Why Your First AI Project Probably Shouldn't Be a Chatbot

**[00:00:00]** Your first AI project should probably not be a chatbot. And I know that sounds strange because chatbots are usually the first thing that people picture when they think about AI at work because of ChatGPT and all that. But a chatbot, though it is easy to imagine and it is visible and it is easy to demo because you can put it on a website, you can ask it questions, and it just feels like you're doing something modern and what other people are doing. It is useful but it might not be the first thing that your business needs to do. A well-designed support assistant can help customers find answers faster and a an internal knowledge base assistant can help employees search for procedures. And a sales assistant can qualify visitors. Yes, of course, these are all great uses and we do some of these all the time. However, for many small and medium-sized businesses, that might not be the best first project. Doesn't mean it's not a project that you're going to do. It's just might not necessarily have to be the first one. Why? Because voice assistants chatbots, these are all non-deterministic. There's conversations coming in from people outside the business. And you don't know exactly what they're going to say and you have to plan for that. And the inputs are not always the same and the outputs are not always the same and there's a lot involved in getting that and a big part of getting a chatbot or voice agent that works well within an environment of a business is having that knowledge layer properly defined and readable like we've talked in previous videos. So, if you don't have everything first then you might want to clean up your data first and get everything in order before you do a chatbot, especially before you put a chatbot out there in the wild. So, for many small and medium businesses, the better first project is **[00:02:00]** usually a type of workflow because a workflow has a beginning, it has an owner, it has inputs, it has decisions, it has outputs, and you plan what those are. So, you understand that flow and you can measure it. It has a place where the work succeeds or it breaks, and that makes it much easier to to test and determine and easier to improve. For your first project, you want something that's easy to understand, test, and measure. That That's the point. Your first AI project should not be the one that looks the most impressive. It should not be the one where the business just shows off that you're doing something modern. It should be the one where you actually get immediate, measurable value because a chatbot often fails that test. Not always, but but it often does. A company adds a chatbot to a website, it answers some questions, and then maybe it collects a few leads. Yeah, that's all great. And again, it works. It's very easy to just put it out there. And it looks very interesting. Then, a month later, leadership asks, "What did this actually improve? Did it increase qualified leads? Did it reduce support tickets? Did it improve conversion? Did it save the team time? Did customers like it? Did it handle the questions safely? Because sometimes it can go off the rails. Did it create better data? Did it connect it to the next step in the process?" Sometimes, the answer is unclear. And they are These are many questions just for what little thing. So, an unclear business value is where that AI project might fail. Gartner actually predicted that at least 30% of generative AI projects will be abandoned. and it's one of the main reasons it's because of the unclear business value. That should make you slow down before **[00:04:00]** starting with the flashiest idea. Again, they work great. We we build them all the time. The problem is a lot of times the idea of a chatbot is great, but the actual deployment of it is very different from a proof of concept that you can easily demo. Something that actually works in the wild is a whole other beast because of all the messy inputs that you're going to get and the protection that you have to put in place such that people don't start asking it wild things and have it do things that you didn't plan for it to do. Again, because you're not the one putting in the inputs for a chatbot unless it's an internal one. Still, that's another fun story. But especially in the wild if you want your customers to use an And also, anything that's LLM related that you put out there, it's something that's going to be using tokens all the time while people are using it. So, it can get very costly as well. So, just something to consider. They're great if you have by all means, I'm going to preface this with if you have a great flow already that's working, you already have your people talking you maybe you already have a chatbot that's not LLM based, but you have a chatbot that's sort of hierarchy based, then yeah, cuz you already have that flow mapped out. If you have a very good process that works, then yes, a chatbot might be a good next step because you already have the groundwork. You laid it before and you know it works and now you can test against it and do AB testing versus your team or your existing knowledge tree based bot or the new one. But if you have not if you have done nothing regarding AI in your business and you've not checked if your business is legible to the AI, then a chatbot might not be the best thing to start with. Chatbot says, "Look, AI can talk, right?" A workflow says, "Look, this process is now faster, cleaner, or **[00:06:00]** easier to manage." For a first project, I would rather have something that you can actually prove that it worked. So, I'll I'll try to give you some examples. Suppose a new lead comes in through your website. A chatbot might answer a few questions for that visitor. And that could help. But look at what happens after the lead is submitted. Someone still has to read the form. Someone has to decide