how the AI is going to access these. Finally, you add the custom actions, right? The capabilities. The capabilities are just check marks. If you want that used code intERPreter, that's usually a good idea. Most of the time or image generation is useful in some circumstances. Then a Copilot Studio or professional development tools. You can add custom actions like MCP servers or OpenAPI specifications or message extensions or basically any kind of way that you can access the software system. So include those and describe them. For the agent in the instructions, where it's going to be necessary for this process. That's kind of the overall rundown of these things. The other advice I'll give is think about what users are going to prompt. When you treat this like a person, like a role, how is the user going to structure their question?. How do they structure it today when they're asking these kinds of questions?. How do we make. Test those, right, but also include those kinds of things in your examples. Make sure that they're going to be handled well.
Samuel
That's a super good framework to follow. You just mentioned you can ground it on either a SharePoint site or on specific document. I got the feedback a lot that the experience is a bit different. The output is a bit different if you point it to a whole SharePoint site versus like individual documents. I will assume that the architecture is a bit different in the background, but can you explain why?
Abram Jackson
Yeah, and we just put out a documentation article on this because it's not very intuitive and we're working on more product improvements to make this more straightforward for users. So the way that language models work. This going to be a little bit technical, but it won't be too bad, so don't worry too much. The way that they work is you give them a whole bunch of text and then they predict the next thing. That would go into this text. Right, so with base models, which we're not really using here, right? But you could say, that the sly fox jumped over the lazy, right? Give that to the model and would immediately return dog with 99 % confidence. Because that is the Word that almost always follows here. So in chat and instruction models, it's more complicated than that. So. What OpenAI learned with ChatGPT was the breakthrough. It was this instruction following, which is if you organize this text as a conversation, like a play, for instance, right? A two-character play. The first user says this, know, or character A says this, character B says this, character A says this, character B says, and you leave it there, the model will fill in an incredibly intelligent answer of what character B would have said at this time. And, this was the emergent behavior, that was talked about, know, in the literature at this time because nobody predicted that this would happen and that conversation between characters. And B can be a translation, right? And it can be instructions of this a translation kind of conversation and, character A will say something in English and character B will say something in Spanish and then English and then you prompt the model and it will translate whatever that was. Incredible. Just because the AI is effectively magic. I just, when it got large enough. Okay, so the way that this works is only based on the information that's in this conversation. And so the way that, and it scales poorly. While the models are getting better at this, the top models often are limited to 128,000. Letters or characters in this conversation before they can't process anymore. So when you include 20,000 characters of really useful information, it bases the completion of the next thing it's going to say on all of that information existing. And that's why embedding this knowledge works really well, but it's limited. You can only get 128,000 characters in there. You still need to get the user's question and the conversation history and the guidelines and everything like that. So it's limited. We limit it to about 100,000 characters right now. Meanwhile, all of Wikipedia, right, is 72 gigabytes of text only. That's a lot bigger than 100,000 characters. And, you we can't just put all of it into the context window and then ask questions. It just doesn't fit. We don't have enough computing power in the world, perhaps in the universe, to do so. With this this AI architecture. So the workaround is called RAG, or retrieval-augmented generation. Where the model basically asks for more information like I can't answer that unless I go get this information over there and our product code or what we call the orchestrator sees the model has asked for this question. Let me go. With that AI and just go get that information, right? Maybe it's go get the article, Mary, Queen of Scots. I don't know, whatever it is from Wikipediand put it into the context now, because it's pretty small. That's only, 5,000 characters. We can put it here into the model and now the AI is going to be great at answering questions about Mary, Queen of Scots. But if our code or if the model, if that article is actually. I should know this history better. Don't know. Mary II, Queen of England or something like that. And it doesn't show up with the query Mary, Queen of Scots. I should probably say Scotland at least. Sorry UK listeners. We might not get it. We might get the wrong article and then the AI is not going to help. Or the article might be, 60 pages long and will only grab a couple paragraphs out of it and it won't be able to answer based on the full context. So.
Samuel
You.
Abram Jackson
When you can, select individual high quality files in your agent and you're aligned well with how the technology works to get much better answers out of..
Samuel
So if I'm pointing it to a whole SharePoint site, there's the risk that it will pick up the wrong information and pass it as context to the agent and then I'll get a bad answer. While if I'm specifying specific documents on any subject, it will be more precise in this answer because the chances that it will extract wrong information from it are less, right?