whether the lead is qualified. Someone has to create or update the CRM record if the chatbot is not doing it. Someone has to write a response. Someone has to assign the next step. Maybe you have to determine who's the next rep. Someone has to remember to follow up with the prospect if they don't reply. Someone then has to route the lead if it belongs to a different person. And that workflow is often where the real money leaks. It the leads sit there too long. Good prospects get slow replies. Bad fit leads consume the attention. And the CRM records at the end of the day stay incomplete. We've been talking about this a lot. We recently had a stream where we were talking about the problem of people having perfectly good leads and nobody following up with them. People showing genuine interest in a thing or a service or a widget and then nobody following up. That's not going to work. And it doesn't matter if you have a fancy chatbot on your website if the flow that happens afterwards is not properly maintained or generated or created. Then it's not going to give you much help. You're just going to create a bunch of extra work to go into a bottleneck of a pipeline that doesn't already work. So, first you might want to see what is happening right now with the current leads. Are your employees, team members, **[00:08:00]** reps handling all of this already? Well, do they have a proper workflow so that all that pipeline runs through efficiently and you can measure it and you can see where the bottlenecks are and what the objections are and all that? And if so, then yeah, maybe you can start thinking about a chatbot. But if you don't, then not only for the chatbot, but but it will continue that every single lead that comes in is not being properly followed up with. And that could be a very bad thing. So, if you start with a chatbot and then you try to improve that workflow, it's it'll just make you not focus on the most important thing, which is actually servicing whatever that queue that the chatbot is creating for you. So, in another scenario, when a lead comes in, AI summarizes the inquiry, it checks the lead against qualification criteria, it drafts a response using approved examples like we've talked about in other videos, it creates or updates the CRM record, it assigns the next step, it flags the leads that need actual manual human review, it triggers a reply, uh it might do a reminder, and then if a human approves anything sensitive before it goes out, then you'll be a little better off because it won't just send things that might make no sense in logical terms. Um and though that part might not be as flashy as a chatbot, it is closer to the revenue. It will fix a real problem that you might already have, and it will actually help your team. And you can measure it. So, the workflow part, if you start with that, then you can measure things like how fast did we respond? How many leads got followed up? How many CRM records **[00:10:00]** were complete or incomplete? How many did we clean up? How much manual copying and pasting disappeared? How many opportunities stopped falling through the cracks? That is where you should focus. So, like we've said in other videos, you want to focus on getting those workflows fixed before adding more to the pipeline. And a chatbot will add more to the pipeline if you have the traffic, right? If you don't have traffic, chatbot's not going to do anything for you. But if you already have traffic coming into your website, or you have calls and coming into your office, then yes, these tools are really good, but they will just create more manual work if the rest of the pipeline has not been worked on. You might not even need AI. You might want a workflow first that handles a lot of that manual copying and moving around of data and assignments, and not even require AI. They could be deterministic. Then, in some parts of it, yes, you might have some AI to help you organize, prioritize, determine if something's going on that shouldn't and should be stopped, something that needs to be flagged, things of that nature. Or you want something to start using the brand's voice, the company's voice, what what best practices you have for follow-up, the ones that actually work. What are the questions that it should ask to make people more engaged and actually get closer to that sale. That's really where you should be focusing yourself on because again, chatbots, they've existed for a long time. Now they've gotten a lot better with LLMs, of course. They can they can talk and answer questions and retrieve information and and actually get data from your knowledge bases and sources, but if you don't lay the groundwork properly, they can again create a lot of additional work for you. And your customers are not going to be happy because you made it very easy for them to send something, but then **[00:12:00]** there's nothing just crickets, right? Just there waiting. And I think that was what I was talking about in the interview the other day. We are now and have been for many years, I think, sometimes afraid of actually filling out Maybe afraid is not the right word, but filling out contact forms on websites cuz you just feel nobody's going to see them. I see a lot of these companies, maybe even providers that I might want to use their system and I see a contact form. I'm like, I don't know. I'm going to send it. Nobody's going to reply to me or might go into a ticketing system and get lost there for weeks. I don't understand why that happens. I mean systems exist for people to be able to be reminded when something's there, but the same way they made it so easy for someone to just send something, then they might be getting spammed and getting a lot of unqualified leads or unqualified comments and so they can't actually even get to your conversation as a customer. Because they're just too saturated and the same thing will happen with a chatbot if you