Abram Jackson
That's right. Yeah, there's two issues with searching broad forums and we're of course handling this as best as we can. One part is just getting the right information and then also when you get the wrong information, it makes the model worse. So if you've been using the same SharePoint site for 17 different projects for 10 years, it's going to have a lot of information in there that's not helpful for what the user's current question is and what this agent is supposed to be doing. Now it does make a good fallback option, right? So you can do both. You can specify individual high quality files and also say, if these qualities, not even say, right? But just set up the agent for, if these files don't have that information, you do have the ability to search. So you can set up both and you probably should, right? But for important agents, yeah, think about what those knowledge sources are, right? Make sure they're set up well and treat them seriously. Right? It's part of the software, frankly, like what that document is.
Samuel
So for all the organization out there that asked me, I have this 10,000 document repository. Can I just plug the AI on it and ask any question? It's not a good idea?
Abram Jackson
You cannot know and know anybody that tells you that the transformers and LLLMs can do that will be wrong. That's actually a search problem. Now AI and agentic search is helpful, right? It can do multiple searches. It can reason over it and it is getting better and so it is going to be better than just a straight lexical or semantic search on that entire SharePoint site, especially if you use Researcher for it. Because it can do a lot of searching. But the AI cannot, it simply cannot process 10,000 documents at once. Not Gemini, not ChatGPT, nobody's AI can process 10,000 documents. It is a search problem.
Samuel
Changing subject, but you mentioned in one of your article that the agents you're creating should be opinionated, right? Should represent your specific way of seeing a challenge or seeing your work, right?. Why is that? And how will you recommend people out there to create their agent based on this idea?
Abram Jackson
Yeah, and.. Yo, the.. The naive way to use AI is to give it, know, write a sonnet about, SharePoint search, right? Do something like that and it will correctly write a sonnet. It will use whatever 17 line form is necessary, but it's also not going to be that good, right? Because there's a lot of things that you could say about SharePoint search. Many of them probably don't fit very well in a sonnet, but anyway. And the. The overuse of this, and this even worse when there's little buttons with magic wands on them, I really wanted to get the feedback to the LinkedIn team. I really don't like the auto generated answers that everybody just clicks on for giving LinkedIn comments and it produces just very generic. Yo, technically accurate, correctly punctuated. Sure, yes, correct English, very modern correct English. But very generic, boring content. I was really hoping the Word.
Samuel
With a rocket. They all have a rocket. Including a LinkedIn post with the magic one.
Abram Jackson
Yeah, the blowing mind blowing mind emoji and the rocket emoji for sure. And. It's so tedious. I was hoping the Word Pablum would take off for a very. It's a very boring cereal that nobody eats anymore. But the term that actually took off is AI slop, right? Where the mental image is just, scooping in a big thing and just. Putting it on the cafeteria tray or the pig trough. It's not appetizing. It's technically food, but nobody wants it. So that's the generic. If you aren't giving more information about this, I think the example that I might have used in that one was answer this question in a few paragraphs, but also pretend that there is a tiny Tyrannosaurus rex dinosaur that is on your counter and is trying to make coffee. Needs your help as you answer. I've turned a very generic answer into whatever I was asking about into a hilarious essay, about, he's knocked over the coffee pot now or whatever. So that's probably not useful at work necessarily to put, your quarter and financials being interrupted by a T-Rex perhaps. But that's.. If you make it opinionated, you're going to avoid the AI slop. And also, if you are developing software for sale, right, or that you want to take off, can.. Anybody can produce this, the slump, right? They'll reverse engineer your prompt or whatever, and they'll produce the same product, and they might out-hustle you or out-distribute you or something like that, but there's no vote. What you can do is have taste, like.. We're using this Riverside recording software right now, which looks very different from Teams and from Zoom. And, it's very clearly intentionally thought about for the needs of interviewing somebody on video. And yeah, it's not generic. Like if you asked AI make a video recording app with two people, like it would not produce this application. The Riverside designers and product managers and engineers might have thought very intentionally about what this experience should be using their expertise of, producing videos. It made a product exactly for that. They might have used AI systems in it, but they gave a lot of detail. They were very opinionated. We need these features and not those ones and it's going to look like this. So as you're building software, you and agents in particular. But any AI software, right? Be opinionated, right? If your opinion is different than others, your software will resonate with some category of users. That's how you'll grow fans and that's moat. Your moat is actually the taste because it's so easy to build stuff now. So apply your opinions to your AI.
Samuel
I'm often hearing this joke of, you all these email will be crafted by AI and then everybody will use it to AI to summarize those long emails. But that's the point I'm personally using it a lot to craft emails. My my you can hear from my accent like English not my primary languages and I wasn't posting a lot on LinkedIn because it was tedious for me to look at. My grammar, my punctuation, all this stuff. And now with AI, I have the opportunity of crafting better email, of crafting cool LinkedIn posts, but I just don't use the magic wand like I mentioned. I'm not using the out-of-the-box, cool palette for Outlook generated email. I'm putting my idea out there and I'm telling which tone to use and I'm putting my twist to it. And I know a lot of people won't necessarily agree that because..