don't plan that properly. Because it's so easy for them to just chat. If you don't have the rules in place and give it the background knowledge that it needs such that that chatbot can actually guide them to where they have to go. And if they're not qualified, then take them somewhere else. Maybe give them some more education, give them some free guides, give them something that they actually need or will be useful for them to get closer to where you need them to be. You don't have to get everyone through the chatbot and into your lead pipeline, because maybe they're not leads. Maybe they're in a completely different part of the triangle. So, just keep that in mind. Maybe you need to build out your knowledge base first and then talk about setting up a chatbot. If you don't map that workflow first, then again, that I think **[00:14:00]** preference here, you might not need the chatbot to be that first thing. It might be the second thing. But the first thing is lay the groundwork first such that you can see how AI and workflows can work in your company. Make sure your data is clean. Make sure your data is legible. And see which boundaries should exist. And that way, when you do put in a voice agent, a chatbot, or anything like that, it's not vague to the AI. So, now the AI can do the good thing or the thing properly, I could say. So, if you are planning to do anything like a chatbot or a voice agent or anything that's public on your site, just ask yourself a few questions first whether that will automatically help you improve your lead to purchase cycle or will it actually help people fix their problems or could it accidentally, because you didn't lay the groundwork, give people the wrong impression or lead them towards the wrong product or give them the wrong information about a product and then you get unhappy customers or give them an incorrect answer in the support question because it didn't have the right access to the right knowledge base. That is something that you need to consider. So, when we think of what we should do first, think about workflows. What an automation could do in your business that will automatically kick off tasks that your team already needs to be doing, but that only some of your team members are doing because they don't know the process. If you can automate it because it's something that's repeatable, then yeah, **[00:16:00]** maybe your first automation or your first workflow is going to be something simple that just puts everyone in the same level playing field so it sets your highest earning reps have the same advantage as someone that just came in, that's just being onboarded because the tool can now allow them to check for urgency, qualify the lead, check if it's accurate, give them the right context so that they can communicate in the same way, flag anything that's sensitive. It's not let's just give I every AI to everything and every process. It's a figure out which ones are leaking and see if you can create a workflow, map it properly such that it will give you a better outcome. Many service businesses spend too much time just copying all proposals. And then, yeah, AI can help, but only if the workflow is clear. A good first project there would be uh it might gather the discovery notes, client goals, scope details, pricing options, and approved language, and then the AI can draft that first version of a proposal, and then a human can read it, and the system checks whether the required sections are included. That's super helpful. And the final proposal still depends on human's judgment, but there's no more blank page, like where do I start? And that just alone can save you hours. That also weekly reporting. And this is not glamorous at all, but it's one of the most useful workflows. Every week someone gathers numbers from one system, another system, another system, and then they paste them into a report or a spreadsheet or multiple spreadsheets, they write a summary, and then they explain what changed, and it takes so much time and mental effort, and then they send it to leadership. And leadership might say, "Oh, can you change these two things?" And they have to go back and change a bunch of other things down the line. Now, a workflow first AI can pull the numbers, highlight the changes, draft a plain English summary, and then ask the manager to review it **[00:18:00]** before sending. That way you have a human in the loop, and that can remove a recurring Friday afternoon burden that they always have to do, and it just takes so much time. It can't stop flashing, but it's very useful. And useful here is what you're going for. Your first AI project should create momentum. It should make people say, "Oh, wow, this this actually helps." And they'll they'll want to do more, and then the second project is going to become easier because now you've proven how AI starts to fit in to real work. And a chatbot can do that if it's tied to a clear workflow, but one that just sits there on the side is not really connected to anything. It's more of a novelty thing that that may or may not answer the questions the right way. And it might just create a new place that you need to monitor. It might just produce transcripts that nobody's going to review, and it just may collect leads that still fall into the same broken up follow-up process that you had before. So, if the workflow after the chatbot is broken, the chatbot has not solved the real problem. It just moved the starting point. It It's uh just something you need to consider. And the fortunately the demos are very exciting. Demos about chatbots and avatars and AI avatars look amazing. And and voice agents are amazing, especially if you compare them to an IVR or just compare them to waiting in line, right? In in a call queue. They're good. Just think of what comes before that. Improve that, whether with workflows, just regular automations, and or AI. And after you've done that, then