Abram Jackson
Mm-hmm. Yeah. Yes.
Samuel
Will say no you're not the creator of those of those texts of those LinkedIn posts but I think that the work that's been put inside of the prompt and giving my twist and most of the time I am iterating right I'm not just like taking it like it is I will do some modification etc it I make it my own right.
Abram Jackson
And. Yeah. That is a great use case here. The bad version of this you write three bullet points and you have AI expanded into three formal paragraphs. And because the AI doesn't have any information to go on besides what you wrote in the bullet points, the rest of it is slop. The actual information was in those three bullet points and then it just added lots of filler words. So that is the of the bad version of. But, I use AI in my writing all the time. Not to do the writing necessarily, right, but to certainly research, to rewrite, to identify problems, to find places that aren't clear, to give suggestions of what I could write next. There's a huge amount of value in doing that for writing. Just be careful that you are the one adding the ideas, not the AI, because the AI will give the most boring average generic ideas. And that's just how it's built. Right, the whole thing, this live fox jumped over the lazy blank. Right, it can't. It's not going to say octopus or zeppelin, right? It's going to give the average answer of dog because that is precisely what it's set up to do. And if you think about that in the rest of your writing, what the AI is going to tell you is the most average version of it. Now you can act. AI can actually be very creative if you ask for 20 versions of it. It's going to give them to you and some of them will be hilarious and those can be really fun to do as well, but. So definitely be using AI in your writing, but if you just clicked on a button, be a little skeptical of the quality of that answer.
Samuel
I'm seeing so many answer on LinkedIn polls that are clearly AI generated, super generic using all the same words, same emojis. Yeah.
Abram Jackson
Yeah. Twitter is just as bad. You can always you can always spot them of basically restating the question or they the guys post in slightly different words and then maybe answering asking a generic question. They're all over the Internet. Unfortunately. What's crazy is it's so easy to like be more interesting even if you're using AI to automate posting, which you probably also shouldn't do. But you could make it way more interesting, and most people don't. So anyway, it's still super easy to tell everybody that's doing it.
Samuel
I think we'll work in with solutions to prevent that at some point. Mean, LinkedIn will improve its algorithm.
Abram Jackson
Yeah. I just saw a study of, there's starting to be some stigma of. Using using AI for writing, which is unfortunate because AI is such a great tool for writing when you're using it well. And also I can understand some stigma for using AI poorly. If you're clicking on the magic wand buttons and just posting it without adding any original thought. Yeah, maybe you shouldn't feel bad wasting everybody's time with that just a little.
Samuel
A lot of feedback I heard about Copilotator AI in general, generative AI actually in general is mostly people don't take the time to understand how it works and thinking that the magic ones is what generative AI is, right? I can put like a three, four words question and it will give me this awesome output that will change my life and it will save me four hours, but it's not the case, right?
Abram Jackson
Yeah. And consume 50,000 gallons of water to do it, right? That is the other complaint, which is false, by the way. I won't get into that. Yeah, that's totally true. And even, you if you're listening to this, if you used AI a year ago, well, first off, if you've got that experience, it did a couple of toy things. There was actually really cool stuff it could do a year ago, but.
Samuel
Yeah, true.
Abram Jackson
The latest Frontier models, they are so incredible. If it's been 12 months, you absolutely need to go back and look at them. Now you are listening to a podcast about Copilot, maybe you don't necessarily need to be told this, but try Researcher and Analyst, right? The reasoning models from Microsoft 365 Copilot. Try Think Deeper in Microsoft Copilot. Try Researcher in ChatGPT. Whatever Gemini and PERPlexity versions of these things are with the Frontier models. I get I buy the process or the plus inscription for ChatGPT every few months Do know to see how it's progressed and how it's being used? I did this in April and the o3 model is so So incredible it maybe you've seen memes of you showed an image right and says you could say GeoGuess this And we'll like inspect like the plants or like I don't even know what it's doing like the type of rock. Is. And then get you down to about a 15 mile square area. That's what it is. I did this, looking at a rock in my neighborhood and it's like, those are western sword ferns. That siding looks like it's, very Pacific Northwest. Alright, so we're in Pacific Northwest and I see a slope on the hill. So you're probably not, well first off this a residential suburb area. We're not in the Snoqualmie Valley. You gotta be a little bit farther east. I bet in you're in the Cascade foothills. And then it gave a couple of towns that left next to. I just took a picture of a rock. Anyway, o3 can do that about anything. Not geogas images, right? But like anything that you want to do. It's really astounding and you owe it to yourself to give it another chance. Researcher and Microsoft 365 Copilot also uses o3, so that's a great place to get going with it if you've got the license. Otherwise, you gotta pay the 20 bucks. Sorry, give it to Microsoft's competitor. You gotta pay the 20 bucks and try it out. Or let's plug consumer, Microsoft Copilot, get Copilot Pro at least and try out those things in your daily life. Whether it's groceries and meal planning, everybody always jokes about travel scenarios being overused. Just whatever your work is, tips on managing your toddler's tantrum. Whatever like it'll research it. It will give you suggestions. You definitely need to check out AI again if it's been a little.