yes, let's consider can we do something that will automate that intake, that will automate the support that we already have down and nailed. How can we use AI now to have that interaction with people? So, try to use **[00:20:00]** this filter. First, pick something repetitive. If it happens once a year, then don't want to automate. Just going to spend a little money on it. Look for work that happens every day or every week. Then, pick something with clear inputs. That's why chatbot doesn't work for this first thing. A form submission, a support ticket, a signed contract, a meeting transcript right after a meeting or do even during a meeting, a set of numbers, a document. Uh because AI needs something to work with that that you can measure so you can make it better. Third, pick something with clear outputs. A draft email, a CRM note, uh task list, a summary, a report, a recommendation, a routed ticket to another team member. Because if the output is very vague, the project will be hard to judge whether it works or not. Then you're going to pick something with a manageable risk. You don't want your first project to be something with the like the most important thing in the company, like legal advice or financial approvals or medical judgments or anything where a mistake could cause serious harm. Just think about something that if it works, great and if it doesn't, then you figure it out and you fix it and then you keep going, but just don't do something that will break the whole company just because you wanted to play with AI. Start with a queue a human actually review the output and the cost of making a mistake is actually manageable. Fifth, you're going to pick something where the team feels pain. So, if there's no one complaining about it, then nobody's going to really care about it. However, if it's solving something that's really annoying and the improvement is obvious, people are going to be more likely to use it and they're going to be on board with using it. And of course, pick something that you can measure. Whether that's time saved, response time, fewer errors, fewer missed follow-ups, more complete records, faster onboarding, less manual **[00:22:00]** copying, because if you can't measure anything, you will just struggle to prove value. And pick something that already has available context. Now, this was huge. We talked about the AI legibility, businesses being legible to the AI. Uh we've gotten into so many projects where you can have everything, but if the company does not have enough context for that workflow, that workflow is not going to be a fit. You should scrap it, do something else. Because you're going to spend so much time trying to gather the available context just to be able to try to implement something that you can measure. If that is not there, that context is not there, please go try something else. While if you want in parallel, get that context so you can do it later, but don't don't choose something where you don't have the initial context. Like let's say you don't have you're going to do something related to incoming calls or invoices, but you don't have enough invoices to train the workflow, then the people, either your team or whoever's doing this for you, the implementation, the deployment, they won't have anything to test against and they won't be able to proceed. So you'll just start with a lot of momentum and then suddenly hit a wall and stop because you don't have the available information to keep going. So find something, apart from everything we already mentioned, that you already have a lot of examples and use cases and good inputs and good outputs and bad outputs and bad inputs, such that you can train it to do things and you can actually measure, "Hey, it did it properly." Or it didn't do it properly. That's really important to consider. Again, comes back to the legibility for the business. Does the system have the knowledge that it needs to proceed? Are those rules and standard operating procedures written down? Are there examples available? Do you know what good actually looks like? **[00:24:00]** Does the data live somewhere that it's accessible? And if not, then the first step may be capturing that knowledge before automating. So, like I said, it's not a delay, it's just part of the work, but don't choose the one where you need to find context first. First, do one where you have it. At the same time, if you want, if you have the resources, go and get the context for that second one or third one. Um because you don't need it to be the most giant project. It just needs to be something that will work and be proven. So for sales workflow, it might mean uh getting those qualification rules, approval examples, objection handling and routing rules. And for support, it might mean policies, answer templates, escalation rules and examples of difficult tickets. For onboarding, it might be like uh checklists, required documents, roles, timelines and kickoff templates. And for reporting, it might mean trusted data sources, metric definitions, thresholds, and a sample report. That is company brain material. That is what turns the generic AI into useful AI. Now, let's talk about what a strong first AI project brief should include. You should be able to answer these questions on one or two pages. What workflow are we improving? Why does it matter? How often does it happen? Who owns it? Where does it start? Where does it end? What tools does it touch? What information is needed? What outputs should be produced? What are examples that show good output and bad output? What should AI draft, summarize, classify, route, check, or recommend? What should simple automation handle without AI? What part should the humans approve, and what part should the AI never do? How will we measure success **[00:26:00]** after 30 days? And if you cannot answer those questions, then that project might not be ready. And it doesn't mean the idea or the project is bad. It just means that the workflow needs more clarity. And this is where an internal AI champion can help. We talked about this in the last video. The champion does not need to build everything, but they can map the workflow because they're there on the ground, and they can see what people are doing. They can gather the examples. They can write the brief, and they can test the output if they're there. They can help train the team and make sure that the project actually deploys instead of just launching. We talked about this as well in another video. Deployment does not mean implementation. Launch means it exists, but deployment means it's actually now part of how the work gets done. And a lot of AI projects will launch, but not all of them actually deploy. So, your first project should be designed for the deployment from the beginning. That means that the workflow must fit the team, and the team must know what changed, and the output needs to actually be useful, and the review points must be clear, and the success metric must be visible. The system must remove work, not just move the work to another part. And that's where chatbots often create trouble, unfortunately. A A chatbot may create a new channel without fixing the underlying process. Now, somebody has to monitor it to make sure it's not saying things that it shouldn't. Someone has to review those conversations. Someone has to update the knowledge. Somebody has to handle those escalations. Somebody has to answer when it fails. You might need a backup people there. If that workflow is not designed, then the chatbot just becomes another responsibility, and small businesses do not need more responsibilities. Just with a better name, it's working systems. That's why I would start with a workflow for something **[00:28:00]** that's already painful, and not necessarily a chatbot. So, do not ask where I can put a chatbot. Ask where does the work currently get stuck. Where does a team copy information between tools? Where do our customers wait? Where does a founder get constantly pulled in to make routine decisions every day? Where do the reports take too long? Where does a quality depend on one person's memory? If they're gone, it's a mess. Where do the leads always fall through the cracks? And where do support tickets get routed incorrectly? All this is very important to ask yourself. That's why your first AI project AI project probably is more aligned to getting your typical day-to-day work improved. Once you can improve one simple loop, then you can improve another. The first one will teach you how to gather the context and get everything ready. It teaches you how to define how human review works. It teaches you how to measure the value. It teaches your team how to trust AI safely. And it teaches your internal champion how to write better briefs. And so, it will help continue to build your company brand just one workflow at a time. And that's how AI capability compounds, not through one giant transformation, not through a chatbot that sits alone on the website and has nothing to do with the rest of your tooling, through repeated practical workflows. That's how you win. So, here's my practical exercise for you today. Before you build a chatbot, write down those five workflows in your business that happen every week. The The I mean, that could be the new lead to first follow-up, the signed contract to kickoff, the support ticket to resolution one, the discovery call to proposal, that's always a big one, meeting ended to tasks assigned, that's also super important, invoice received to actually approved, weekly numbers to leadership report, maybe it's repeated **[00:30:00]** internal question to trusted answer. Just pick one. Then ask which one is repetitive, painful, measurable, and low enough risk to improve first. Don't choose the most impressive-sounding one. Choose the one that will make your team's life easier in the next 2 weeks. That is usually a good starting point. Then map it. Draw it. Use tools to figure out how to how to align it. Find the trigger. Find the owner. Find the inputs. Find the outputs. Find the handoffs. Find the rules. Find the examples. Again, good examples and bad examples. Find the review point. Find the metric, and after that, the AI decision gets much easier. May you still need a chatbot. Great. But, now the chatbot will be a part of a workflow. It will have a job. It will have boundaries. It will have a handoff. It will have a way to improve because you can measure it. That is very different from let's add a chatbot just because it's fun or it looks impressive. So, the takeaway is simple. Your first AI project should not be a symbol that you're modern. It should be a system, not a toy, not a demo, not a vague system floating around the business. It has to be a workflow with clear inputs, clear outputs, clear rules, human review, and a measurable result. Start there. The chatbot can wait its turn. And when it's the right time, you'll be ready to deploy. And if you want help finding the right first AI workflow, check out our AI opportunity map. It's on our website paracosano.com. And it will help you identify those three to five workflows that you should look at. It would help you find the risks to look out for, and then **[00:32:00]** you're it will give you a two-week action plan. It's designed for a small and medium businesses that want practical AI value, and it's really cool. Just go through it, ask you a bunch of questions. It's not very long, and it will give you actionable insights to see how much money you might be leaving on the table, where you should start based on your team size, and where the data is being collected, and how the data is being managed. So, it's very similar to what we've been discussing today, and that might actually give you a first idea to how to move forward. Then, go and actually build the project that deserves to go first.