Samuel
The pace of innovation is just crazy. And it's so funny you talked about this geolocalization or geosearching because I found a video yesterday of a YouTuber doing that and I'm just thinking you might be using AI finally where you challenge anyone in the world to take a picture and it will tell them exactly where they are. Probably he's using AI.
Abram Jackson
Well, those. GeoGuessers were doing it before o3. Maybe he is now. But they were very impressive last year too. Just unbelievable stuff that some people can do there. But now that I can, that guy's out of a job.
Samuel
Yeah, true. I think we'll continue to have his audience. How many agents are you using?
Abram Jackson
Yes. I probably have 40 or 50 agents installed for myself and I use them variously. Some of these are because I am testing out the features that we are building so some of them are certainly that they're not necessarily daily use agents. I get a lot of use out of my critical feedback agent for sure. I get a lot of use out of agents that act like. My bosses or peers that review my documents and I can have them have the agent review the document first. Really saves me a whole lot of time doing that. I've got other agents that help me specifically with writing. Know, like strategy documents or something like that. Make sure that these things are laid out correctly and making the right points and also I can tip you. Standardize that by creating an agent for them. And then just other stuff that come up. It's so easy to make agents in Agent Builder and Microsoft 365 Copilot. It only takes a few extra seconds beyond prompting, and then I can share it with somebody. The other things that I try, I've got an agent that asks me questions. I demand that it asks me several questions until it completely understands what I'm trying to ask and then answer. And I find that often gives me better answers if it does take a little bit more time to use. So I've got maybe a dozen that I use pretty regularly. Certainly I get a lot of you sort of Researcher. I mentioned that multiple times here, but I'm very excited about it for sure. That is a good one. And then a lot of these are just pretty standard for me. And then Microsoft has agents as well that help out with your specific Microsoft. Processes. So these ones are just easier than whatever other system that we had. So those ones are always great, I don't. Yeah, I don't. I'm in need to order a new computer every few years, so I don't interact with those that often.
Samuel
Yeah, I do agree with the Researcher one. Mean, I'm using it almost daily. It's crazy good. Mean, to be honest, prior to the Researcher agent, I was using PERPlexity because it was really good at that. Researching using a reasoning model, et cetera. But now that I have Researcher, it's not only searching the web, it's searching my organizational data. So if I need to prepare for a specific customer, can look at all my transcription, all my email, and it's really, really, really good. I'm really totally impressed by the Researcher agent. It's very powerful.
Abram Jackson
Yeah. Yeah, for. Sure. Sometimes I'll use it just for a single Word answer just because it's so good at searching. Even if I do feel a little bit bad about the energy usage. No, it can't, but the but the answer will be somewhere in that report. If I simply cannot find it and I can't work with Copand prompted to find it will be like alright, I'll ask Researcher. I'll come back in 10 minutes and then I'll stand through it and find that answer for that.
Samuel
Can it provide single Word, can it provide single Word answer? Okay. Okay.
Abram Jackson
That name or that project that I looking for. It's very good at that.
Samuel
In one your articles you were mentioning, starting sometimes building the plan of what you want to execute with GPT-4.0, For instance, and then asking to o3 to expand on that plan. Because one, GPT-4.0 Is faster than o3. Now I can see the same with using standard Copilot and then the Researcher.
Abram Jackson
Mm-hmm. Yeah. Totally, yeah, that's absolutely true. This actually just a pattern that I think more people should be using of multiple agents in one conversation. When we designed and built these agents in Microsoft 365 Copilot, were pretty intentional. Other platforms weren't doing this. Like ChatGPT, couldn't do this at the time. In ChatGPT, you'd open up the custom GPT and you would. Interact with it and then you'd be done and you'd close the conversation and you start a new conversation. What we wanted to do was recognize that, in order for you to complete like a large project, there's a lot of roles that you're going to interact with. As I work on shipping Researcher capabilities, to the world, I mentioned privacy earlier, right? So privacy experts are one of them, but security experts is.. Another person that I work with, the responsible AI people, our deployment safety board. There's several different people that I want to work with in order to just get the full gamut of reviews, of the compliance reviews of this feature. So all of that's the same project and it's all related. And if I.. Go interact with each of these people individually or which each of these agents individually I've got to repeat all of the context of what I'm doing and what I've already just worked out with the privacy expert so that the responsible AI expert can can know it So anyway, most work is like this right your boss decides your project you work on that project for days weeks months. And if you keep that conversation going, you don't have to keep repeating yourself. You're going to start getting better answers. And you can bring in specialist agents for individual pieces of it as necessary. And they've all got the context of the whole conversation. So that is a very helpful pattern of working across many agents, is keeping the conversation going and having all of that history. So totally recommend.
Samuel
I'm seeing time's flying. I prepared like 14 questions. We're at number five. Most of the question I asked you just came naturally. Almost like I can talk to you for hours. It's really interesting to see how you address it, how you add it to your daily workflow. But for my last question, I'd like to know, where do you see that evolving?. How do you see cool pilot evolving or agent evolving? In five years. It seems like a lot in our world.
Abram Jackson
Five years. Yeah. Well, what's really interesting in this entERPrise space is the technology is moving so fast, but entERPrises take time. And certainly I've seen this for the previous AI features that I've worked on before this. Adoption is going to take time. Now this I think is going to be adopted in this already being adopted faster than you any technology in history. But still, it's going to take some time. So five years in AI progress is really going to be incredible. We've you we kind of hit the limit on your pre training scale, right? You GPT 4.5 Was kind of the last one for pre training scale, but OpenAI discovered inference time scale and these reasoning models, and there's so much more headroom there. So there's a clear path for. Continued model improvements for at least a couple of years. And if it's anything like the last couple of years, it's hard to predict what even those things are. So the technology is going to keep moving fast. I think what I can predict is, certainly in five years, the entERPrises that are working hard to adopt now are. Going to be leaps and bounds ahead of their entERPrises that are not and have not been working for the last couple of years to adopt this. Because if you don't start until 2030, it's going to be too late, right? Your competitors, they'll be 30%, 50%, 100 % more productive than you, which means they are getting so much more done. They are going to be out shipping, out distributing, out selling you at every corner for whatever your business is. So what the individual work looks like? Everybody is going to have to be using personal AI assistance. Because their peers are just going to be so much more successful with it, then they are if they're not using it so everybody is going to be using this. So I think this advice goes to individuals too, like put in those 10 hours. You're be way ahead. There might be new things to learn next year. There's almost certainly going to be new things to learn next year. You AI is going to keep getting better. You'll be able to use it for more things. You're going to be so far ahead of, I'll say, if you're not using AI and you're not learning about it now, you're going to be so far behind in one year, certainly five years. It's going to be hard to compete at that time.
Samuel
Yeah, and it's a muscle that needs to be built in the sense that people don't add up AI day one. So if you're too long, you'll have to start from scratch. If you start in 2030, I mean, the curve of adoption will take a year or two. You'll be really, really behind. Agree. But like, how do you keep, sorry, with my last question, but this the last question. How do you keep with the pace of innovation?
Abram Jackson
Yes. Okay. I work a lot. It's probably the main answer here. And when I'm not working, I am learning. As many complaints as we can have about X or Twitter and its current owner, it remains the best place to learn what's going on. Like what people are finding is working in AI, what the latest papers are, what the trends are, what the latest tools are to try. And so on weekends, you all try out the new tools too. So I spend several hours a week scrolling. I research for my newsletter, some of these new tools that are available. So I work a lot. It's kind of the answer. Maybe that's not great. For everybody. Yeah, it is what it is. It takes a lot of work to stay up to date.
Samuel
Yeah, I'm the same spot trying to just keep on. Awesome. Thank you so much for this discussion. It's been super, super insightful. If people want to hear more of you, you have a blog, you have a podcast, you have a newsletter, you're active on LinkedIn, on Twitter. So there's a lot of place to follow.
Abram Jackson
That's right. Yes, those are the main ones.
Samuel
Am I missing somewhere?
Abram Jackson
No, no, that's it. I've also got mastodon and threads installed, but I don't use those so yeah.
Samuel
Awesome, I'll put all those links inside the show notes. Thank you so much for joining me today. And maybe a part two, I don't know, can continue asking you questions for hours, but thank you.
Abram Jackson
My pleasure Samuel. Had a lot of fun chatting with you today.
Samuel
Great, take care.
Abram Jackson
Bye.